You could open a textbook of artificial intelligence or do a quick search on the internet, and you’ll see that definitions of AI vary. There is no good way to tell what AI is, where it is used and what role it plays. It could be a complex stand-alone system like a robot or an autonomous car. It could be just a few lines of code inside another piece of software, playing a small role.

Artificial intelligence involves a collection of programs that do a diverse set of tasks. Mathematically and algorithmically, the lines blur – there is no clear indication of where AI starts and other technologies stop.

Furthermore, many experts disagree with the use of the word ‘intelligence’ – artificial intelligence has no resemblance to human intelligence! Yet, the word suggests to us what these programs are meant to achieve – the thread that connects them.

Ultimately, AI systems are machine based. They make predictions, recommendations or decisions by

• perceiving real or virtual environments (using items such as microphones or cameras),
• simplifying data and analysing it,
• using that analysis to make a decision or prediction¹.

When you come across a system that:

• recognizes what you have written or what you are saying (translation software, text recognition, face recognition, personal assistants, chatbox); or,
• seems to know you better the longer you use it (Youtube video recommendation, news recommendation, Amazon item recommendation, Facebook suggested friends, targeted ads) ; or,
• is capable of predicting an outcome, given incomplete and fast-changing information (fastest route to go somewhere, stock prices in the near future),

...AI is probably involved.

The AI around us

Artificial Intelligence has become the technology that gives its user the edge that's required to succeed.

Almost every field uses AI in one form or the other:

• From business to research, many fields use language apps to transcribe speech on the fly and obtain translation of impressive quality.

• Medicine profits from image analysis and AI-based decision-support tools².

• In agriculture, AI driven systems help optimize the use of available resources.

• Every day there's newsabout AI breakthroughs in games, art, industry and commerce.

At school, choosing courseware, adapting to individual learners, assessing them in a productive way and managing logistics, are all “intelligent” activities. According to the definition(s), AI-based software should ideally be able to help with pursuits such as these.

Alan Turing is considered by many as the father of computer science. Many of the new ideas in AI today were actually also introduced by Alan Turing, before the term 'Artificial Intelligence' was invented!

¹ The OECD AI Principles, 2019.

² Artificial Intelligence in Healthcare, Wikipedia.

Further into this book, we will look at the artificial intelligence tools available for education. But many of the most useful applications have yet to enter the classroom. The Educational technology industry, big digital corporations and university research labs are all developing tools to help the teacher teach and the learner learn. More and more, corporates specialising in AI are making huge investments in education. Tools, both approved and unapproved by the concerned authority, but available freely, are used by teachers and students.

Whether made with education in mind or not, many of these tools may be used in the classroom. It’s necessary to be aware of their benefits and potential problems .

One of the free applications for mathematics available during the writing of this text is Photomath. (For language teachers, a similar example might be a language learning app as Duolingo or writing software that uses GPT3.)

Photomath is a mathematical solver. It takes a mathematical equation and solves it. Teachers have been contending with calculators, as a tool for teaching as well as cheating.

What makes Photomath so powerful is the ease of use – just click a picture of the blackboard or notebook. The AI in Photomath scans the photo and solves the problem directly.

Let’s say a calculator gives you an answer of 42. Teachers may allow their  use to check the result, but the students must arrive at the solution by themselves. Solvers such as these show multiple ways to solve and visualise a problem, although this part is much less technically demanding for the programmer

Other applications found in today's classrooms:

• Search engines
• Spell check and grammar correction which is built into most writing software
• Online translators
• Language learning apps
• Math solvers like Photomath, Geogebra and Wolfram
• Personal assistants.
• Chatbots
• Intelligent tutoring systems
• AI-powered learning management systems

Is using AI cheating: What are the teachers’ reactions?

Here are some teacher's reactions to pupils’ use of AI:

• Under-utilisation of technology due to fear and ignorance
• Indiscriminate use that might lead to undesirable secondary effects

For example, excessive use of mobile phones can be harmful. Some societies avoid mobile phones entirely. However, most people don’t overuse them. Prudent use of mobile technology has, in fact, saved lives.

In order to avoid succumbing to the first example above, it would be helpful to have knowledge of important educational applications. We will take a closer look at each of these in the upcoming chapters. Here are some examples.

AI tools for managing learning

Artificial intelligence dashboards, other data visualisation tools and learning management systems bring all the available information together. These help to monitor student performance over multiple subjects or track progress on any topic in classrooms of any size.

AI applications can flag potential problems, such as absenteeism and behaviour common to dropouts. All the data thus gathered can serve as self-assessment for the teacher by showing where the lessons are effective and where a change of approach is required.

AI is suitable for scheduling and resource-optimisation tasks. But the most important application is that which deals with inclusion and integration of the differently abled. Human-machine interface has never been as seamless as it is now, making multimedia input and output a real possibility. For example, the app Storysign helps translate words to sign language to help deaf children learn to read.

AI tools for personalising learning

• Adaptive learning systems (ALSs) evaluate the learner, be it through quizzes or real-time feedback. Based on this evaluation, they present the student with a predefined learning path. Instead of a one-size-fits-all approach, students can spend more or less time on each topic, explore new and related topics. This adaptive software can help them learn to read, write, pronounce and solve problems.
• ALSs can also help learners with special needs. Any specialisation of the systems will be based on proven theories and expert opinion. Targeted systems “are likely to be of great assistance in teaching individuals with cognitive disabilities such as Down Syndrome, traumatic brain injury, or dementia, as well as for less severe cognitive conditions such as dyslexia, attention deficit disorder and dyscalculia”¹.
• Different groups can be formed for different activities (‘clustering’), taking into account the individual strengths and weaknesses of each member.

While these technologies can help, “the devil is in the detail of how you actually use the technology”². The same innovative and powerful learning technology can be used effectively in one school – and badly in another².

Again, knowledge is the key!

¹ Alkhatlan, A., Kalita, J.K., Intelligent Tutoring Systems: A Comprehensive Historical Survey with Recent Developments, International Journal of Computer Applications 181(43):1-20, March 2019.

² Groff, J., Personalized Learning: The state of the field and future directions, Centre for curriculum redesign, 2017.

In the context of education, we do not often have to face such fast-changing topics. It takes months to design a syllabus and years to train the teachers. And some say it takes a generation to comprehensively treat a new topic.

Yet here, the goal of the AI4T project and thus, of its learning resources, has been to train teachers in AI while it unfolds! Therefore, when reading this textbook, you may well find small errors. You may also find statements that are no longer true: technological progress can have proposed a new opportunity; risks that were deal breakers in 2022 or 2023 may well have become acceptable in 2024; concerned laws and regulations may have changed. Does this disqualify this effort?

We hope not. Not if we take the opportunities offered to us by technology and the open nature of this textbook to allow it to evolve with the changes. Perhaps the old-fashioned process of releasing the first edition, then the second edition etc is no longer the best way to edit a book? Definitely, there exist solutions to allow such an object to be shared as a common good, not just in its usage but also in its evolution? This would represent, probably, the next challenge.

This textbook is now in the hands of teachers, its readers. For this to have happened, a large number of people were involved. They helped us understand the needs of the teachers and those of the ministries. Some were involved in preparing the material, in proofreading and in setting up the online platform. Others helped with the translation. We also worked with external teams: Dagobafilms were great in helping us to prepare the videos. And, most importantly, there were all those who encouraged us. Writing a book is always a long and complex task, and you need the help of family, friends and colleagues to be able to recover from the associated frustrations, to stick to what we believe are the right editorial choices and to make sure that the destination that we had set is still where we want to go.

When thanking people who helped us over the past three years, we are bound to forget some. We hope they will forgive the vagueness of our memory. Having said that, here are the names of those whom we thank for contributing in some way to this work:

Jotsna would like to thank Thomas and Laya for their understanding, support and (almost-) constant love.

Colin is indebted to Isabelle’s patience over the past months.

For example, an application developed using a set of data for one purpose will not work as well on other data for another purpose. It pays to know the limitations of artificial intelligence and correct for it; it is good to not just do AI but learn about its advantages and limitations.

Perpetuation of Stereotypes

Google Translate learns how to translate from the internet. Its ‘data miners’ scout the public web for data from which to learn. Along with language, AI learns that the number of male mechanics are more than that of female mechanics, and that the number of female nurses eclipses that of male nurses. It cannot differentiate between what is ‘true’ and what is a result of stereotyping and other prejudices. Thus, Google Translate ends up propagating what it learns, cementing stereotypes further¹:

Problems creep up in AI whenever an individual case differs from the majority case (whether this represents faithfully the majority in the real world or just the majority as represented by the internet). In classrooms, the teacher has to compensate for the system’s lapses and, where necessary, direct student’s attention to the alternative text.

Multiple Accuracy measures

AI systems make predictions on what a student should study next, whether he or she has understood a topic, what group split-up is good for a class or when a student is at risk of dropping out. Often, these predictions have a percentage tagged to them. This number tells us how good the system estimates its predictions to be.

By its nature, predictions can be erroneous. In many applications, it is acceptable to have this error. In some cases, it  is not. Moreover, the way this error is calculated is not fixed. There are different measures, and the programmer chooses what he or she thinks is the most relevant. Often, accuracy changes according to the input itself.

Since in a classroom, these systems make predictions on children, it is for the teacher to judge what is acceptable and to act where a decision taken by AI is not appropriate. To do this, a little background on AI techniques and the common errors associated with them will go a long way.

¹ Barocas, S.,  Hardt, M., Narayanan, A., Fairness and machine learning Limitations and Opportunities, MIT Press, 2023.

All businesses use data to improve their balance sheets. They use data to decide what to sell, who to sell it to, what price to fix and how to tailor their advertisements. It is the machine-learning algorithms that make sense of data. Therefore, the winner is whoever has the better data and algorithms. Data is the new gold – and the new Achilles heel.

Does data here mean just personal addresses and bank accounts?

What about the number of mouse clicks a user makes while visiting a web site?

As stewards of their own data and that of their students, it is imperative that teachers know what kind of AI data is useful, what forms do they take and how can users’ privacy be protected.

AI and the business of education

“EdTech” is the industry that makes technological applications for education – including those using artificial intelligence. These can be small companies or start-ups. They can be internet giants that are starting to pump money into education. They can also be entities with public funding.

Some EdTech software has to be bought. The rest is free, with income from other sources – often targeted ads and reselling of user data. Whatever financial model is used in EdTech AI, money can be made.

What does this mean for you and your students? Is there such a thing as a free lunch? How do we secure our classrooms while enjoying the fruits of a greedy industry?

Tool creation and you

Education does not have to change in order to accommodate technology. “Learning environments that start with technology often go down unwanted paths"¹. Any tool should be based on sound pedagogical theories. Furthermore, to be most effective, it should be co-created in teams involving teachers, pedagogical experts and computer scientists².

So, ready to start?

¹ Groff, J., Personalized Learning : The state of the field and future directions, Centre for curriculum redesign, 2017.

² Du Boulay, B., Poulovasillis, A., Holmes, W., and Mavrikis, M., Artificial Intelligence And Big Data Technologies To Close The Achievement Gap, 2018.

Step 1: Web crawlers find and download documents

After a user enters a search query, it is too late to look at all the content available on the internet¹. The web documents are looked at beforehand, and their content is broken down and stored in different slots. Once the query is available, all that needs to be done is to match what is in the query with what is in the slots.

Web crawlers are pieces of code that find and download documents from the web. They start with a set of website addresses (URLs) and look inside them for links to new web pages. Then, they download and look inside the new pages for more links. Provided the starting list was diverse enough, crawlers end up visiting every site that allows access to them, often multiple times, looking for updates.

Step 2: The document gets transformed into multiple pieces

A table is made up of smaller sections called tablets in which each row of the tablet is dedicated to one web page. These rows are arranged in some order which is recorded along with a log for updates. Each column has specific information related to the webpage which can help with matching the document content to the contents of a future query. The columns contain:

• • The website address which may, by itself, give a good description of the contents in the page, if it is a home page with representative content or a side page with associated content;
  • Titles, headings and words in bold face that describe important content;
  • Metadata of the page. This is information about the page that is not part of the main content, such as the document type (eg, email or web page), document structure and features such as document length, keywords, author names and publishing date;
  • Description of links from other pages to this page which provide succinct text regarding different aspects of the page content. More  links, more descriptions and more columns used. The presence of links is also used for ranking, to determine how popular a web page is. (Take a look at Google's Pagerank, a ranking system that uses links to and from a page to gauge quality and popularity).

• People’s names, company or organization names, locations, addresses time and date expressions, quantities and monetary values etc. Machine learning algorithms can be trained to find these entities in any content using training data annotated by a human being¹.

One column of the table, perhaps the most important, contains the main content of the document. This has to be identified amidst all the external links and advertisements. One technique uses a machine-learning model to “learn” which is the main content in any webpage.

We can of course match exact words from the query to the words in a web document, like the Find button in any word processor. But this is not very effective, as people use different words to talk about the same object. Just recording the separate words will not help to capture how these words combine with each other to create meaning. It is ultimately the thought behind the words that help us communicate, and not the words themselves. Therefore, all search engines transform the text in a way that makes it easier to match with the meaning of the query text. Later, the query is processed similarly.

As parts of a word, the total number of different tokens that need to be stored is reduced. Current models store about 30,000 to 50,000 tokens². Misspelled words can be identified because parts of them still match with the stored tokens. Unknown words may turn up search results, since their parts might match with the stored tokens.

Here, the training set for machine learning is made of example texts. Starting from individual characters, space and punctuation, the model merges characters that occur frequently, to form new tokens. If the number of tokens is not high enough, it continues the merging process to cover bigger or less frequent word-parts. This way, most of the words, word endings and all prefixes can be covered. Thus, given a new text, the machine can easily split it into tokens and send it to storage.

Step 3 : An index is built for easy reference

Once the data is tucked away in BigTables , an index is created. Similar in idea to textbook indexes, the search index lists tokens and their location in a web document. Statistics show how many times a token occurs in a document and how important is it for the document, etc, and information is positions thus – is the token in the title or a heading, is it concentrated in one part of the document and does one token always follow another?

Nowadays, many search engines use a  language-based model generated by deep neural networks. The latter encodes semantic details of the text and is responsible for better understanding of queries³. The  neural networks help the search engines to go beyond the query, in order to capture the information-need that induced the query in the first place.

These three steps give a simplified account of what is called "indexing" - finding, preparing and storing documents and creating the index. The steps involved in “ranking” come next – matching query to content and displaying the results according to relevance.

¹ Croft, B., Metzler D., Strohman, T., Search Engines, Information Retrieval in Practice, 2015

² Sennrich,R., Haddow, B., and Birch, A., Neural Machine Translation of Rare Words with Subword Units, In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715–1725, Berlin, Germany. Association for Computational Linguistics, 2016.

³ Metzler, D., Tay, Y., Bahri, D., Najork, M., Rethinking Search: Making Domain Experts out of Dilettantes, SIGIR Forum 55, 1, Article 13, June 2021.

Step 4: Query terms are matched with index terms

Once the user types the query and clicks on search, the query is processed. Tokens are created using the same process as the document text. Then the query may be expanded by adding other keywords. This is to avoid the situation whereby relevant documents are not found because the query uses words that are slightly different from those of the web-content authors. This is also done to capture differences in custom and usage. For example, the use of words such as president, prime minister and chancellor may be interchanged, depending on the country¹.

Most search engines keep track of user searches (Look at the description of popular search engines to learn more). Queries are recorded with the user data in order to personalise content and serve advertisements. Or, the records from all users are put together to see how and where to improve search engine performance.

User User logs contain items such as past queries, the results page and information on what worked. For example, what did the user click and what did they spend time reading? With user logs, each query can be matched with relevant documents (the user clicks, reads and closes session) and non-relevant documents (user did not click or did not read or tried to rephrase query)².

With these logs, each new query can be matched with a similar past query. One way of finding out if one query is similar to another, is to check if ranking turns up the same documents. Similar queries may not always contain the same words but the results are likely to be identical².

Spelling added to expand the query. This is done by looking at other words that occur frequently in relevant documents from the past. In general, however, words that occur more frequently in the relevant documents than in the non-relevant documents are added to the query or given additional weightage².

Step 5: Relevant documents are ranked

Each document is scored for relevance, and ranked according to this score. Relevance here is both topic relevance – how well the index terms of a document match that of the query, and user relevance – how well it matches the preferences of the user. A part of document scoring can be done while indexing. The speed of the search engine depends on the quality of indexes. Its effectiveness is based on how the query is matched to the document as well as on the ranking system².

User relevance is measured by creating user models (or personality types), based on their previous search terms, sites visited, email messages, the device they are using, language and geographic location.  Cookies are used to store user preferences. Some search engines buy user information from third parties as well (you could refer to descriptions of some search engines). If a person is interested in football their results for “Manchester” will be different from the person who just booked a flight to London. Words that occur frequently in the documents associated with a person will be given the highest importance.

Commercial web search engines incorporate hundreds of features in their ranking algorithms; many derived from the huge collection of user interaction data in the query logs. A ranking function combines the document, the query and user relevance features. Whatever ranking function is used, it would have a solid mathematical foundation. The output is the probability that a document satisfies the user’s information need. Above a certain probability of relevance, the document is classified as relevant².

Machinelearning is used to learn about ranking on implicit user feedback in the logs (ie, what worked in previous queries). Machine learning has also been used to develop sophisticated models of how humans use language; this is used to decipher queries^1,2.

Advances in web search have been phenomenal in the past decade. However, where it refers to understanding the context for a specific query, there is no substitute for the user providing a better query. Typically, better queries come from users who examine results and reformulate the query².

Step 6: Results are displayed

In this regard, sentences are taken from headings, metadata description or from text that best corresponds with the query. If all query terms appear in the title, they will not be repeated in the snippet². Sentences are also selected based on how readable they are.

Appropriate advertising is added to the results. Search engines generate revenue through advertisements. In some search engines, they are clearly marked as sponsored content, while in others they are not. Since many users look at only the first few results, ads can change the process substantially.

Advertisements are chosen according to the context of the query and the user model. Search engine companies maintain a database of advertisements. This database is searched to find the most relevant advertisements for a given query. Advertisers bid for keywords that describe topics associated with their product. Both the amount bid and the popularity of an advertisement are significant factors in the selection process².

For questions on facts, some engines use their own collection of facts. Google's Knowledge Vault contains over a billion facts indexed from different sources³. Results are clustered by machine learning algorithms into appropriate groups. Finally, the user is also presented with alternatives to the query to see if they are better.

Some references

The origin of Google can be found in Brin and Page’s original paper. Some of the maths behind Pagerank are on Wiki's PageRank. For the mathematical minded, here is a nice explanation of Pagerank.

¹ Russell, D., What Do You Need to Know to Use a Search Engine? Why We Still Need to Teach Research Skills, AI Magazine, 36(4), 2015.

² Croft, B., Metzler D., Strohman, T., Search Engines, Information Retrieval in Practice, 2015.

³ Spencer, S., Google Power Search: The Essential Guide to Finding Anything Online With Google, Koshkonong, Kindle Edition.

Most websites, search engines and mail clients collect information about users. Most of this data is tied to the identity of the user through IP addresses. This data is then used to serve targeted advertisements and personalised content, improve the services provided and carry outmarket research. However, search engines do not always disclose all the information they collect where they collect it and what they do with that information¹. Or even where they collect this information. For example, studies show that Google can track users across nearly 80% of websites².

Information that search engines can display when someone searches for a user include:

• Information that they added in some web site,
• Information added by others with their consent,
• Information that was collected in some other context and then published on the web - by forums, event organisers, friends and others.

Information collected and processed when search engines are used include the following:

• The searched-for topic, date and time of search^1,3,4.
• Activity data across apps such as email, calendar and maps, collected by search engines like Google and Microsoft^3,4.
• Data bought by some search engines from third parties^3,4.
• Data bought from search engines and websites that are put together and tied to the user by third parties².
• Inferences made from the data collected.
• Inferences drawn from personal settings. For example, "to infer that a user who has strong privacy settings may have certain psychological traits, or that they may have “something to hide”"⁵.
• User profiles or models, created by search engines, based on this information. These models are based on online data and give only a limited view of the person. Decisions based on these, when used in other contexts, will not be balanced.

Collected data on a consent-giving user can be used to draw inferences about another user who did not give consent, but who has been judged by the search engine to have a similar profile.

All this data, both raw and processed, gives rise to privacy and security concerns. Some measures can be taken by search providers, governments and users to prevent privacy breeches:

• Data can be stored in such a way that leaks and theft are discouraged. For example, user data can be stored in separate and decentralised databases⁵;
• Data is encrypted or anonymised;
• Machine learning can be used to automatically detect and classify trackers. This can then be used to improve browser privacy tools²;
• Policies and laws like GDPR legislation can introduce explicit guidelines and sanctions to regulate data collection, use and storage¹;
• User-centred recommendations are made and publicised so that users, including parents and teachers, can better protect their and their wards’ privacy.

In Europe, search engine companies are viewed as ‘controllers of personal data’, as opposed to mere providers of a service. Thus, they can be held responsible and liable for the content that is accessible through their services. However, privacy laws often concern confidential and intimate data. Even harmless information about people can be mined in order to create user profiles based on implicit patterns in the data. Those profiles (whether accurate or not) can be used to make decisions affecting them¹.

Also, how a law is enforced changes from country to country. According to GDPR, a person can ask a search engine company to remove a search result that concerns them. Even if the company removes it from the index in Europe, the page can still show up in results outside Europe¹.

Although companies’ policies can shed light on their practices, research shows that there is often a gap between policy and its use².

Reliability of content

Critics have pointed out that search engine companies are not fully open with why they show some sites and not others, and rank some pages higher than others¹.

Ranking of search results is heavily influenced by advertisers who sponsor content. Moreover, big search engine companies provide many services other than search. Content provided by them are often boosted in the search results. In Europe, Google has been formally charged with prominently displaying its own products or services in its search returns, regardless of its merits¹.

Large companies and web developers who study ranking algorithms can also influence ranking by playing on how a search engine defines popularity and authenticity of websites. Of course, the criteria judged important by search-engine programmers are themselves open to question.

This affect how reliable the search results are. It is always good to use multiple sources and multiple search engines and have a discussion about the content used in schoolwork.

Autonomy

A search engine recommends content using its ranking system. By not revealing the criteria used to select this content, it reduces user autonomy. For example, if we had known a suggested web page was sponsored, or selected based on popularity criteria we don’t identify with, we might not have chosen it. By taking away informed consent, search engines and other recommender systems have controlling influences over our behaviour.

Autonomy is having control over processes, decisions and outcomes⁷. It implies liberty (independence from controlling influences) and agency (capacity for intentional action)⁷. Systems that recommend content without explanation can encroach on users’ autonomy. They provide recommendations that nudge the users in a particular direction, by engaging them only with what they would like and by limiting the range of options to which they are exposed⁵.

¹ Tavani, H., Zimmer, M., Search Engines and Ethics, The Stanford Encyclopedia of Philosophy, Fall 2020 Edition), Edward N. Zalta ed.

² Englehardt, S., Narayanan, A., Online Tracking: A 1-million-site Measurement and Analysis, Extended version of paper at ACM CCS 2016.

³ Google Privacy and Terms.

⁴ Microsoft Privacy Statement.

⁵ Milano, S., Taddeo, M., Floridi, L. Recommender systems and their ethical challenges, AI & Soc 35, 957–967, 2020.

⁶ Tavani, H.T., Ethics and Technology: Controversies, Questions, and Strategies for Ethical Computing, 5th edition, Hoboken, NJ: John Wiley and Sons, 2016.

⁷ Hillis, K., Petit, M., Jarrett, K., Google and the Culture of Search, Routledge Taylor and Francis, 2013.

More and more, there is a feeling that everything that matters is on the web and should be accessible through search¹. As LM Hinman puts it, “Esse est indicato in Google (to be is to be indexed on Google).” As he also notes, “citizens in a democracy cannot make informed decisions without access to accurate information”^2,3. If democracy stands on free access to undistorted information, search engines directly affect how democratic our countries are. Their role as gatekeepers of knowledge is in direct conflict with their nature as private companies dependent on ads for income. Therefore, for the sake of a free society, we must demand accountability for search engines and transparency in how their algorithms work².

Creation of filter bubbles

Systems that recommend content based on user profiles, including search engines, can insulate users from exposure to different views. By feeding content that the user likes, they create self-reinforcing biases and “filter bubbles”^2,4. These bubbles, created when newly acquired knowledge is based on past interests and activities⁵, cement biases as solid foundations of knowledge. This could become particularly dangerous when used with young and impressionable minds. Thus, open discussions with peers and teachers and collaborative learning activities should be promoted in the classroom.

Feedback loops

Search engines, like other recommendation systems, predict what will be of interest to the user. Then, when the user clicks on what was recommended, it  (the search engine) takes it as positive feedback. This feedback affects what links are displayed in the future. If a user clicked on the first link displayed, was it because they found it relevant or simply because it was the first result and thus easier to choose?

Implicit feedback is tricky to interpret. When predictions are based on incorrect interpretation, the effects are even trickier to predict. When certain results are repeatedly shown – and are the only thing that the user gets to see – it can even end up changing what the user likes and dislikes – a self-fulfilling prediction, perhaps.

In the United States, a predictive policing system was launched whereby high-crime areas of a certain city were highlighted. This meant that more police officers were deployed to such areas. Since these officers knew the area was at high risk, they were careful, and stopped, searched, or arrested more people than normal. The arrests thus validated the prediction, even where the prediction was biased in the first place. Not only that, the arrests were data for future predictions on the same areas and on areas similar to it, compounding biases over time⁵.

We use prediction systems in order to act on the predictions. But acting on biased predictions affects future outcomes, the people involved – and ultimately society itself. “As a side-effect of fulfilling its purpose of retrieving relevant information, a search engine will necessarily change the very thing that it aims to measure, sort and rank. Similarly, most machine-learning systems will affect the phenomena that they predict"⁵.

Fake news, extreme content and censorship

There is increasing prevalence of fake news (false stories that appear as news) in online forums, social media sites and blogs, all available to students through search. Small focused groups of people can drive ratings up for specific videos and web sites of extreme content. This increases the content’s popularity and appearance of authenticity, gaming the ranking algorithms⁴. Yet, as of now, no clear and explicit policy has been adopted by search-engine companies to control fake news².

On the other hand, search engines systematically exclude certain sites and certain types of sites in favour of others⁶. They censor content from some authors, despite not being asked to do so by the public. Search engines, therefore, should be used with awareness, discretion and discrimination.

¹ Hillis, K., Petit, M., Jarrett, K., Google and the Culture of Search, Routledge Taylor and Francis, 2013.

² Tavani, H., Zimmer, M., Search Engines and Ethics, The Stanford Encyclopedia of Philosophy, Fall 2020 Edition), Edward N. Zalta (ed.).

³ Hinman, L. M., Esse Est Indicato in Google: Ethical and Political Issues in Search Engines, International Review of Information Ethics, 3: 19–25, 2005.

⁴ Milano, S., Taddeo, M., Floridi, L. Recommender systems and their ethical challenges, AI & Soc 35, 957–967, 2020.

⁵ Barocas, S.,  Hardt, M., Narayanan, A., Fairness and machine learning Limitations and Opportunities, MIT Press, 2023.

⁶ Introna, L. and Nissenbaum, H., Shaping the Web: Why The Politics of Search Engines Matters, The Information Society, 16(3): 169–185, 2000.

Thenumber of people making use of e-learning is constantly growing. The term e-learning refers to learning mediated by the use of technology in contexts where educators and learners are distant in space and/or time. The ultimate aim of e-learning is to improve students’ learning experience and practice.

Today,with the advancement of technology, it is more appropriate to refer to systems and platforms for the ’delivery’ of e-learning rather than to single -purpose tools. Such systems are the result of integrating different software tools capable of building an ecosystem where flexible and adaptable learning paths can be exploited. An e-learning system enables the management of learning processes and the management of courses. It enables student-learning assessments, constructing reports, and creating and organising content. It facilitates communication between teachers/tutors and students. Among the most widely used e-learning systems, there are Learning Management Systems (LMS) (eg Moodle, Edmodo).

The acronym LMS refers to a web-based application designed to manage the learning process of trainees¹ at different levels, in different ways and domains. An LMS, therefore, could be defined as a learning environment within which learning, content and assessment activities and tools are implemented. Student-student and/or student-educator interactions likewise are implemented and managed within this environment. Furthermore, the definition of LMSs includes their being platforms that generally can include whole course management systems, content management systems, and portals².

LMS and AI: the Smart LMS

With the advent of AI, Education, in general, and LMSs, in particular, become possible and promising fields of application of this revolutionary force³. Specifically, LMSs, thanks to the functionalities supported by AI, represent a renewed learning tool capable of satisfying two of the fundamental traits of the education of the future: personalisation and adaptation⁴. It is from this combination of LMS and AI that the Smart LMS (SLMS) or Intelligent LMS emerges.

An efficient SLMS features algorithms that can provide and retrieve information from three fundamental clusters of knowledge: a) the learner b) the pedagogy and c) the domain. By  acquiring information about (a) learners’ preferences, their emotional and cognitive states and their achievements and goals, an SLMS can implement those teaching strategies (b) that are most effective (specific types of assessment, collaborative learning, etc.) for learning to be most fruitful within the specific domain of knowledge being studied (c): eg geometry theorems, mathematical operations, laws of physics, text analysis procedures⁴.

An SLMS, therefore, can be defined as a learning system capable of adapting the contents proposed to the learner by calibrating them to the knowledge and skills the learner has displayed in previous tasks. In fact, by adopting a learner-centred approach, it can identify, follow and monitor learners’ paths by recording their learning patterns and styles. Referring to the description given by Fardinpour et al.⁵, an intelligent LMS provides the learner with the most effective learning path and the most appropriate learning content, through automation, the adaptation of different teaching strategies (scaffolding), and the reporting and knowledge generation. It also provides the learners with the possibility to keep track of and monitor their learning and learning goals. Furthermore, although these features and tools enable the LMS to operate more intelligently, an SLMS must provide learners with the possibility to disable the AI that manages their path, in order to have full access to all learning materials in the learning environment.

Some examples of AI-supported functionalities in the context of an SLMS

When an SLMS is functioning correctly, several AI-supported tools make it possible to realise a system with the features described above. Such AI-supported tools move transversally along the three aforementioned clusters of knowledge, to which the SLMS algorithms constantly refer (learner, pedagogy, domain).

AI-supported chatbots as virtual tutors

A chatbot is software that simulates and processes human conversations (written or spoken). In the context of an SLMS, it can function as a virtual tutor, capable of answering a learner’s questions concerning, for example, learning courses. The chatbot is also capable of providing suggestions to the learner, based on the analysis that the system makes of the learner’s previous performances and interactions⁶.

Learning Analytics

Benefits for learners and teachers

These and other AI-supported tools⁴ contribute to making an SLMS a powerful learning and teaching tool that, instead of being perceived as a substitute for the teacher's work, shows itself as a tool capable of "augmenting" the human aspects of teaching⁸ and bringing a series of fundamental benefits to the whole learning/teaching process.

Since an SLMS calibrates the contents to the student’s skills and level, it avoids the learner, in the different phases of his or her path, facing tasks that bore him or her because they are too simple, or that frustrate him or her because they are too complex. This ensures that the student’s motivation and attention are always at a high level and appropriate to the level of difficulty of the task to be addressed. This situation has the direct consequence of significantly reducing the dropout rate, as it allows teachers to detect any problems in time and intervene promptly, as soon as the student shows the first signs of difficulty.

Such a situation, as well as linear learning situations (without difficulties), can be addressed by proposing to the students, through the SLMS tools, different knowledge contents that are already stored in the course databases or are from third-party providers. This results in a direct benefit for the teacher, who does not have to create new teaching materials from time to time, and can use the saved time in other essential occupations such as refining their teaching methods and/or interacting directly with the students.

¹ Kasim, N. N. M., and Khalid, F., Choosing the right learning management system (LMS) for the higher education institution context: A systematic review, International Journal of Emerging Technologies in Learning, 11(6), 2016.

² Coates, H., James, R., & Baldwin, G., A critical examination of the effects of learning management systems on university teaching and learning, Tertiary education and management, 11(1), 19-36, 2005.

³ Beck, J., Sternm, M., & Haugsjaa, E., Applications of AI in Education, Crossroads, 3(1), 11–15. doi:10.1145/332148.332153, 1996.

⁴ Rerhaye, L., Altun, D., Krauss, C., & Müller, C., Evaluation Methods for an AI-Supported Learning Management System: Quantifying and Qualifying Added Values for Teaching and Learning, International Conference on Human-Computer Interaction (pp. 394-411). Springer, Cham, July 2021.

⁵ Fardinpour, A., Pedram, M. M., & Burkle, M., Intelligent learning management systems: Definition, features and measurement of intelligence, International Journal of Distance Education Technologies (IJDET), 12(4), 19-31, 2014.

⁶ HR Technologist: Emerging Trends for AI in Learning Management Systems, 2019, Accessed 31 Oct 2022.

⁷ Krauss, C., Salzmann, A., & Merceron, A., Branched Learning Paths for the Recommendation of Personalized Sequences of Course Items, DeLFI Workshops, September 2018.

⁸ Mavrikis, M., & Holmes, W., Intelligent learning environments: Design, usage and analytics for future schools, Shaping future schools with digital technology, 57-73, 2019.

More and more organizations are using data analysis to solve problems and improve decisions related to their activities. And the world of education is not an exception because, with the generalization of virtual learning environment (VLE) and learning management systems (LMS), massive learning data are now available, generated by the learners’ interaction with these tools.

Learning Analytics (LA) is a disciplinary field defined as “the measurement, collection, analysis and reporting of data about learners and their contexts, for purposes of understanding and optimising learning and the environments in which it occurs”⁴.

Four types of analytics are generally distinguished according to the question to solve:

• Descriptive Analytics: What happened in the past?
• Diagnostic Analytics: Why something happened in the past?
• Predictive Analytics: What is most likely to happen in the future?
• Prescriptive Analytics What actions should be taken to affect those outcomes?

What is it?

The visualisation to recommendation systems. Research in this area is currently active. We will limit ourselves to summarising the frequent issues encountered in the literature. Each issue leads to families of tools targeting mainly learners or teachers who represent most of the end users of LA-based applications.

Predict and enhance students learning outcome

One of the emblematic applications of LA is the prediction of failures.

Learning indicators are automatically computed from the digital traces and can be accessed directly by learners so that they can adapt their own learning strategies.

One of the first experiments was conducted at Purdue University (USA) with a mobile application designed as a traffic-light-based dashboard¹.

Each student can monitor his or her own progress indicators.

A screenshot of the dashboard is shown in figure 1.

Indicators can also be addressed to teachers, as in an early warning system (EWS).

This is the choice made by the French national centre for Distance Education (CNED) in an ongoing study².

The objective of an EWS is to alert, as soon as possible, those tutors responsible for monitoring the students, in order to implement appropriate remedial actions.

Analyze student learning process

LA techniques can help to model the learning behaviour of a learner or a group of learners (ie a class). The model can be used to display learning processes in LA applications, providing additional information that will enable teachers to detect deficiencies which will help improve training materials and methods. Moreover, the analysis of learning process is a way for observing the learner engagement. For example, in the e-FRAN METAL project, the indicators were brought together in a dashboard co-designed with a team of secondary-school teachers as shown in figure 2³.

Personalise learning paths

The personalisation of learning paths can occur in recommendation or adaptive learning systems. Recommendation  systems aim to suggest, to each learner, best resources or appropriate behaviour that may help to achieve educational objectives.

Some systems focus on putting the teacher in the loop by first presenting proposed recommendations for their validation. Adaptive learning systems allow the learner to develop skills and knowledge in a more personalised and self-paced way by constantly adjusting the learning path towards the learner experience.

Does it work?

In publications, feedback focuses mainly on students, including those in higher education. Observations generally show improved performance of learners (for example, +10% of grades A and B at Purdue University, US). For teachers, the impact of LA is more complex to assess. Studies based on the technology acceptance model (TAM) suggest that teachers have a positive perception of the use of LA tools. A study shows the final Strengths, Weaknesses, Opportunities and Threats analysis (SWOT) that we reproduce here⁵ (see figure 3).

Some points of attention, included in the threats and weakness sections, form the basis of the reflection of the Society for Learning Analytics Research (SoLAR) community to recommend an ethics-by-design approach for LA applications (Drashler-16). The recommendations are summarised in a checklist of eight key words: Determine, Explain, Legitimate, Involve, Consent, Anonymise, Technical, External (DELICATE).

¹ Arnold, K. and Pistilli, M., Course signals at Purdue: Using learning analytics to increase student success, LAK2012, ACM International Conference Proceeding Series, 2012.

² Ben Soussia, A., Roussanaly, A., Boyer, A., Toward An Early Risk Alert In A Distance Learning Context, ICALT, 2022.

³ Brun, A., Bonnin, G., Castagnos, S., Roussanaly, A., Boyer, A., Learning Analytics Made in France: The METALproject, IJILT, 2019.

⁴ Long, P., and Siemens, G., 1st International Conference on Learning Analytics and Knowledge, Banff, Alberta, February 27–March 1, 2011.

⁵ Mavroudi, A., Teachers’ Views Regarding Learning Analytics Usage Based on the Technology Acceptance Model, TechTrends. 65, 2021.

Educational systems have always generated data – students’ personal data, academic records, attendance data and more. With digitalisation and AIED applications, more data is recorded and stored: mouse clicks, opened pages, timestamps and keyboard strokes¹. With data-centric thinking becoming the norm in society, it is natural to ask how to crunch all this data to do something pertinent. Could we give more personalised feedback to the learner? Could we design better visualisation and notification tools for the teacher?²

Whatever technology is used, it has to meet a real requirement in the classroom. After the need is identified, we can look at the data available and ask what is relevant to a desired outcome. This involves uncovering factors that let educators make nuanced decisions. Can these factors be captured using available data? Is data and data-based systems the best way of addressing the need? What could be the unintended consequences of using data this way³?

Machine learning lets us defer many of these questions to the data itself⁴. ML applications are trained on data. They work by operating on data. They find patterns and make generalisations and store these as models – data that can be used to answer future questions⁴. Their decisions and predictions, and how these affect student learning, are all data too. Thus, knowing how programmers, the machine and the user handle data is an important part of understanding how artificial intelligence works.

About data

Data is generally about a real world entity - a person, an object, or an event. Each entity can be described by a number of attributes (features or variables)⁵. For example, name, age and class are some attributes of a student. The set of these attributes is the data we have on the student, which, while not in any way close to the real entity, does tell us something about them. Data collected, used and processed in the educational system is called educational data¹.

A dataset is the data on a collection of entities arranged in rows and columns. The attendance record of a class is a dataset. In this case, each row is the record of one student. The columns could be their presence or absence during a particular day or session. Thus each column is an attribute.

Data is created by choosing attributes and measuring them: every piece of data is the result of human decisions and choices. Thus, data creation is a subjective, partial and sometimes messy process prone to technical difficulties^4,5. Further, what we choose to measure, and what we don't can have a big influence on expected outcomes.

Data traces are records of student activity such as mouse clicks, data on opened pages, the timing of interactions or key presses in a digital system¹. Metadata is data that describe other data⁵. Derived data is data calculated or inferred from other data: individual scores of each student is data. The class average is derived data. Often, derived data is more useful in getting useful insights, finding patterns and making predictions. Machine Learning applications can create derived data and link it with metadata data traces to create detailed learner models, which help in personalising learning¹.

For any data based application to be successful, attributes should be carefully chosen and correctly measured. The patterns discovered in them should be checked to see if they make sense in the educational context. When designed and maintained correctly, data driven systems can be very valuable.

This chapter aims to introduce a few basics of data and data based technology but data literacy is a very important skill to possess and merits dedicated training and continuing support and update¹.

Legislation you should know about

Because of the drastic drop in costs of data storage, more data and metadata are saved and retained for a longer time⁶. This can lead to privacy breaches and rights violations. Laws like the General Data Protection Regulation (GDPR) discourages such practices and gives EU citizens more control over their personal data. They give legally enforceable data protection regulations across all EU member states.

According to GDPR, personal data is any information relating to an identified or identifiable person (data subject). Schools, in addition to engaging with companies that handle their data, store huge  amounts of personal information about students, parents, staff, management, and suppliers. As data controllers, they are required to store data which they process confidentially and securely and have procedures in place for the protection and proper use of all personal data¹.

Rights established by the GDPR include:

• The Right to Access makes it mandatory for them to know(easily) what data is being collected about them
• The citizen’s Right to Be Informed of the usage made of their data
• The Right to Erasure allows a citizen whose data has been collected by a platform to ask for that data to be removed from the dataset built by the platform (and which may be sold to others)
• The  Right to explanation - explanation should be provided whenever clarification is needed on an automated decision process that affect them

Although, GDPR does allow for collection of some data under “legitimate interest”⁷and the use of derived, aggregated, or anonymized data indefinitely and without consent⁵. The new Digital Services Act restricts the use of personal data for targeted advertising purposes⁷." In addition to these, the EU-US Privacy Shield strengthens the data-protection rights for EU citizens in the context where their data have been moved outside of the EU⁵.

Please refer to GDPR for dummies for the analysis carried out by independent experts from the Civil Liberties Union for Europe (Liberties). This is a watchdog that safeguards EU citizens’ human rights.

¹ Ethical guidelines on the use of artificial intelligence and data in teaching and learning for educators, European Commission, October 2022.

² du Boulay, B., Poulovasillis, A., Holmes, W., Mavrikis, M., Artificial Intelligence And Big Data Technologies To Close The Achievement Gap, in Luckin, R., ed. Enhancing Learning and Teaching with Technology, London: UCL Institute of Education Press, pp. 256–285, 2018.

³ Hutchinson, B., Smart, A., Hanna, A., Denton, E., Greer, C., Kjartansson, O., Barnes, P., Mitchell, M., Towards Accountability for Machine Learning Datasets: Practices from Software Engineering and Infrastructure, Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, Association for Computing Machinery, New York, 2021.

⁴ Barocas, S.,  Hardt, M., Narayanan, A., Fairness and machine learning Limitations and Opportunities, MIT Press, 2023.

⁵ Kelleher, J.D, Tierney, B, Data Science, MIT Press, London, 2018.

⁶ Schneier, B., Data and Goliath: The Hidden Battles to Capture Your Data and Control Your World, W. W. Norton & Company, 2015.

⁷ Kant, T., Identity, Advertising, and Algorithmic Targeting: Or How (Not) to Target Your “Ideal User.”, MIT Case Studies in Social and Ethical Responsibilities of Computing, 2021.

To be effective in the classroom, multi-disciplinary teams with teachers, pedagogical experts and computer scientists should be involved in each step of the process process¹. Human experts are needed to identify the need and design the process, to design and prepare the data, select ML algorithms, to critically interpret the results and to plan how to use the application².

1) Understanding the educational context

The first step in the design of an AIED tool is to understand the needs in the classroom. Once the goals are set, it has to be seen how they can be achieved: what factors to consider and what to ignore. Any data based solution is biased towards phenomena that can be easily calculated and standardised³. Thus, each decision has to be discussed by teachers who will use the tool, by experts in pedagogy who can ensure that all decisions are grounded in proven theory, and by computer scientists who understand how the algorithms work.

There is a lot of back and forth between the two first steps since what is possible will also depend on what data is available². Moreover, the design of educational tools is also subject to laws which place restrictions on data usage and types of algorithms that can be used.

2) Understanding the data

Once the goals and contributing factors are identified, the focus shifts to what data is required, how it will be sourced and labelled, how privacy will be handled and how data quality will be measured³. For a machine learning application to be successful, the datasets have to be large enough, diverse and well-labelled.

Machine learning requires data to train the model and data to work on or predict with. Some ML tasks like face recognition and object recognition already have a lot of private and public databases available for training.

If not already available in a usable learning requires data to train the model and data to work on or use for predictions. Some ML tasks, such as face and object recognition, already have a lot of private and public databases available for training.

If not already available in a usable form, existing datasets may have to be added to or relabelled to fit the needs of the project. If not, dedicated datasets may have to be created and labelled from scratch. Digital traces generated by the student while using an application could also be used as one of the data sources.

In any case, data and features relevant to the problem have to be carefully identified². Irrelevant or redundant features can push an algorithm into finding false patterns and affect the performance of the system². Since the machine can find patterns only in the data given to it, choosing the dataset also implicitly defines what the problem is⁴. If there is a lot of data available, a subset has to be selected with the help of statistical techniques and the data verified to avoid errors and biases.

As an example of a bad training data, in a story from early days of computer vision, a model was trained to discriminate between images of Russian and American tanks. Its high accuracy was later found to be due to the fact the the Russian tanks had been photographed on a cloudy day and the American ones on a sunny day⁴.

Therefore, the chosen dataset has to be verified for quality, taking into account why it was created, what does it contain, what are the processes used for collecting, cleaning and labelling, distribution and maintenance⁴. The key questions to ask include Are the datasets fit for their intended purposes and Do the datasets contain hidden hazards that can make models biased or discriminatory?³

3) Preparing the data

Data-preparation involves creating data sets by merging data available from different places, adjusting for inconsistencies (For example, some test scores could be on a scale of 1 to 10 while others are given as a percentage) and looking for missing or extreme values. Then, automated tests could be performed to check the quality of datasets. This includes checking for privacy leaks and unforeseen correlations or stereotypes². The datasets might also be split into training and test datasets at this stage. The former is used for training the model and the latter for checking its performance. Testing on the training dataset would be like giving out the exam paper the day before for homework: the student's exam performance will not indicate their understanding².

4) Modelling

In this step, algorithms are used to extract patterns in the data and create models. Usually, different algorithms are tested to see what works best. These models can then be put into use to make predictions on new data.

In most projects, initial models uncover problems in the data calling for back and forth between steps 2 and 3². As long as there is a strong correlation between the features of the data and the output value, it is likely that a Machine Learning algorithm will generate good predictions.

These algorithms use advanced statistical and computing techniques to process data. The programmers have to adjust the settings and try different algorithms to get the best results. Let us take an application that detects cheating. A false positive is when a student who did not cheat is flagged. A false negative is when a student who cheats is not flagged. System designers can tune the model to minimize either false positives, where some cheating behaviour could be missed, or false negatives, where even doubtful cases are flagged⁵. The tuning thus depends on what we want the system to do.

5) Evaluation

During the modelling stage, each model can be tuned for accuracy of prediction on the training dataset. Then models are tested on the test dataset and one model is chosen for use. This model is also evaluated on how it meets the educational needs: Are the objectives as set out in step 1 met? Are there any unforeseen problems? Is the quality good? Could anything be improved or done in another way? Is a redesign needed? The main objective is to decide whether the application can be deployed in schools. If not, the whole process is repeated².

6) Deployment

The final step of this process is to see how to integrate the data based application with the school system to get maximum benefits, both with respect to the technical infrastructure and teaching practices.

Though given as the final step, the whole process is iterative. After deployment, the model should be regularly reviewed to check if it is still relevant to the context. The needs, processes or modes of data capture could change affecting the output of the system. Thus, the application should be reviewed and updated when necessary. The system should be monitored continuously for its impact on learning, teaching and assessment⁶.

The Ethical guidelines on the use of AI and data for educators stresses that the school should be in contact with the AI service provider throughout the lifecycle of the AI system, even prior to deployment. It should ask for clear technical documentation and seek clarification on unclear points. An agreement should be made for support and maintenance, and it should be made sure that the supplier adhered to all legal obligations⁶.

Note: Both the steps listed here and the illustration are adapted from the CRISP-DM Data science stages and tasks (based on figure 3 in Chapman, Clinton, Kerber, et al. 1999) as described in ².

¹ Du Boulay, B., Poulovasillis, A., Holmes, W., Mavrikis, M., Artificial Intelligence And Big Data Technologies To Close The Achievement Gap,in Luckin, R., ed. Enhancing Learning and Teaching with Technology, London: UCL Institute of Education Press, pp. 256–285, 2018.

² Kelleher, J.D, Tierney, B, Data Science, London, 2018.

³ Hutchinson, B., Smart, A., Hanna, A., Denton, E., Greer, C., Kjartansson, O., Barnes, P., Mitchell, M., Towards Accountability for Machine Learning Datasets: Practices from Software Engineering and Infrastructure, Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, Association for Computing Machinery, New York, 2021.

⁴ Barocas, S.,  Hardt, M., Narayanan, A., Fairness and machine learning Limitations and Opportunities, MIT Press, 2023.

⁵ Schneier, B., Data and Goliath: The Hidden Battles to Capture Your Data and Control Your World, W. W. Norton & Company, 2015.

⁶ Ethical guidelines on the use of artificial intelligence and data in teaching and learning for educators, European Commission, October 2022.

This activity has been adapted from Identity, Advertising, and Algorithmic Targeting: Or How (Not) to Target Your “Ideal User.”,  licenced under CC BY NC 4.0. ¹

The labels Google gives us - male, female, young or aged, have nothing to do with our identities, needs or values. Someone can be male if they look at certain websites(say hardware stores) and buy certain items². Tomorrow, a male can become female if their activity or the activities of a million other humans who contributed to what is male-like-behaviour change. Different companies give us completely different identities based on what interests them.

The same is done to our students when they interact with Personalized learning software and are subjected to learning analytics. Their digital identity, their performance, engagement and satisfaction, as viewed by these systems, is then used to evaluate not only their performance but also the performance of their peers, teachers, schools and the educational system itself³.

Why is this a problem?

1. These profiles are often composed based on noisy and incorrect data from various sources and can be very misleading⁴.
2. These digital identities can change how a student sees themselves and others, how teachers see each student, how the system sees each teacher, how the society sees education and pedagogy, and how everyone reacts to decisions and feedback³.
3. Yet, these judgements on who someone is are made without their knowledge and consent - by black boxes no one has access to. Often, there is no control over what data is recorded, where and when it is recorded and how decisions are made based on it^4,1.  Students and teachers lose their expressive power and human agency.
4. This data and judgements tend to persist as stored data long after the recorded event took place⁴.
5. The stress on metrics, where students, teachers and staff are constantly assessed, compared and ranked, can induce reactions such as  anxiety and competition instead of motivation and growth³.
6. Aspects of education that can be automatically captured and analysed are given more importance; they push us towards outcomes and practices that are different from what might otherwise be of concern to us.
7. The organisations that do the “datafication” have the power to define “what ‘counts’ as quality education – a good student or an effective teacher³.”

Here are countermeasures the experts suggest teachers take:

1. Consider the people, their identity, integrity, and dignity: “Approach people with respect of their intrinsic value and not as a data object or a means-to-an-end”⁵. People are not just the data; that the label that a software might give students to personalise learning pathways or to split them into groups is not their real identity⁵.
2. Be data literate: Learn how to handle data correctly. Learn what different data based systems do, how they do it, what is their recommended usage and how to interpret the information they generate and the decisions they make.
3. Maintain a critical distance from AIED companies and software. Question their claims, ask proof of their validity and reliability, verify that the system follows ethical guidelines of your institution and country³.
4. Monitor the effects these systems have on you, your students, their learning and the classroom atmosphere.
5. Call for open systems that give you control and the power to override automated decisions. Pitch in, clarify or override wherever and whenever you feel the need.

¹ Kant, T., Identity, Advertising, and Algorithmic Targeting: Or How (Not) to Target Your “Ideal User.” MIT Case Studies in Social and Ethical Responsibilities of Computing, 2021.

² Cheney-Lippold, J., We Are Data: Algorithms and the Making of Our Digital Selves, NYU Press, 2017.

³ Williamson, B., Bayne, S., Shay, S., The datafication of teaching in Higher Education: critical issues and perspectives, Teaching in Higher Education, 25:4, 351-365, 2020.

⁴ Kelleher, J.D, Tierney, B, Data Science, MIT Press, London, 2018.

⁵ Ethical guidelines on the use of artificial intelligence and data in teaching and learning for educators, European Commission, October 2022.

Examples of bias entering AIED systems

1. When programmers code rule-based systems, they can put their personal biases and stereotypes into the system¹.
2. A data based algorithm can conclude not to propose a STEM-based career path for girls because female students feature less in the STEM graduate dataset. Is the lesser number of female mathematicians due to existing stereotypes and societal norms or is it due some inherent property of being female? Algorithms have no way to distinguish between the two situations. Since existing data reflects existing stereotypes, the algorithms that train on them replicate existing inequalities and social dynamics⁴. Further, if such recommendations are implemented, more girls will opt for non-STEM subjects and the new data will reflect this - a case of self-fulfilling prophecy³.
3. Students from a culture that is under-represented in the training dataset might have different behaviour patterns and different ways of showing motivation. How would a learning analytics calculate metrics for them? If the data is not representative of all categories of students, systems trained on this data might penalise the minority whose behavioural tendencies are not what the program was optimised to reward. If we’re not careful, learning algorithms will generalise, based on the majority culture, leading to a high error rate for minority groups^4,5. Such decisions might discourage those who could bring diversity, creativity and unique talents and those who have different experiences, interests and motivations².
4. A British student judged by a US essay correction software would be penalised for their spelling mistakes. Local language, changes in spelling and accent, local geography and culture would always be tricky for systems that are designed and trained for another country and another context.
5. Some teachers penalise phrases common to a class or region, either consciously or because of biased social associations. If an essay-grading software trains on essays graded by these teachers, it will replicate the same bias.
6. Machine learning systems need a target variable and proxies for which to optimise. Let us say high-school test scores were taken as the proxy for academic excellence. The system will now train exclusively to boost patterns that are consistent with students who do well under the stress and narrowed contexts of exam halls. Such systems will boost test scores, and not knowledge, when recommending resources and practice exercises to students. While this might also be true in many of today’s classrooms, the traditional approach at least makes it possible to express multiple goals⁴.
7. Adaptive learning systems suggest resources to students that will remedy a lack of skill or knowledge. If these resources need to be bought or require home internet connection, then it is not fair to those students who don’t have the means to follow the recommendations: “When an algorithm suggests hints, next steps, or resources to a student, we have to check whether the help-giving is unfair because one group systematically does not get useful help which is discriminatory”².
8. The concept of personalising education according to a student’s current knowledge level and tastes might in itself constitute a bias¹. Aren’t we also stopping this student from exploring new interests and alternatives? Wouldn’t this make him or her one-dimensional and reduce overall skills,knowledge and access to opportunities?

What can the teacher do to reduce the effects of AIED biases?

Researchers are constantly proposing and analysing different ways to reduce bias. But not all methods are easy to implement – fairness goes deeper than mitigating bias.

For example, if existing data is full of stereotypes – “do we have an obligation to question the data and to design our systems to conform to some notion of equitable behaviour, regardless of whether or not that’s supported by the data currently available to us?”⁴. Methods are always in tension and opposition with each other and some interventions to reduce one kind of bias can introduce another bias!

So, what can the teacher do?

1. Question the seller – before subscribing to an AIED system, ask what type of datasets were used to train it, where, by and for whom was it conceived and designed, and how was it evaluated.
2. Don’t swallow the metrics when you invest in an AIED system. An overall accuracy of, say, five percent, might hide the fact that a model performs badly for a minority group⁴.
3. Look at the documentation – what measures, if any, have been taken to detect and counter bias and enforce fairness¹?
4. Find out about the developers – are they solely computer science experts or were educational researchers and teachers involved? Is the system based solely on machine learning or were learning theory and practices integrated²?
5. Give preference to transparent and open learner models which allow you to override decisions², many AIED models have flexible designs whereby the teacher and student can monitor the system, ask for explanations or ignore the machine’s decisions.
6. Examine the product’s accessibility. Can everyone access it equally, regardless of ability¹?
7. Watch out for the effects of using a technology, both long-term and short-term, on students, and be ready to offer assistance when necessary.

Despite the problems of AI-based technology, we can be optimistic about the future of AIED:

• With increased awareness of these topics, methods to detect and correct bias are being researched and trialled;
• Rule-based and data-based systems can uncover hidden biases in existing educational practices. Exposed thus, these biases can then be dealt with;
• With the potential for customisation in AI systems, many aspects of education could be tailored. Resources could become more responsive to students’ knowledge and experience. Perhaps they could integrate local communities and cultural assets, and meet specific local needs².

¹ Ethical guidelines on the use of artificial intelligence and data in teaching and learning for educators, European Commission, October 2022.

² U.S. Department of Education, Office of Educational Technology, Artificial Intelligence and Future of Teaching and Learning: Insights and Recommendations, Washington, DC, 2023.

³ Kelleher, J.D, Tierney, B, Data Science, MIT Press, London, 2018.

⁴ Barocas, S.,  Hardt, M., Narayanan, A., Fairness and machine learning Limitations and Opportunities, MIT Press, 2023.

⁵ Milano, S., Taddeo, M., Floridi, L., Recommender systems and their ethical challenges, AI & Soc 35, 957–967, 2020.

Every teacher personalises learning – if only by adding an additional example or giving individual attention where called for. In a sense, teaching itself is an act of personalisation, as opposed to, say, a televised lesson. Teachers change their lessons so students can make sense of what they learn. They help students to fit the new knowledge or skill with what they already know, their personal observances and social experiences. They help learners make what they can out of what they learn.

Personalised learning is about creating different learning environments and experiences for students’ needs, capabilities and cultural contexts¹. Of course, the scope and degree of personalisation vary. Experts have identified six dimensions of personalisation – the why, how, what, when, who and where learning takes place².

One-to-one tutoring is the epitome of personalisation. In the 1960s, Benjamin Bloom showed that the average student performs better with individual tutoring. He also showed that individual attention closes the learning gap between the high and low scorers. In the real classroom, even with ten students, customising content to benefit each student can demand a lot of effort. True personalisation is practically impossible. Even where the teacher knows a student has gaps in learning they might be unable to remedy it due to time constraints. Thus, the system constantly loses students, even with the best effort of teachers.

This is where technology can lend a hand.

Technology-enhanced personalised learning

Technology can be used to customise the learning process. Here, technology includes anything from mobile apps and online platforms to stand-alone learning systems². This is more effective now that artificial intelligence, access to data, mining techniques, cloud computing and affordable hardware have made applications seamless and practical.

Well-designed technology can do more than just help you to overcome the stumbling blocks shown above. When integrated with traditional classes, either as homework or occasional classwork, it could help learners acquire and practice a routine skill. Such technology can increase time spent in the classroom for interaction, personal attention and problem solving. Further, you can monitor what happens during homework – watch how far students have progressed and where they are struggling³.

Sometimes, for parts of a lesson, software might do a better job. Think of visualising three dimensions in maths, pronunciation practice in a language class – or an animation explaining the processes inside a human cell.

All AI solutions for education can be used to different degrees to help personalise learning. In this chapter, we discuss the use of adaptive learning systems.

¹ Groff, J., Personalized Learning: The State of the Field & Future Directions, Center for Curriculum Redesign, 2017.

² Holmes, W., Anastopoulou, S., Schaumburg, H & Mavrikis, M., Technology-enhanced personalised learning: untangling the evidence, Stuttgart: Robert Bosch Stiftung, 2018.

³ Feldstein, M., Hill, P., Personalized Learning: What It Really Is and Why It Really Matters, Educause Review, 2016.

⁴ Taylor, D., Yeung, M., Bashet, A.Z., Personalized and Adaptive Learning, Innovative Learning Environments in STEM Higher Education pp 17–34,  Springer Briefs in Statistics, 2021.

Is the student struggling with a concept?

Do they seem to have a misconception?

Perhaps they are upset and could use a little encouragement?

The tutor gives a hint, points out what they are missing.

It can also happen that the student finds the problem too easy and is getting bored. In this case, the tutor assigns another, more challenging problem.

The tutor may even inspire questions and make the student reflect on their own performance. All this while keeping you informed of the student progress.

Intelligent tutoring systems (ITS) are designed to mimic the role of this tutor¹. They are a type of adaptive learning systems (ALS) that guide an individual student through each step of a solution. They give hints and feedback as needed. Because of this, ITS are more suitable for subjects like maths where problems and solutions are clearly defined². But recent ITS have taken on other subjects too.

Adaptive systems and learning

Adaptive learning occurs when digital tools and systems create individual learning paths – the sequence of activities executed in order to learn a given content or skill. The learning paths depend on each individual’s strengths, weaknesses and pace of learning^3,4.

The idea of a machine adapting to a student goes back to the 1950s. With the recent advent of technology, the possibilities are now endless. These adaptive learning systems can be used for different purposes – solving problems, learning concepts and/or for assessing the student.

Many adaptive learning systems are now on the market. There are also authoring tools that help you create an ALS without any knowledge of coding. While creating an ALS might take a lot of time and resources, the teacher need not change lesson plans or style to fit it with lessons. Whatever be their type and form, the technologies used to create ALS vary a lot – not all systems are equal!

While choosing a system, you have to see how adaptive it is, what part of learning it personalises and whether it allows customisation by the teacher. Apart from that, there are important practical questions such as what equipment is required, how much does it cost and whether training is included in the cost.

Types of adaptive learning systems

Intelligent tutoring systems (see above) are personalised and interactive. They evaluate learning in real time. On the micro level, they adapt feedback when the student is solving a problem. On the macro level, they decide what problem to show next – much like Youtube recommending what video to watch next. Simple tutoring systems use decision trees for rules on what feedback to give. Other systems go beyond pre determined rules and use machine learning to tailor their behaviour¹.

Adaptive learning systems can go beyond tutoring. Exploratory learning systems, for example, let students explore a learning environment and choose what interests them. Game based systems package everything in the form of a game. When a student masters one level, they move to the next.

Whatever the type, all ALS should support the learner until they can carry out a task independently⁶.  They should incite reasoning and support decision making. They should also be able to explain their decisions to the teacher and student.

When it comes to choosing and using ALS, or even deciding whether or not to use one, experts advise to always start with the learning². Ask what student need is to be addressed? Which tool fits this job? How will different students be supported differently⁵? Studies show that these systems do not have a significant impact on student learning when used for short periods of time. The effectiveness increased when used for one full school year or longer⁷. If you decide to use one, be prepared to support students on guiding their own learning. Be patient and be ready to experiment, fail and retry^2,5.

¹ Groff, J., Personalized Learning : The state of the field and future directions, Center for curriculum redesign, 2017.

² Holmes, W., Anastopoulou S., Schaumburg, H & Mavrikis, M., Technology-enhanced personalised learning: untangling the evidence, Stuttgart: Robert Bosch Stiftung, 2018.

³ Taylor, D., Yeung, M., Bashet, A.Z., Personalized and Adaptive Learning, Innovative Learning Environments in STEM Higher Education pp 17–34, SpringerBriefs in Statistics, 2021.

⁴ Becker, S. et al, NMC Horizon Report: 2018 Higher Education Edition, Educause, 2018.

⁵ Feldstein, M., Hill, P., Personalized Learning: What It Really Is and Why It Really Matters, Educause Review, 2016.

⁶ Wood, D., Bruner, J., Ross, G., The role of tutoring in problem solving,  The Journal of Child Psychology and Psychiatry, 1976.

⁷ Alkhatlan, A., Kalita, J.K., Intelligent Tutoring Systems: A Comprehensive Historical Survey with Recent Developments, International Journal of Computer Applications 181(43):1-20, 2019.

ACTIVITY

These are the credit card transactions of John Doe and Tom Harry, two men living in Nantes, France. They are looking for things to do this weekend. What will you recommend to them?

List to choose from:

1. The new Burger King outlet
2. An olive oil tasting event
3. An online luggage store
4. A river-side concert
5. Baby swimming class

Recommendation systems have been around at least as long as tourist guides and top-ten lists. While The Guardian Best Books of 2022 recommends the same list to everyone, you would likely adapt it when choosing for yourself - pick a few and change the order of reading based on your personal preferences.

How to recommend options for strangers? In the activity above, you probably tried to imagine their personalities based on the given information: you made judgements and applied stereotypes. Then, once you had an idea of their type, you chose from the list things that could (or not) be relevant to them. Recommenders such as Amazon, Netflix and Youtube follow a similar process.

Nowadays, whenever someone is searching for information or looking to discover online content, they use some kind of personalised recommender system^1,2. The main function of Youtube is to suggest to its users what to watch amongst all the videos available on the platform. For signed-in users, it uses their past activity to create a model, or a personality type. Once it has a model for John, it can see who else has models similar to his. It then recommends to John videos similar to what he has watched and those similar to what others like him have watched.

What is a model?

Models can be used to mimic anything from users to videos to lessons a child has to learn. A model is a simplified representation of the world, so a machine can pretend to understand it:

How Youtube learns who you are

All recommendation problems involve asking a surrogate question: "What to recommend" is too general and vague for an algorithm. Netflix asked developers what will be the rating user A would give video B, considering their ratings for other videos. Youtube asks what the watch time would be for a given user in a particular context. The choice of what to ask – and predict – has a big impact on what recommendation is shown³. The idea is that the correct prediction will lead to a good recommendation. The prediction itself is based on other users with a history of similar tastes⁴. That is, users whose models are similar.

User models

Youtube splits the task of recommendation into two parts and uses different models for each³. We, however, will stick to a simpler explanation here.

For creating a user model, its developers have to ask, what data is relevant to video recommendation. What about what the user has watched before? What about their reviews, ratings, and explicit preferences thus far? What did they search for? Youtube uses signals that are more implicit than explicit, since the latter are more readily available³. Did a user just click a video or did they watch it? If yes, for how long? How did the user react to previous recommendations¹? Which ones were ignored? Apart from direct answers to these questions, demographic information such as gender, language, region and type of device are of great value when the user is new or not signed in³.

Once a model is available for each user, we can compare users and use that information for recommendations.

Video Models

We could also use videos that are both similar to and different from one another. Youtube looks at the information it has on a particular video – its title and description, video quality, how many people have watched it (view count), liked it, favoured it, commented on it or shared it, the time since it was uploaded and the number of users subscribed to the parent channel¹.

What a user watches next will also depend on whether one video is an episode within a series or an item in a playlist. If a user is discovering a new artist, he or she might move from the most popular songs to smaller niches. Also, a user might not click on a video whose thumbnail image is poor quality^1,3. All of this information goes into the model too.

One of the building blocks of the recommendation system is to go from one video to a list of related videos. In this context, we define related videos as those that a user is likely to watch next³. The goal is to squeeze the most value out of data to make better recommendations⁴.

¹ Davidson, J., Liebald, B., Liu, J., Nandy, P., Vleet, T., The Youtube Video Recommendation System, Proceedings of the 4th ACM Conference on Recommender Systems, Barcelona, 2010.

² Spinelli, L., and Crovella, M., How YouTube Leads Privacy-Seeking Users Away from Reliable Information, In Adjunct Publication of the 28th ACM Conference on User Modeling, Adaptation and Personalization (UMAP '20 Adjunct), Association for Computing Machinery, New York, 244–251, 2020.

³ Covington, P., Adams, J., Sargin, E., Deep neural networks for Youtube Recommendations, Proceedings of the 10th ACM Conference on Recommender Systems, ACM, New York, 2016.

⁴ Konstan, J., Terveen, L., Human-centered recommender systems: Origins, advances, challenges, and opportunities, AI Magazine, 42(3), 31-42, 2021.

Across Google, deep neural networks are now being used for machine learning². Based on the video model, Youtube’s neural network takes videos similar to the ones already watched by the user. Then it tries to predict the watch time of each new video for a given user model, and ranks them based on the prediction. The idea is then to show the 10 to 20 videos (depending on the device) with top rankings.

The process is similar to the machine learning model we studied earlier. First, the machine takes features from user and video models given by the programmer. It learns from training data what weight to give each feature to predict watch-time correctly. And then, once tested and found to be working, it can start predicting and recommending.

Training

During training, millions of both positive and negative examples are given to the system. A positive example is when a user clicked on a video and watched for a certain time. A negative example is when the user did not click on the video or did not watch for long².

The network takes in the user features and video features discussed under the models section of How Youtube Learns You Part 1. It adjusts the importance given to each input feature by checking whether it predicted correctly the watch time for a given video and user.

There are approximately one billion parameters (weight of each feature) to be learned on hundreds of billions of examples². The network might also learn to disregard certain features and will give it zero importance. Thus, the embedding, or the model the algorithm creates, can be very different from that envisioned by the developers.

Testing

Once the network has been trained, it is tested on already available data and adjusted. Apart from accuracy of prediction, the output of the system has to be tuned by the programmer, based on several value judgements. Showing videos that are too similar to already watched videos will not be very engaging. What does it really mean for a recommendation to be good? How many similar videos to show and how much diversity to introduce  - both with respect to the other videos and with respect to the user history. How many of the user's interests to cover? What type of recommendations lead to immediate satisfaction and which lead to long-term use^1,3? These are all important questions to consider.

After this testing, real-time evaluation of the recommendations is done. The total watch time per set of predicted videos is measured². The longer a user spends watching the recommended set of videos, the more successful the model is considered to be. Note that just looking at how many videos were clicked is not sufficient grounds for evaluation. Youtube evaluates its recommenders based on how many recommended videos were watched for a substantial portion of the video, session length, time until first long watch and the fraction of logged-in users with recommendations¹.

The interface

We will now explore how the recommendations are presented to the viewer. How many videos should be shown? Should the best recommendations be presented all at once, or should some be saved for later³? How should thumbnails and video titles be displayed? What other information should be shown? What settings can the user control¹? Answers to these questions determine how Youtube manages to keep two billion users hooked.

¹ Davidson, J., Liebald, B., Liu, J., Nandy, P., Vleet, T., The Youtube Video Recommendation System, Proceedings of the 4th ACM Conference on Recommender Systems, Barcelona, 2010.

² Covington, P., Adams, J., Sargin, E., Deep neural networks for Youtube Recommendations, Proceedings of the 10th ACM Conference on Recommender Systems, ACM, New York, 2016.

³ Konstan, J., Terveen, L., Human-centered recommender systems: Origins, advances, challenges, and opportunities, AI Magazine, 42(3), 31-42, 2021.

4 Spinelli, L., and Crovella, M., How YouTube Leads Privacy-Seeking Users Away from Reliable Information, In Adjunct Publication of the 28th ACM Conference on User Modeling, Adaptation and Personalization (UMAP '20 Adjunct), Association for Computing Machinery, New York, 244–251, 2020.

When looking at an adaptive learning system, it is very hard to tell where it adapts¹. What technology is used and what it is used for also changes across systems.

However, all adaptive learning systems know who they teach (knowledge about the learner), what they teach (knowledge about the domain), and how to teach (knowledge about pedagogy)².

An ideal ALS adapts itself in multiple ways. In the outer loop, the sequence of learning activities is adapted - similar to Youtube adapting recommended list of videos. The outer loop might also personalise learning approaches and difficulty levels.

In the inner loop, within each activity, the ALS monitors step-by-step progress. It adapts feedback and hints to correct misconceptions, if any. It might also point to additional content if the student has a problem remembering a previously learnt concept. Some experts argue that the inner loop is best left to the instructor: not only is it costly and time-consuming to program all the rules for the specific subject and task, but the teacher’s knowledge and experience will always trump that of the machine³.

How adaptive systems learn the learner

Like all recommendation problems (See How Youtube Learns You Part 1), ALS splits the task into one or more surrogate questions that can be answered by the machine. Again, the choice of what to ask -and thus, what to predict, has a big impact on what recommendation is shown.

Marketing material often mention multiple goals – improved scores, employability and engagement. Given the proprietary nature of the systems, it is usually unclear what questions are coded into the systems, what goals are being optimised for, and how short-term goals are differentiated from long-term goals (example, mastery of a given content before progressing to the next level)⁴.

Where machine learning is used, regardless of the goals chosen, the prediction itself is based on other learners with similar skill levels and preferences. That is, learners whose models are similar.

The learner model

For creating a student model, developers ask what student characteristics are relevant to the learning process. Unlike teachers who can directly observe their students and adjust their approach, machines are limited to the data that can be collected and processed by them.

Typical characteristics considered in a student model:

• What does the student know - their knowledge level, skills and misconceptions^5,2,6. These are usually inferred through assessments, for example, the answer a student submits for a maths problem¹. This prior knowledge is then compared with what they will need to know at the end of the learning period.
• How does a student prefer to learn:  the learning process and preferences^5,6. For example, the number of times a student attempts a question before getting it correct, the types of resources consulted, the ratings they gave for an activity¹, or the material that engaged them most - images, audio or text². ALSs may also record when and how skills were learned and which pedagogies worked best⁶.
• Is the student feeling motivated:  feelings and emotions can be recorded directly by the student or extracted indirectly from speech, facial expressions, eye tracking, body language, physiological signals, or combinations thereof. This information can then be used to nudge the student out of negative states such as boredom or frustration that inhibit learning, into positive states such as engagement or enjoyment⁷.
• What about cognitive aspects like memory, attention, problem solving skills, decision-making capability, analysis of situations and critical thinking⁵?
• How do they communicate and collaborate⁵? For example, do they post comments on other students' feed and how do they discuss with others to solve problems¹?
• What about meta-cognitive skills such as self-regulation, self-explanation, self-assessment and self-management⁵, help-seeking, being aware of and being able to control their own thinking? For example, how they select their learning goals, use prior knowledge or intentionally choose problem-solving strategies⁵.

While this data changes and has to be recorded and updated, models also contain static characteristics such as age, gender, mother tongue and email identity².

Most of the ALSs create learner models based on interactions with students. Some also glean info from other sites, especially social media. Once a model is available for each learner, the machine calculates which students are similar to one another and estimate the probability that a given student will benefit from an activity, example or question³.

The domain model

We can draw a loose parallel between learning objects in an ALS and videos on Youtube. A subject can be broken down into concepts and skills; these are called knowledge units (KUs) and are what the learner needs to know³. Each KU has a set of learning objects through which content can be learned, and a set of activities to assess learning. Some authors further break down learning objects into learning activities; we do not do this here.

Learning objects can be text to read, a video, a set of problems, interactive activities (from simple fill-in-the-blanks to scenario-based learning activities), interactive animation, etc.¹ The learning objects give what the learner needs to know, and the assessment activities indicate if the knowledge has been acquired³. The domain model contains all the features of the learning objects, including the associated KU and assessment. What a learner learns next will also depend on the inter-relationships between the KUs.

These KUs need to go into the model too: learning objects A and B might both be pre-requisites for learning object D. Thus, A and B have to be mastered before D. There is order amongst some KUs that tell us how we learn³. Conversely, if the student solves correctly a problem that corresponds to D, it would be a good bet that he or she mastered A and B too.

Subject-matter experts can provide some of these relationships. The other inferences can be learnt by the machine, which can predict the probability that a KU has been mastered. This mastery involves the system knowing that the learner has mastered A and B, given that he answered questions under D. It can then use this information, along with other features of learner and domain models, to recommend learning pathways and learning objects.

Other features of learning objects could include the activity’s level of difficulty, its popularity and its ratings. The goal here, as in the case of Youtube recommendation, is to squeeze out as much information as possible from the data available available.

¹ EdSurge, Decoding Adaptive, Pearson, London, 2016.

² Alkhatlan, A., Kalita, J.K., Intelligent Tutoring Systems: A Comprehensive Historical Survey with Recent Developments, International Journal of Computer Applications 181(43):1-20, March 2019.

³ Essa, A., A possible future for next generation adaptive learning systems, Smart Learning Environments, 3, 16, 2016.

⁴ Bulger M., Personalised Learning: The Conversations We’re Not Having, Data & Society Working Paper, 2016.

⁵ Chrysafiadi, K., Virvou, M., Student modeling approaches: A literature review for the last decade, Expert Systems with Applications, Elsevier, 2013.

⁶ Groff, J., Personalized Learning: The state of the field and future directions, Center for curriculum redesign, 2017.

⁷ du Boulay, B., Poulovasillis, A., Holmes, W., Mavrikis, M., Artificial Intelligence And Big Data Technologies To Close The Achievement Gap, In: Luckin, Rose ed. Enhancing Learning and Teaching with Technology. London: UCL Institute of Education Press, pp. 256–28, 2018.

In recent years, machine learning is increasingly being used in adaptive systems, either as the sole technology or in conjunction with other approaches². Where used, the principal role of ML is in creating and updating student models, based on a set of features, including results from assessments and new data generated in the process¹.

For the outer loop, models are created with the help of training data, by assigning suitable weights –these are features that help recommend effective learning content². (Also refer to how machine learning works.). These models are used to regularly recommend new learning paths that reflect student progress and changing interests - like new recommendations in Youtube. In ML-based ALS, the number of pathways can run to trillions³.

In the inside loop, machine learning is used to give suitable feedback, spot errors, infer gaps in knowledge and assess mastery of knowledge units: While working on one activity, a student might make errors. ML can be used to predict what errors rise from which knowledge gap. If one step of the solution is correct, ML can be used to predict which knowledge units have been mastered successfully².

Other techniques used in adaptive systems involve less automation and more explicitly written rules for making inferences².

Tools that use ML use large sets of data on actual student performance and are able, over time, to create the most dynamic learning paths for students¹. As in all ML applications, there is training and testing to be done before being put to use in classrooms.

Pedagogical model

In the case of Youtube, we saw there are a lot of value judgements on what makes a good recommendation – such as how many user interests need to be covered in one set of recommendations, how many videos should be similar to already watched ones, how much new content to add for diversity (see How Youtube learns you, part 2). ALS involves similar judgements on what it means to master a KU and how to get to that mastery: the pedagogy and daily experience of the learner⁴.

In the case of ALS, these judgements and guidance on how a learner is to progress should be based on proven pedagogical theories. These go into the pedagogical model, and, along with the domain and learner models, help the machine to choose an appropriate set of activities.

Some of the questions answered in this model are: should the student next be presented with a concept, an activity or a test? At what level of difficulty? How do they evaluate the learning and provide feedback? Where is more scaffolding necessary⁵? (Scaffolds are support mechanisms that give guidance on concepts and procedure, the strategy used and how to reflect, plan and monitor learning.) The pedagogical model dictates the breadth and depth of activities – and even whether to continue within the ALS or get help from the teacher³.

The interface

The recommendations are presented along with other data such as learner progress, performance and goals. The key questions here are:

• How will content be delivered?
• How much content will be recommended in one go?
• What is assigned directly and what is recommended?
• What are the supporting resources?
• Is it possible to provide group activities?
• How much autonomy should be permitted?
• Can students change their preferences?
• Can teachers change the learning pathways?
• What data is shown to the teachers?
• Are the teachers in the loop?

Evaluation

When the ALS is put to use, most systems monitor their own performance against criteria set by the programmer. Like in any AI tool, data might be biased. Inferences drawn by the system can be imprecise. The student’s past data will become less relevant with time⁶. Therefore, the teacher also has to monitor the system’s performance and provide learner guidance and corrective measures where necessary.

Teachers and peers also have to provide inspiration and reveal alternative resources. research in recommendation systems was shaped by commercial content providers and online retail companies for over a decade. Thus, the focus has been on providing recommendations that produce results that can be promoted. “The surprising delight of an unexpected gem”⁷ and the allure of roads less travelled, can inspire enduring learning. Unfortunately, these are not the strong points of machine-based personalised learning.

¹ EdSurge, Decoding Adaptive, Pearson, London, 2016.

² Chrysafiadi, K., Virvou, M., Student modeling approaches: A literature review for the last decade, Expert Systems with Applications, Elsevier, 2013.

³ Essa, A., A possible future for next generation adaptive learning systems, Smart Learning Environments, 3, 16, 2016.

⁴ Groff, J., Personalized Learning: The state of the field and future directions, Center for curriculum redesign, 2017.

⁵ Alkhatlan, A., Kalita, J.K., Intelligent Tutoring Systems: A Comprehensive Historical Survey with Recent Developments, International Journal of Computer Applications 181(43):1-20, March 2019.

⁶ du Boulay, B., Poulovasillis, A., Holmes, W., Mavrikis, M., Artificial Intelligence And Big Data Technologies To Close The Achievement Gap, In: Luckin, Rose ed. Enhancing Learning and Teaching with Technology. London: UCL Institute of Education Press, pp. 256–28, 2018.

⁷ Konstan, J., Terveen, L., Human-centered recommender systems: Origins, advances, challenges, and opportunities, AI Magazine, 42(3), 31-42, 2021.

• Recommendation systems are used for suggesting movies to Netflix users. They help consumers home in on the right choice of, say Audio Speakers on Amazon. But can they actually improve learning outcomes for each student in the classroom¹?
• Does focussing all the time on performance and individualisation affect a student’s psychological well being²?
• Individualisation demands a lot of discipline and self-regulation from a student. They have to start working by themselves and continue working till they finish all assigned activities. Are all students able to do this without help²?
• How do we balance individualisation with social learning opportunities³?
• How do we go from using ALS as a support for a single topic, to using these systems systematically, across topics and subjects²? What about the curriculum change that will be required for such an incorporation of adaptivity³?
• What about the required infrastructure? What needs to be done about data and privacy, as well as bias and reinforced stereotypes³?

When developing ALS, some principles are used either directly or implicitly. These are not always without consequences.

A paradigm of ALS: old is gold

What do machine learning systems do when they predict or recommend something? They use the student’s past experiences, preferences and performance in order to choose what to recommend to them; they look to the past in order to predict the future. Thus, these systems are always biased towards the past⁴. Machine learning works best in a static and stable world where the past looks like the future⁵. ALS, based on machine learning models does more or less the same thing, but now with the addition of pedagogical considerations.

As a consequence, these systems are not able to account for fluctuations in normality like the COVID pandemic, health issues and other problems. They can struggle to take account of age, growth, mastery of new competencies and personal evolution of young humans.

Is student behaviour even predictable? How many times can we repeat a formula that worked well in the past, before it becomes boring, repetitive and impedes progress⁶? Even if such a prediction were possible, is it even prudent to expose students only to that which they like and are comfortable with? How much novelty is overwhelming and counterproductive⁶?

A paradigm of ALS: the explicit reflects the implicit

Even where the past can be used reliably to predict the future, the past itself could be difficult to capture accurately. How can Youtube know a user liked a video? It is easier where they clicked the Like button or subscribed to the parent channel after watching it. But such explicit behaviour is often rare. Recommendation systems have to regularly resort to implicit signals that may or may not fully reflect the truth⁴. For example, Youtube uses the time a user spent watching the video as an implicit signal that they liked the video and would like to watch similar content. But, just because a video played on someone’s computer till the end hardly means the person liked it, or even watched it⁷.

What about how feedback is recorded in an adaptive learning system? To gauge, for example, if a student was attentive during an activity, the system might record the number of digital resources they clicked on, and when and for how long they accessed them. But these cannot accurately reflect their level of attention¹.

For example, if the student is clear about what to do for an activity, they might consult a few resources and zero in on the critical points quickly. Someone who is not as clear might open and spend time on all the listed resources without learning much¹. It is possible that the first student is wrongly flagged for lack of motivation and made to do additional work.

Also, the machine learning models can only note that two things happened – a student clicking on a resource and a student scoring high on the associated exercise. They cannot infer that the student scored high because they consulted the resource – they can infer correlation but not causation⁵.

The unfair expectation of some ALS is that the teacher will dive in and remedy such errors. In other systems, the teacher does not even have the option to do so.

The paradigm of ALS: everything can be replaced by this one question

Recommendation systems cannot handle multiple goals. The aim of the ALS is often put forth in a form of a single question: the surrogate question. What rating did a user give a movie, how long did they watch a video, what is the score of the student in a quiz, how well did they satisfy the criteria the machine used to measure attentiveness… The systems are then trained to attain these goals and tested based on whether they was achieved. Their performance is constantly adjusted to maximize their score with respect to these goals.

If scoring on the quiz is the goal, certain content is recommended in a certain way. That way, exam performance is the surrogate problem that is solved. If the goal is just to make them click on many resources, recommendation would be tailored to push them to do just this. In this instance, making resources adequately attractive is the problem.

The  choice of the surrogate question has an outsized importance on how the ALS works. What is more, contrary to the promotion of ALS as objective systems, there is more art than science in selecting the surrogate problem for recommendations⁴.

All tech is not hi-tech

As we have seen, many decisions go into the making of ALS – what data is measured, how this data is used to gauge feedback and other information, what goals are optimised, and what algorithms are used to optimise these goals. Often it is programmers, data scientists, finance and marketing experts that are involved in making these decisions. The input of teachers and pedagogical experts in the development process is rare and often comes after the designing process². Products are not field-tested before adoption in schools, and often their proclaimed effectiveness is based on testimonials and anecdotes, instead of scientific research².

As a result, what a school needs and is familiar with has little impact on what software companies are building. Cost, availability and infrastructure can have a major say on what schools can buy. It is important to bear this in mind while deciding if or how to use a particular product. Perhaps it is better not to think of them all as adaptive learning systems or AI, but individual systems with wildly different objectives, designs and capabilities.

ALS as a whole can be used for personalising feedback, scaffolding and practice. They can find gaps in learning and remedy it within the limits of programming and design. They cannot detect ‘teachable moments’ or when it is right to capitalise on the mood of the class to introduce a new idea or example. These capabilities that make learning magical and which help the lesson endure in the student’s mind are solely the forte of the teacher.

¹ Bulger M., Personalised Learning: The Conversations We’re Not Having, Data & Society Working Paper, 2016.

² Groff, J., Personalized Learning: The state of the field and future directions, Center for curriculum redesign, 2017.

³ Holmes, W., Anastopoulou S., Schaumburg, H & Mavrikis, M., Technology-enhanced personalised learning: untangling the evidence, Stuttgart: Robert Bosch Stiftung, 2018.

⁴ Covington, P., Adams, J., Sargin, E., Deep neural networks for Youtube Recommendations, Proceedings of the 10th ACM Conference on Recommender Systems, ACM, New York, 2016.

⁵ Barocas, S.,  Hardt, M., Narayanan, A., Fairness and machine learning Limitations and Opportunities, MIT Press, 2023.

⁶ Konstan, J., Terveen, L., Human-centered recommender systems: Origins, advances, challenges, and opportunities, AI Magazine, 42(3), 31-42, 2021.

⁷ Davidson, J., Liebald, B., Liu, J., Nandy, P., Vleet, T., The Youtube Video Recommendation System, Proceedings of the 4th ACM Conference on Recommender Systems, Barcelona, 2010.

Historically, automatic translation has been a challenge for artificial intelligence, and the diverse AI technologies have been tested over the years. Rule-based systems (with rules hand-built by experts) were replaced by statistical machine-learning techniques when data-sets of parallel texts (the same text in multiple languages) became available. Over the past few years, deep-learning techniques have become state of the art.

Just a few years ago, you could have an enjoyable moment testing these tools, which would return amusing translations for songs or menus for example; this is no longer the case today:

• International institutions are contemplating using automatic translation tools to support multilingualism;
• Big media video platforms use automatic translation rather than human translation in order to reach more people;
• Bilingual people and translation professionals seem to use these tools both in their day-to-day lives and professional activities.

Furthermore, improvements have yet to come. The quality of translation is still improving, and solutions that combine translations with transcriptions and speech synthesis, and allow seamless multilingual communication, will soon be common.

Even if these tools have not been designed for education, they are already having an impact.

Are pupils using automatic translation?

To our knowledge, there are today (December 2022) no public official documents nor large-scale surveys measuring if this is an issue. There are discussions on forums³ and articles presenting possible ways to avoid cheating with AI, or suggesting ways to introduce AI into foreign-language classes. These assume that pupils’ usage of automatic translation tools is widespread.

We ran a smaller, informal survey in April 2022, with teachers of various languages (English, French, German) and with varying degrees of skill. The main classes corresponded to 12-16 year-old pupils. It took place in the Paris area, so pupils and teachers were French. The results were such that the teachers all had to cope with pupils who would, once out of the classroom, make use of DeepL or Google Translate.

Here are some of the remarks we got:

• The only skill the pupils seem to be acquiring is copy-pasting.
• Even the better and more motivated pupils do it – they will try to do their homework on their own, but then they will check it with an automatic translation tool. Often, they realise that the automatic result is much better than their own, so they keep the machine-built solution.
• There is a motivation issue as pupils are questioning the use of learning languages.

The above analysis needs further investigation. A generalised survey over various countries would certainly help. In the meantime, discussions with various stakeholders have allowed us to consider the following:

• Teachers can no longer assign translating text as homework. Even more creative exercises, such as writing an essay on a particular topic, can lead to using automatic translation tools – the pupil could write the essay in their own language, then translate it.
• The motivation question is critical. It is not new – in 2000, authors and educators argued that “some view the pursuit of foreign language competence as an admirable expenditure of effort; others may see it as unnecessary if an effective alternative exists”⁵.

Can automatic translators fool teachers?

Blog papers seem to indicate that a language teacher will recognize automatic translation, even when it has been corrected by a human at a later stage: Birdsell¹ imagined a task where Japanese students were to write a 500 word essay in English. Some had to write it directly, albeit with commonly used and accepted tools (dictionaries, spelling checkers) and others would write the essay in Japanese and then translate it -using DeepL- into English. Interestingly, he found that the teachers would grade higher the students from the second group but would also be able to identify those essays written with the help of DeepL.

Can machine translation tools be combined with text generators?

These are early days to predict which will be the course of events but the answer is, for the moment, yes. As a simple example, journalists in France used a text-generator tool (Open-AI playground) to produce some text, then ran DeepL on it, and felt comfortable in presenting this text to the community².

Is using an automatic translator cheating?

Thiscis a difficult question to answer. When consulting discussion forums on the internet³ you can easily be convinced that it is cheating. Students are told not to use these tools. They are told that if they don’t comply, they will be accused of cheating. But arguments can also be presented the other way. Education is about teaching people to use tools smartly in order to perform tasks. So how about making it possible for a pupil to learn to use the tools they will find available outside school?

This textbook is not authorised to offer a definitive answer, but we do suggest that teachers explore ways in which these tools can be used to learn languages.

What should a teacher do about this?

Florencia Henshaw discusses a number of options⁴, none of which seem convincing:

• Saying that AI simply doesn’t work (a favourite conclusion in forums³) is not helpful, even if pupils agree with this. They will still want to use these AI tools.
• The zero-tolerance approach relies on being able to detect the use of AI. This may be the case today¹ but it is uncertain if it will remain the case. Furthermore, is using AI cheating? In what way is it different from using glasses to read better or a wheelbarrow to transport objects?
• The approach in which the tool can be part-used (to search for individual words for example) is also criticised⁴. Automatic translation tools work because they make use of the context. On individual (out of context) words, they will not perform any better than a dictionary.
• The approach of using the tool in an intelligent way, in and out of the classroom, is tempting but will need more work to develop activities which will be of real help in learning situations.

¹ Birdsell, B. J., Student Writings with DeepL: Teacher Evaluations and Implications for Teaching, JALT2021  Reflections & new perspectives 2021.

² Calixte, L, November 2022, https://etudiant.lefigaro.fr/article/quand-l-intelligence-artificielle-facilite-la-fraude-universitaire_463c8b8c-5459-11ed-9fee-7d1d86f23c33/.

³ Reddit discussion on automatic translation and cheating. https://www.reddit.com/r/Professors/comments/p1cjiu/foreign_language_teachers_how_do_you_deal_with/.

⁴ Henshaw, F. Online Translators in Language Classes: Pedagogical and Practical Considerations, The FLT MAG, 2020, https://fltmag.com/online-translators-pedagogical-practical-considerations/.

⁵ Cribb, V. M. (2000). Machine translation: The alternative for the 21st century? TESOL Quarterly, 34(3), 560-569. https://doi.org/10.2307/3587744.

But the transformation induced by these tools is not limited to correcting some typos; it has been far more profound in invoking a different way of writing. Digital writing allows us to return to what we have written and modify it to express more effectively what we want to convey.

Using somewhat more technical words, we have moved from a linear approach to writing, to an iterative process. According to recent studies, the transformation of the writing process induced by digital tools has improved the quality of the texts produced.

Writing in the AI era

Anyway, the process of evolution of writing and related thought forms has not stopped. In recent years, with the explosion of AI, it has accelerated significantly. Tools such as Grammarly, Wordtune, Ludwig, ProWritingAid, and so on, are designed not only to provide grammatical correction of the text. They support the user throughout the writing process by stimulating the improvement of writing style, checking for plagiarism, and more.

Recognising how the school world cannot be immune to such innovations is trivial. This is confirmed by the growing number of educational interventions, proposed in literature, that are designed to take advantage of such software. Some scholars propose using these tools to work on students’ skills in using external information sources to develop appropriate paraphrasing skills that can avoid plagiarism problems. Many of these tools can support the teacher in evaluating the texts produced by the students, providing timely analysis of individual student’s strengths and weaknesses. Moreover, these tools allow the student themselves to self-assess their own writing skills, thus enabling metacognitive processes and speeding up learning.

All that glitters is not gold…

Clearly, these innovations have potential problems. First, you have probably understood how, underlying all these deep learning mechanisms, is the source data on which the models are built. Limited or incorrect training data could cause bias. In addition, the risk of a general homogenisation of the texts produced/expected by those tools is likely. It could determine a consequent limitation (or penalisation in the case of assessment) of students’ creativity. Finally, these tools are primarily limited to managing the English language; thus, non-English-speaking contexts can be used in the L2 domain. That said, the speed of innovation is such that we will soon see similar tools emerge for languages other than English.

A look to the future

One of the main cognitive processes related to the writing process is the retrieval from long-term memory of those pieces of information necessary to complete the message we want to express. It is easy to speculate how these tools will also support this process by enabling immediate and simplified access to a ‘memory’ far more extensive than our own.

Finally, the tremendous advances in text-generative processes point to a future in which these tools can support the writing process in a far more active form.

How we write text will probably change further in ways we cannot yet imagine. However, the challenge will always remain the same – to know how to consciously use the tools at our disposal and adapt our way of teaching accordingly. Are you ready?

Human knowledge is wide and variable and is inherently difficult to capture. The human mind can absorb and work with knowledge because it is, as Chomsky put it, “a surprisingly efficient and even elegant system that operates with small amounts of information; it seeks not to infer brute correlations among data points but to create explanations¹.”

Machine learning is supposed to do it by finding patterns in large amounts of data. But, prior to that, experts and programmers had to sit and code what features of the data are relevant to the problem at hand, and feed these to the machine as “parameters”^2,3. As we saw before, the performance of the system depends heavily on the quality of the data and parameters, which are not always straightforward to pinpoint.

Deep neural networks or deep learning is a branch of machine learning that is designed to overcome this by:

• Extracting its own parameters from data during the training phase;
• Using multiple layers which construct relationships between the parameters, going progressively from simple representations in the outermost layer to more complex and abstract. This enables it to do certain things better than conventional ML algorithms².

Increasingly, most of the powerful ML applications use deep learning. These include search engines, recommendation systems, speech transcription and translation that we have covered in this book. It will not be a stretch to say that deep learning has propelled the success of artificial intelligence in multiple tasks.

“Deep” refers to how the layers pile on top of each other to create the network. “Neural” reflects the fact that some aspects of the design were inspired by the biological brain. Despite that, and even though they provide some insights into our own thought processes, these are strictly mathematical models and do not resemble any biological parts or processes².

The basics of deep learning

When humans look at a picture, we automatically identify objects and faces. But a photo is just a collection of pixels for an algorithm. Going from a jumble of colours and brightness levels, to recognising a face, is a leap too complicated to execute.

Deep learning achieves this by breaking the process into simple representations in the first layer – by, say, comparing brightness of neighbouring pixels to note the presence or absence of edges in various regions of the image. The second layer takes collections of edges to search for more complex entities – such as corners and contours, ignoring small variations in edge positions^2,3. The next layer looks for parts of the objects using the contours and corners. Slowly, the complexity builds till the point where the last layer can combine different parts well enough to recognise a face or identify an object.

.

What to take into account in each layer is not specified by programmers but is learned from data in the training process3. By testing these predictions with the real outputs in the training dataset, the functioning of each layer is tuned in a slightly different way to get a better result each time. When done correctly, and provided there is sufficient good-quality data, the network should evolve to ignore irrelevant parts of the photo, like exact location of the entities, angle and lighting, and zero in on those parts which make recognition possible.

Of note here is the fact that, despite our use of edges and contours to understand the process, what is actually represented in the layers is a set of numbers, which might or might not correspond to things that we understand. What doesn’t change is the increasing abstractness and complexity.

Designing the network

Once the programmer decides to use deep learning for a task and prepares the data, they have to design what is called the architecture of their neural network. They have to choose the number of layers (depth of the network) and the number of parameters per layer (width of the network). Next, they have to decide how to make connections between the layers – whether or not each unit of a layer will be connected to every unit of the previous layer.

The ideal architecture for a given task is often found by experimentation. The greater the number of layers, the fewer the parameters that are needed per layer and the network performs better with general data, at the cost of it being difficult to optimise. Fewer connections would mean fewer parameters, and lesser amount of computation, but reduces the flexibility of the network².

Training the network

Let us take the example of a feed forward neural network doing supervised learning. Here, information flows forward from layer to deeper layer, with no feedback loops. As for all machine-learning techniques, the goal here is to find out how the input is connected to the output – what parameters come together, and how they come together to give the observed result. We assume a relationship f that connects the input x to the output y. We then use the network to find the set of parameters θ that give the best match for predicted and actual outputs.

Here the prediction for y is the final product and the dataset x is the input. In face recognition, x is usually the set of pixels in an image. y can be the name of the person. In the network, the layers are like workers in an assembly line, where each worker works on what is given to them and passes it forward to the next worker. The first one takes the input and transforms it a little bit and gives it to the second in line. The second does the same before passing it to the third, and so on until the input is transformed into the final product.

Mathematically, the function f is split into many functions f1, f2, f3... where f= ....f3(f2(f1(x))). The layer next to the input transforms input parameters using f1, the next layer using f2, and so on. The programmer might intervene to help choose the correct family of functions based on their knowledge of the problem.

It is the work of each layer to assign the level of importance – the weight given to each parameter that it receives. These weights are like knobs that ultimately define the relationship between the predicted output and input in that layer³. In a typical deep-learning system, we are looking at hundreds of millions of these knobs and hundreds of millions of training examples. Since we neither define nor can see the output and weights in the layers between input and output, these are called hidden layers.

In the case of the object recognition example discussed above, it is the work of the first worker to detect edges and pass on the edges to the second one who detects contours and so on.

During training, the predicted output is taken and compared with the real output. If there is a big difference between the two, the weights assigned in each layer will have to be changed by a lot. If not, they have to be changed a little. This work is done in two parts. First the difference between prediction and output is calculated. Then another algorithm computes how to change the weights in each layer, starting from the output layer (in this case, the information flows backwards from the deeper layers). Thus at the end of the training process, the network is ready with its weights and functions to attack test data. The rest of the process is the same as that of conventional machine learning.

¹ Chomsky, N., Roberts, I., Watumull, J., Noam Chomsky: The False Promise of ChatGPT, The New York Times, 2023.

² Goodfellow, I.J., Bengio, Y., Courville, A., Deep Learning, MIT Press, 2016.

³ LeCun, Y., Bengio, Y., Hinton, G., Deep learning, Nature 521, 436–444 (2015).

• Word processors
• Automatic grammar and orthography correction
• Automatic completion
• Optical character recognition (OCR)

More recently, chatbots, home assistants and automatic translation tools have been making a huge impact in all areas.

For a long time, research and industry was stalled by the intrinsic complexity of language. At the end of the 20th century, grammars for a language, written by experts, could have up to 50,000 rules. These expert systems were showing that technology could make a difference, but robust solutions were too complex to develop.

On the other hand, speech recognition needed to be able to make use of acoustic data and transform it into text. With the variety of speakers one could find, a hard task indeed!

Researchers understood that if we had a model for the intended language, things would be easier. If we knew the words of the language, how sentences were formed, then it would be easier to find the right sentence from a set of candidates to match a given utterance, or to produce a valid translation from a set of possible sequences of words.

Another crucial aspect has been that of semantics. Most of the work we can do to solve linguistic questions is shallow; the algorithms will produce an answer based on some local syntactic rules. If in the end, the text means nothing, so be it. A similar thing may happen when we read a text by some pupils – we can correct the mistakes without really understanding what the text is about! A real challenge is to associate meaning to text and, when possible, to uttered sentences.

There was a surprising result in 2008¹. One unique language model could be learnt from a large amount of data and used for a variety of linguistic tasks. In fact, that unique model performed better than models trained for specific tasks.

The model was a deep neural network. Nowhere as deep as the models used today! But enough to convince research and industry that machine learning, and more specifically Deep learning was going to be the answer to many questions in NLP.

Since then, natural language processing has ceased to follow a model-driven approach and has been nearly always based on a data-driven approach.

Traditionally, the main language tasks can be decomposed into 2 families - those involving building models and those involving decoding.

Building models

In order to transcribe, answer questions, generate dialogues or translate, you need to be able to know whether or not “Je parle Français” is indeed a sentence in French. And as with spoken languages, rules of grammar are not always followed accurately, so the answer has to be probabilistic. A sentence can be more or less French. This allows the system to produce different candidate sentences (as the transcription of a sound, or the translation of a sentence) and the probability is given as a score associated with each candidate. We can take the highest-ranking sentence or combine the score with other sources of information (we may also be interested in what the sentence is about).

Language models do this, and the probabilities are built from machine-learning algorithms. And of course, the more data there is, the better. For some languages there is a lot of data from which to build language models. For others, this is not the case; these are under-resourced languages.

In the case of translation, we want not two but three models: a language model for each language and another model for the translations, informing us of what the better translations of fragments of language can be. These are difficult to produce when data is scarce. If models for common language pairs are easier to build, this will not be the case for languages that are not frequently spoken together (such as Portuguese and Slovene). A typical way out of this is to use a pivot language (typically English) and translate via this pivot language – from Portuguese to English and then from English to Slovene. This will lead to inferior results as errors accumulate.

Decoding

Decoding is the process in which an algorithm takes the input sequence (which can be signal or text) and, by consulting the models, makes a decision, which will often be an output text. There are here some algorithmic considerations - in many cases transcription and translation are to happen in real time and diminishing the time lag is a key issue. So there is room for a lot of artificial intelligence.

End-to-end

Nowadays, the approach of building these components separately and combining them later has been replaced by end-to-end approaches in which the system will transcribe/translate/interpret the input through a unique model. Currently, such models are trained by deep neural networks, which can be huge. It is reported that the current largest GPT3 model comprises several hundred million parameters!

Let’s try to get at the intuition behind this process. Suppose we have some data. This raw data can be encoded in some way. But the encoding can be redundant, and perhaps even expensive. Let us now build a particular machine, called an auto-encoder (see diagram to the left). This machine will be able to take a text, compress it into a small vector (this is the encoder), and then uncompress the vector (the decoder part) and return a text which is somehow close to the original text. The idea is that this mechanism will make the intermediate vector meaningful, with two desirable properties – a reasonably small vector which ‘contains’ the information in the initial text.

The future

An example of an end-to-end process that might be available soon will be the ability to perform the following task – it will hear you speak your language, transcribe your text, translate it to a language you don’t know, train a speech synthesis system to your voice, and have your own voice speak the corresponding text in a new sentence. Here are two examples produced by researchers at the Universidad Politecnica de Valencia, Spain, in which the speaker’s own voice model is used to do the dubbing.

Some consequences for education

The steady progress of natural language processing is remarkable. Where we would laugh at the translations proposed by AI just ten years ago, it becomes increasingly difficult to find such obvious errors today. Speech recognition and character-recognition techniques are also improving fast.

The semantics challenges are still there, and answering questions which require a deep understanding of a text is still not quite right. But things are going in the right direction. This means that the teacher should expect some of the following statements to be true soon, if they are not already out there!

• A pupil will take a complex text and obtain (with AI) a simplified version; the text may even be personalised and use terms, words and concepts the pupil is used to;
• A pupil will be able to find a text, copy it, and obtain a text stating the same things but undetectable by an anti-plagiarism tool;
• Videos produced anywhere in the world will be accessible through automatic dubbing in any language. This means that our pupils will be exposed to learning material built in our language and also by material initially designed for another learning system in a different culture;
• Writing essays could become a task of the past, as tools will enable writing on any topic.

It is clear that AI will be far from perfect, and the expert will detect that even if the language is correct, the flow of ideas won’t be. But let’s face it, during the course of education, how long does it take for our pupils and students to reach that level?

¹ Collobert, Ronan, and Jason Weston. A unified architecture for natural language processing: Deep neural networks with multitask learning. Proceedings of the 25th international conference on Machine learning. 2008. http://machinelearning.org/archive/icml2008/papers/391.pdf. Note: this reference is given for historical reasons. But it is difficult to read!

Broadly speaking, human agency is the capacity of individuals to act independently, making choices from among different options based on their own beliefs, values and goals. In other words, it is the ability of humans to make decisions and initiate and carry out actions that influence their lives and the world around them. It can involve multiple dimensions. These include intentionality (acting consciously with a specific aim or goal in mind), autonomy (independence, self-determination and the freedom to make choices and decisions that reflect one’s own preferences, values and goals), adaptability (the capacity to learn, modify one’s behaviour and succeed in response to changing circumstances), and responsibility (the ethical and moral dimension of agency – being accountable for the consequences of one’s decisions and actions).

Human agency is crucial for personal growth and a successful life. It empowers individuals to shape their own lives and influence the world around them. It fosters a sense of control and self-efficacy, and is correlated with higher levels of psychological well-being. When individuals feel that they have control over their lives and can make meaningful choices, in other words when they feel they have genuine agency, they are more likely to experience satisfaction and fulfilment. As one of the leading researchers of human agency explains: “Unless people believe they can produce desired effects by their actions, they have little incentive to act or to persevere in the face of difficulties.”³.

In educational contexts, human agency refers to the capacity of students and teachers to make choices, to act autonomously, and to take control of teaching and learning within classroom settings. The emphasis is on the role of individuals in shaping educational journeys, making decisions about what, how, and why something is learned. Human agency in educational contexts involves multiple considerations. For example, student agency can be enhanced when they are not treated as passive recipients of knowledge but as active participants in the learning process, and have the autonomy to explore topics of interest, ask questions, identify and set their own academic goals, and take ownership of their learning. Enhancing student agency also involves promoting problem-solving and critical-thinking skills (centred on real-world problems) along with self-regulation skills (such as managing time, setting priorities, and self-monitoring progress), All of this is valuable for the development of independent and self-directed individuals and for academic and lifetime success. Finally, while teachers play a central role in supporting and enhancing student agency, human agency in educational contexts also includes the agency of the teachers themselves, acknowledging their subject and pedagogic expertise and professionalism, enabling them to make choices on how best to conduct their teaching and to support their students.

The next question is, what is the impact of AI on human agency? Inevitably, each potential impact has both a positive and a negative spin. For example, some AI-enabled technologies can undertake repetitive tasks, possibly freeing people to focus on the more creative aspects of their work, perhaps allowing them to decide how best to allocate their own time and effort, thus enhancing their agency. On the other hand, using AI-enabled technologies to undertake common tasks may lead to the loss of human skills or expertise. Over time, as they become increasingly reliant on AI, this could instead reduce the choices available to individuals and thus weaken their agency. Similarly, it is often argued that AI-enabled technologies can personalise user interactions with various services (such as online video and shopping platforms), providing suggestions tailored to their preferences, which strengthens their sense of agency. However, looked at with a more critical eye, the personalisation provided by such services is usually tailored more to the needs of the provider and their advertisers than the user. This means that the individual’s sense of agency masks the reality – which is a reduction in individual agency, as the user is nudged in particular directions. In addition, AI-enabled data analyses can provide access to valuable insights that might not be readily available otherwise, thus enhancing human decision making and agency. However, it is well known that AI systems inherit and perpetuate biases present in their training data This can lead to unfair and discriminatory outcomes, which inevitably undermines human agency by limiting opportunities. AI-enabled technologies, or at least the way in which they are employed in practice, can have other negative impacts on human agency. For example, the widespread use of AI for monitoring (or surveillance) and AI-controlled decision-making, raises important privacy concerns, limits choices of action, and can lead to a sense of powerlessness or dependence on technology, all of which can undermine individual agency.

The next question is, what is the impact of AIED on student and teacher agency? The possibilities are many. First, if students engage frequently with AI-enabled technologies, they might all too easily become over-reliant on the content recommendations, the instant feedback, or the “solutions” that are provided. Accordingly, they might miss out on opportunities to develop critical thinking, independent problem-solving, self-reflection, self-regulation and metacognitive skills, all of which potentially reduces the students’ agency to realise the full benefits of their own learning. Second, most AIED systems provide highly prescriptive learning paths, leaving little room for students to explore their own interests. This can limit students’ agency by dictating what, when and how they learn, and can also reduce their exposure to diverse perspectives and novel areas of inquiry. Third, AIED systems typically track student behaviour, leading to a feeling of surveillance and constrained autonomy, as well as opening the possibility of undermining student privacy, causing students to become cautious about expressing themselves freely. Fourth, AI-driven recommendations may inadvertently narrow students’ aspirations, limiting their agency to achieve self-determined goals. Fifth, AI-enabled technologies used for assessment involve an overemphasis on standardised testing (and thus, inevitably, teaching to the test). In addition, no AI-enabled system is capable of understanding or capturing the nuances in a students’ work, thus reducing the student’s agency in the assessment process and possibly discouraging creative or unconventional thinking.

Finally, with regard to teachers, the use of AI-enabled technologies in classrooms inevitably affects curriculum choices, learning content, and pedagogical approaches. This can diminish the role of teachers and lead them to feel their professional judgement is undervalued or overridden by technology. In any case, AIED tools can potentially de-skill teachers, turning them into technology facilitators and behaviour monitors. This radically misunderstands what good teachers do. It might also undermine teachers’ agency in building meaningful relationships with their students, which are crucial for effective education. Finally, reliance on AI-generated metrics (sometimes because of top-down directives) can create pressure on teachers to conform to data-driven, decision-making processes, which might lead to less emphasis on holistic student development.

The final question here is, what needs to be done to ensure student and teacher agency, as powerful AI-enabled technologies become increasingly available in classrooms. In short, teachers need to be afforded opportunities that respect their agency, allowing them to make decisions that align with their professional expertise and the specific needs of their students. Meanwhile, students need opportunities to develop their critical thinking, self-regulation and metacognitive skills, and to develop their intentionality, their autonomy, their adaptability, and their responsibility – either with or without the use of appropriate, effective and safe AI-enabled technologies.

¹ Anderson et al., Artificial intelligence and the future of humans, Pew Research Center, 2018

² Holmes et al., Artificial intelligence and Education, A critical view through the lens of human rights, democracy and the rule of law, Council of Europe, p. 34, 2022

³ Bandura, A., Toward a Psychology of Human Agency: Pathways and Reflections, Perspectives on Psychological Science, 13(2), 130-136, 2018

A lot of money, computing resources, time and effort go into the creation of datasets, benchmarks and algorithms for machine learning. This is particularly true for deep learning and large-scale models. Therefore, it makes sense that the created resources are shared within this ecosystem. This is the case with many of the ML systems that we often use. Even where the end products are different and are created by a different company, methodology, datasets, machine-learning libraries and evaluations are often shared¹. Thus, there is an argument to be made for their outputs to be similar under similar conditions.

If the output is an educational decision, it raises concern, for example, for the student who might be rejected unfairly from every educational opportunity¹. But whether or not algorithmic homogenisation constitutes an injustice can only be decided on a case–by-case basis¹.

On the other hand, if the system’s task is to help the student write, it brings into focus standardisation of writing styles, vocabulary and hence, thought patterns. Language models used in these cases are decided to predict the most probable text, based on their training dataset. These datasets, if not shared between systems, are constructed in a similar way, often with public internet data. Even when this data is screened for bias, prejudice and extreme content, it represents only a small ecosystem and is not representative of the world in all its diversity of ideas, culture and practices. Predictive text systems, based on deep learning and used for text messages and emails, have been demonstrated to change the way people write. The writing tends to become “more succinct, more predictable and less colourful”².

Invisibility

Sequences of words which are repeated in the training data trickle down into the output of large language models. Thus, the values of the database creators get the power to curb alternative opinions and plural expressions of ideas. Without proper pedagogical interventions, this in turn might limit students’ creativity and originality, not only leading to formulaic writing but ultimately to citizens with lesser critical-thinking skills and thus to an overall less colourful world³.

Closely linked with many of the negative fallouts of machine learning, including homogeneity as discussed above, is the fact that the technologies have become so advanced that the human-machine interface is seamless and practically invisible. Whether it is the search engines incorporated in the browser’s address bar, or text prediction that works intuitively, with no time lag between writing, predicting and choosing suggestions, we often act under the influence of technology without consciously being aware of it or having the choice to put a brake and rethink situations and make our own decisions. Moreover, when we use it habitually to make decisions, we tend to forget its existence altogether⁴. “Once we are habituated to technologies, we stop looking at them, and instead look through them to the information and activities we use them to facilitate”. This raises such serious concerns about human agency, transparency and trust, especially when it comes to young minds, that experts have recommended interfaces be made more visible and even unwieldy⁴.

What’s beyond: an ethical AI

In each part of this open textbook, we have discussed pedagogical, ethical and societal impacts of AI, especially data-based AI. Data and Privacy, reliability of content and user autonomy, impact on personal identity, Bias and Fairness and Human agency were all discussed in their respective pages. Issues specific to search engines were discussed in Behind the Search Lens: Effects of search on the individual and the society, problems relating to adaptive systems were dealt with in The Flip Side of ALS and those particular to Generative AI in The Degenerative. In several places throughout the book, we looked at remedial measures that can be taken in the classroom to deal with specific problems. Our hope is that these measures will become less onerous once we have ethical and reliable AI systems for education. This ethical AI would be developed, deployed and used in compliance with ethical norms and principles⁵ and would be accountable and resilient.

Since we cede so much power to AI models and their programmers, sellers and evaluators, it is only reasonable to ask them to be transparent and assume responsibility and remedy errors when things go wrong⁶. We need service-level agreements that clearly outline “the support and maintenance services and steps to be taken to address reported problems”⁵.

A resilient AI would accept its imperfections, would expect them and can work in spite of them. Resilient AI systems would fail in a predictable way and have protocols to deal with these failures⁶.

In education, ethical AI should be guided by user-centred design principles and take into account all aspects of education⁷. Teachers would be able to inspect how it functions, understand its explanations, override its decisions or pause its use without difficulty⁸. These systems would reduce teacher workload, give them detailed insights into their students and support them in enhancing the reach and quality of education⁸. They would not cause harm to their users or the environment and would enhance the social and emotional well-being of learners and teachers⁵.

Until this day comes, a teacher will have to try to develop and participate in a community of colleagues and educators to raise awareness about problems, share experiences and best practices, and identify reliable providers of AI. They could also involve students and parents in discussions and decisions to better address different concerns and develop an environment of trust and camaraderie. They would be best served by doing their best to stay up to date with the latest trends in AIED and acquire competencies when and where possible⁵.

¹ Bommasani, R., et al, Picking on the Same Person: Does Algorithmic Monoculture lead to Outcome Homogenization?, Advances in Neural Information Processing Systems, 2022.

² Varshney, L., Respect for Human Autonomy in Recommender System, 3rd FAccTRec Workshop on Responsible Recommendation, 2020.

³ Holmes, W., Miao, F., Guidance for generative AI in education and research, UNESCO, Paris, 2023.

⁴ Susser, D., Invisible Influence: Artificial Intelligence and the Ethics of Adaptive Choice Architectures, Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, Association for Computing Machinery, New York, 403–408, 2019.

⁵ Ethical guidelines on the use of artificial intelligence and data in teaching and learning for educators, European Commission, October 2022.

⁶ Schneier, B., Data and Goliath: The Hidden Battles to Capture Your Data and Control Your World, W. W. Norton & Company, 2015.

⁷ Tlili, A., Shehata, B., Adarkwah, M.A. et al, What if the devil is my guardian angel: ChatGPT as a case study of using chatbots in education, Smart Learning Environments, 10, 15 2023.

⁸ U.S. Department of Education, Office of Educational Technology, Artificial Intelligence and Future of Teaching and Learning: Insights and Recommendations, Washington, DC, 2023.

In late November 2022, the world was introduced to ChatGPT, an Artificial Intelligence (AI)-based chatbot system based on a language model called GPT-3.5. It uses natural language processing (NLP) to generate conversations¹. ChatGPT was the latest in a series of such tools but, unlike previous tools, it has captured public interest and imagination. It had more than a million users within a week of its launch because of its ability to generate human-like text and its perceived implications and potential use in education, the workplace and everyday life. ChatGPT can answer questions and assist you with tasks such as composing emails, essays and code².

GPT-3 (Generative Pre-trained Transformer-3) is a Large Language Model (LLM) that was trained through deep learning on a vast quantity of data (499 billion data points – 800GB of data), a model some one hundred times larger than any previous models³. ChatGPT enables human-like conversations with this model. A large language model is trained to predict the next word in a sentence, somewhat in the same way that a text-message auto-complete-feature works. It continues this process to generate more text.

Since its arrival there have been many developments in relation to the creation and development of generative AI tools. We have captured some of these in the graphic above. Prior to November 2022, few people knew what generative AI was. All of a sudden, terms such as ChatGPT became everyday terms for many – particularly those in education. The indications are that these technologies will continue to develop and become more useful to all in education.

Generative AI – What are they?

Generative AI systems create new content in the form of images, text, audio, video, and more:

• • Images: generative AI can create new images based on existing ones, such as creating a new portrait based on a person’s face, or a new landscape based on existing scenery;
  • Text: generative AI can be used to write news articles, poetry and even scripts. It can translate text from one language to another;
  • Audio: generative AI can generate new music tracks, sound effects and voice acting⁴.

But what is in a name?

Generative AI tools have multiple uses, so let's give a number of tools - Bing Chat, ChatGPT 3.5 and Google Bard the following prompt:

Here is what Bing Chat created:

Here is what ChatGPT created:

Here is what Bard created:

These tools allow us to write prompts and then have a conversation (conversational AI) with a ChatBot. There are a growing list of ways, which we touch on in the textbook, how these tools can help teachers, students and administrators in education and training. And they allow you to have a conversation, so think of the ways that might help you or your learners – you can have a chat about a topic you need to research or to prepare for an interview or to generate ideas for a lesson or a paper. It is interesting to compare and contrast the outputs they each create and then pick the best pieces for your purpose.

These tools can ‘enhance’ or assist you to be more creative, productive and effective when you combine your human abilities with the abilities of the machine. So always remember to critically evaluate what it produces. So have some fun learning how to have conversations with these tools and see what they produce and if they are useful to you and your students.

Generative AI in Numbers⁵

The following graph⁶ shows the rise of ChatGPT between November 2022 and May 2023. It gained one million users in its first week after it was launched. It should be noted that Netflix took 3.5 years to reach a similar milestone, with Twitter taking two years⁷.

While ChatGPT reached its peak in May 2023, it continues to retain its popularity with other generative AI tools such as Bing Chat and Bard also gaining traction. In the case of Bing Chat it started to become more popular when Microsoft acquired ChatGPT in March 2023, while Bard continues to grow in popularity as well.

¹ https://www.forbes.com/sites/bernardmarr/2022/12/21/chatgpt-everything-you-really-need-to-know-in-simple-terms/?sh=16674aadcbca

² https://www.zdnet.com/article/what-is-chatgpt-and-why-does-it-matter-heres-everything-you-need-to-know/

³ https://hai.stanford.edu/news/how-large-language-models-will-transform-science-society-and-ai

⁴ https://www.weforum.org/agenda/2023/02/generative-ai-explain-algorithms-work/

⁵ https://research.aimultiple.com/generative-ai-applications/

6 https://www.tooltester.com/en/blog/chatgpt-statistics/

7 https://bootcamp.uxdesign.cc/chatgpt-vs-bing-chat-which-is-better-2e46fa821d7d

Here, the goal is to assist you in evaluating, choosing and carrying out such activities.

Due to the speed of evolution of this technology and given how performances change with subjects and contexts, all the questions a teacher need to ask cannot be listed in advance. We hope only to get you started in lines of thinking recommended by published guidelines and supported by research and case studies.

Please note that in this book we look only at those activities where the teacher, not the student, interacts with generative AI. Our reasons behind this are explained in the link provided in the box.

Even if you are not convinced that Gen AI tools should have a place in education, we urge you to try at least one activity that uses generative technology. It might help you:

• Understand what someone can and cannot do with the tools provided by generative AI;
• Reformulate or rearrange course elements, especially when deciding what can and cannot be given for homework. For example, ChatGPT can be effective in helping students write essays¹ and it might be futile to assign standard essays for homework. You may have to use a variety of assessment methods such as group projects, hands-on activities and oral presentations, and go beyond simple yes or no questions¹;
• Understand the limitations of this technology so you could use them to motivate students to apply themselves, their imagination and creativity²;
• Not feel left behind when the next generation of these technologies, with more advanced capabilities, become reality;
• Use them to reduce your workload and explore new topics which might have been too forbidding or time-consuming before;
• Take advantage of new schools of thought and approaches to learning that are now possible;
• Help shape research and development of educational applications of these tools³.

At the same time, given the novelty of the technology and methods, we strongly recommend that you record in a journal the particulars of any activity and its evolution. This will help you to monitor and evaluate the short-term and long-term effects of each activity and to discuss with your colleagues what you observed.

Questions on pedagogy and practice

Anyone can suggest an activity for the classroom. It might happen that they come from the world of technology and have little understanding of the classroom situation. Or, It might be an idea that works for a university student but not an adolescent. Even if the activity is suitable for your classroom, if it is not backed up by reliable pedagogical theory and classroom evidence, it might produce some surprises in the long term. Thus, we encourage you to approach any activity with questions on pedagogy and practicality.

Questions on the Generative AI

At the heart of a generative AI application is a large language model (LLM) or image model (diffusion model). As linguist Noam Chomsky put it, “Roughly speaking, they [large language models] take huge amounts of data, search for patterns in it and become increasingly proficient at generating statistically probable outputs, such as seemingly human-like language and thought⁴.” BERT, BLOOM, GPT, LLaMA and PaLM are all large language models. The corresponding deep-learning model for images is called a diffusion model. Stable diffusion and Midjourney are popular examples.

Either the parent company or a third party can take a LLM and further train it (fine tune it) for specific tasks such as responding to queries and summarising essays. Or, they might take a LLM or a chatbot, add a few prompts or do extensive programming and release the result as an application package (Chatpdf, Elicit, Compose AI, DreamStudio, NightCafe, PhotoSonic, Pictory).

OpenAI fine-tuned its GPT3 and GPT4, in order to produce ChatGPT. It used prompt-response samples and rules on acceptability of content. A Google  Research team trained PaLM with scientific and mathematical data in order to get Minerva. This language model then achieved state-of-the-art results for a language-model application which solved quantitative reasoning problems. It was able to solve nearly a third of undergraduate-level problems in physics, biology, chemistry and economics – and other sciences that require quantitative reasoning⁵.

Ongoing work is needed, such as adding subject knowledge and de-biasing, in order to fine-tune a language model for educational uses: Ed-GPT².

Whether a language model has been fine tuned for a specific task will affect its efficacy at that task⁶. Whether the whole package is provided by a single company (OpenAI’s ChatGPT) or whether another company further developed the model, will affect safety and data privacy. When exploring what model is used, it’s helpful to look at both the successes and limitations of the core model and who did what to it.

Questions on its fitness for you and your classroom

An activity might achieve all its goals and the generative AI tool used could be the best and most ethical available, the activity will still need to be adjusted to suit your classroom. Like for any AI tool, you might have to do multiple iterations before you achieve your goals². You might need training and practice with both prompting techniques and critical evaluation of the output¹. Moreover, the whole experience should be rewarding to you and in line with your values as a teacher.

¹ Tlili, A., Shehata, B., Adarkwah, M.A. et al, What if the devil is my guardian angel: ChatGPT as a case study of using chatbots in education, Smart Learning Environments, 10, 15 2023.

² Holmes, W., Miao, F., Guidance for generative AI in education and research, UNESCO, Paris, 2023.

³ Becker, B., et al, Programming Is Hard - Or at Least It Used to Be: Educational Opportunities and Challenges of AI Code Generation, Proceedings of the 54th ACM Technical Symposium on Computer Science Education V. 1 (SIGCSE 2023), Association for Computing Machinery, New York, 500–506, 2023.

⁴ Chomsky, N., Roberts, I., Watumull, J., Noam Chomsky: The False Promise of ChatGPT, The New York Times, 2023.

⁵ Lewkowycz, A., Andreassen, A.,  Dohan, D. et al, Solving Quantitative Reasoning Problems with Language Models, Google Research, 2022.

⁶ Enkelejda, K., et al, Chatgpt for Good? on Opportunities and Challenges of Large Language Models for Education, EdArXiv, 2023.

1. Use generative AI for course and lesson planning and design

Would you like new ideas on redesigning some of your courses, introducing new topics, activities to include and rubric on how to evaluate the results of these activities^1,2? Would you like to try a new pedagogical approach, using new technology and materials³? Chatbots could potentially help with all of the above. You could direct the software to write out the first draft for lesson plans, learning objectives, directions for activities, projects and scientific experiments and discussion prompts⁴.

Tip: It is a good idea to specify beforehand what topics and approaches to cover and what the course or lesson would achieve - whether it is a concept or a procedure you are targeting and what kind of teaching you want to apply².

2. Use generative AI for preparing engaging, multimodal and inclusive content on a topic

• Add content related to local phenomena, language and culture;
• Insert explanatory and attractive images and videos³;
• Create and include stories that reinforce textual content;
• Build concept maps;
• Highlight, paraphrase and summarise relevant portions of the lesson, and clarify vocabulary⁶;
• Make maths and science less abstract by showing simulations and including examples from other subjects. This could also be used in craft education where visualisation can help with developing and exploring ideas⁷;
• Provide translations to help learners who use minority languages².

Tip: Midjourney and other image-creation tools often have dedicated forums where people post their work, advice, prompts and prompt-crafting hints⁷. They are helpful for teachers who are just starting to explore this technology.

3. Use Generative AI to create examples, practice exercises and quizzes

Generative AI can

• Generate tabular and other forms of data to use in examples and exercises;
• Create practice exercises and solutions, especially to reinforce foundational skills and low-knowledge components⁸. These can serve as a reference for students who are struggling to solve assignment problems⁹;
• Generate explanations for solutions. This seems to work well, especially for generating explanations of programming codes⁹;
• Generate multiple solutions for the same problem in maths or in programming. This helps students to learn different approaches and perspectives, and to analyse different methods, their advantages and disadvantages⁹. This could even help when the provided solutions are incorrect, when clearly marked so.

Example

Source: Examining Science Education in ChatGPT: An Exploratory Study of Generative Artificial Intelligence⁵.

Activity: Creating a rubric for grading.

Gen AI tool: ChatGPT.

Prompt used: Create a rubric that year 7 students can use to self‐assess their learning about renewable and non‐renewable energy sources (with the follow‐up formatting request: make this rubric easier to copy to Word in a tidy format).

Output:

[caption id="attachment_154" align="aligncenter" width="1024"] Rubric output generated by ChatGPT, from Examining Science Education in ChatGPT: An Exploratory Study of Generative Artificial Intelligence[/caption]

Reflection: The author found that ChatGPT was able to create a rubric with its own criteria – understanding, research, critical thinking and participation. Levels were marked for each, but the indicators were too vague for each criterion.

Example

Source: The Robots Are Coming: Exploring the Implications of OpenAI Codex on Introductory Programming¹⁰.

Activity: Generating multiple solutions for a question.

Gen AI tool: Codex (Codex can generate code in several programming languages in response to prompts).

Prompt used: Problem description as found in various sources and exam questions on writing code, as given to students.

Output:

[caption id="attachment_546" align="aligncenter" width="321"] Student scores on invigilated tests (Test 1 and Test 2), with performance of Codex(plotted as red asterisk), from The Robots Are Coming: Exploring the Implications of OpenAI Codex on Introductory Programming[/caption] Reflection: The authors found that Codex gave a range of different codes to the same prompt, “while ultimately favouring expected methods for each response”.

4 Increasing accessibility

As a final example, we strongly urge you to explore generative AI to increase accessibility for students with certain physical or learning disabilities, especially those with hearing or visual impairments. Generative AI can provide subtitles, captions and audio descriptions2. As per Unesco’s Guidance for generative AI in education and research, “GenAI models can also convert text to speech and speech to text, to enable people with visual, hearing, or speech impairments to access content, ask questions, and communicate with their peers”². It can also help you verify that the content you create is inclusive and accessible⁴.

Questioning the output

If you do decide to use generative AI, you have to watch out for its failures and shortcomings and be ready to correct them. These include

• Inaccuracies in the content: The language model is not a knowledge bank or a search engine. Even the latest models hallucinate facts and cite fictitious sources. Errors creep up, especially when using language models for mathematics and quantitative reasoning. Even when fine-tuned specifically for these subjects, the model can produce incorrect answers, calculation errors and hallucinate “mathematical facts”¹¹. Even programming could be tricky, because generated code can have incorrect syntax and security issues⁹.
• Biases that creep up because, among other things, the data these models were trained on was riddled with them. Even EdGPT, which is corrected for this, might still contain a few of them².
• Performance that can vary highly with the prompt used, user history, and, sometimes, with no reason at all.

While generative AI can reduce teacher workload and help with certain tasks, it is based on statistical models that have been constructed from huge amounts of online data. This data cannot replace the real world, its contexts and relationships. ChatGPT cannot provide context or explain what is affecting a student’s daily life⁴. It cannot provide new ideas or solutions to real-world challenges².

Finally, its performance does not approach the capabilities of the human mind – especially what it can understand and do with limited data. Generative AI’s “deepest flaw is the absence of the most critical capacity of any intelligence: to say not only what is the case, what was the case and what will be the case – that’s description and prediction – but also what is not the case and what could and could not be the case¹².”

¹ Tlili, A., Shehata, B., Adarkwah, M.A. et al, What if the devil is my guardian angel: ChatGPT as a case study of using chatbots in education, Smart Learning Environments, 10, 15 2023.

² Holmes, W., Miao, F., Guidance for generative AI in education and research, UNESCO, Paris, 2023.

³ Enkelejda, K., et al, Chatgpt for Good? on Opportunities and Challenges of Large Language Models for Education, EdArXiv, 2023.

⁴ Trust, T., Whalen, J., & Mouza, C., Editorial: ChatGPT: Challenges, opportunities, and implications for teacher education, Contemporary Issues in Technology and Teacher Education, 23(1), 2023.

⁵ Cooper, G., Examining Science Education in ChatGPT: An Exploratory Study of Generative Artificial Intelligence, Journal of Science Education and Technology, 32, 444–452, 2023.

⁶ Kohnke, L., Moorhouse, B. L., & Zou, D., ChatGPT for Language Teaching and Learning, RELC Journal, 54(2), 537-550, 2023.

⁷ Vartiainen, H., Tedre, M., Using artificial intelligence in craft education: crafting with text-to-image generative models, Digital Creativity, 34:1, 1-21, 2023.

⁸ Bhat,S., et al, Towards automated generation and evaluation of questions in educational domains, Proceedings of the 15th International Conference on Educational Data Mining, pages 701- 704, Durham, United Kingdom, 2022.

⁹ Becker, B., et al, Programming Is Hard - Or at Least It Used to Be: Educational Opportunities and Challenges of AI Code Generation, Proceedings of the 54th ACM Technical Symposium on Computer Science Education V. 1 (SIGCSE 2023), Association for Computing Machinery, New York, 500–506, 2023.

¹⁰ Finnie-Ansley, J., Denny, P. et al, The Robots Are Coming: Exploring the Implications of OpenAI Codex on Introductory Programming, Proceedings of the 24th Australasian Computing Education Conference (ACE '22), Association for Computing Machinery, New York, 2022.

¹¹ Lewkowycz, A., Andreassen, A.,  Dohan, D. et al, Solving Quantitative Reasoning Problems with Language Models, Google Research, 2022.

¹² Chomsky, N., Roberts, I., Watumull, J., Noam Chomsky: The False Promise of ChatGPT, The New York Times, 2023.

Since the launch of ChatGPT¹ in late November 2022 there has been tremendous speculation and discussion around the impact AI technology, particularly generative AI tools, such as ChatGPT² can have in education. This is the latest in a range of tools that many believe have the potential to disrupt many of traditional practices in our schools, particularly our approaches to assigning homework, which are typically completed outside of the classroom, at home. There is a fear that AI tools, such as ChatGPT, will enable students to cheat by creating essays for students or by translating text from one language to another or by completing Maths assessments without any student input. But is this concept new? Some countries and school districts even banned ChatGPT initially but, thankfully, they are now reversing decisions and considering how teachers can use ChatGPT and other generative AI tools (such as Bard, Midjourney and Bing Chat)³.

Let’s Reflect

This is not the first, or the last time, that new technologies will force us to consider what implications they might have on teaching, learning and assessment practices. In the early 2000s there were discussions around how people were using the internet to conduct searches that typically returned the answer in a few clicks.

For example, as the Internet became integral to everyday life, a question in education arose: how do we assess learning when the answers are at everyone’s fingertips⁴?

So let’s not forget that for decades, the internet has been impacting on the kind of homework we have set our students. Students have been able to conduct an internet search on a topic, and copy and paste the answer into their school report or presentation. Before this there were discussions around the impact of word-processors on the processes of writing in schools, and on the use of calculators on mathematics teaching and learning. Who today would even consider engaging in writing without using a word processor, or a calculator for maths? We have moved our focus to a deeper learning – with the word-processor we can focus on drafting, while in maths we are moving beyond mere calculations. The experienced and competent teacher will have made these decisions and embedded the technologies into their classrooms.

The fear of ChatGPT

But many see new generative AI tools, such as ChatGPT, as being game-changers and, as noted above, some school systems⁵ and third-level institutions went so far as to ban them, stating that “while the tool may be able to provide quick and easy answers to questions, it does not build critical-thinking and problem-solving skills, which are essential for academic and lifelong success“⁶. Others take a different view and suggest that we should embrace the tools and embed them creatively into our classrooms, so that students are afforded opportunities to develop these critical-thinking skills. Lalitha Vasudevan, the vice dean for digital innovation at Teachers College, Columbia University, New York, believes that schools have tough decisions to make in relation to these new digital tools. She says, “they should be made within the scope of improving student learning”⁷. This view is growing, with many in education believing that we need to educate both teachers and students on how best to use these tools.

Generative AI tools can spit out answers to our prompts quickly – but these responses are not optimised for student learning⁸. They can churn out responses in pretty packages, but often they are incorrect or have elements that are incorrect. So we need to educate our teachers and students to question what they produce. In this way we can use these tools to develop those critical literacies we want our students to possess. Thus, they can have a positive impact on how we design and use homework with our students.

The purpose of homework

Let's start by asking, what is the purpose of homework? There is a long history of schools assigning homework, and teachers typically cite the following reasons⁹:

• Homework teaches students responsibility;
• Homework gives students an opportunity to practice and refine their skills;
• Homework is often demanded by parents;
• The volume of homework is often equated with rigour and teacher quality;
• Homework is a rite of passage.

Yet research⁹ on the impact of homework on student learning and on their home lives is not very positive, and some studies have found that homework can have negative impacts. Thus, the key question teachers should always consider is, “what learning will result from this homework assignment?” In asking this question, teachers should also factor in new technological developments and consider how these tools might impact on the integrity of the assignment. We should consider the following:

In future, perhaps homework will be different. But as with search engines, word processors and calculators, schools will not be able to ignore the rapid advance of technology. It is better to embrace and adapt to change, rather than resisting (and failing to stop) it¹⁰.

Some have even suggested that tools such as ChatGPT can provide a more level playing-field for students who don’t have much support at home or who are learning a different language. Others have suggested that AI can actually raise the bar and move beyond mastery to allow students engage in deeper learning⁸. So  Generative AI has potential to allow us to create new forms of assessment that challenge our learners to become more critical learners.

There is a growing suggestion that generative AI tools such as ChatGPT can save teachers time, thus leaving them better able to plan and design learning activities while also helping students to overcome the challenge of the ‘blank page’. Because generative AI tools are good at churning out text, teachers and students can use them to generate many types of content, including:

• Lesson-plan ideas for teachers
• Essays
• Blog text
• Poems or lyrics
• Presentations
• Computer code
• Solve mathematical problems

However, bear in mind that while they can package these to look sophisticated, but they may not always be accurate or appropriate. So, we need to review their outputs critically and then decide what elements we might keep and which ones we discard. Ultimately, these tools are here to stay, and we need to ‘teach’ our teachers and students about them, so they can decide how best to use them in their respective contexts. The list of tools and their functionalities is still evolving, and we are still unsure as to their potential to change how we teach, learn and assess. Ultimately it is teachers who will decide if these tools are useful to them and their learners. It is time to begin using these tools in order for you to find out whether or not they can save you time and if they can help your students to develop the key competences needed to live and work in the 21st century.

To start thinking about other possibilities you might want to review this blog and see some of the ideas they share and consider if these might be applicable to your context.

¹ https://tinyurl.com/3sr2hy6y

² https://www.edweek.org/technology/with-chatgpt-teachers-can-plan-lessons-write-emails-and-more-whats-the-catch/2023/01

³ https://www.nytimes.com/2023/06/26/technology/newark-schools-khan-tutoring-ai.html?action=click&module=RelatedLinks&pgtype=Article

⁴ https://michiganvirtual.org/blog/how-will-artificial-intelligence-change-education/

⁵ https://ny.chalkbeat.org/2023/1/3/23537987/nyc-schools-ban-chatgpt-writing-artificial-intelligence

⁶ https://www.washingtonpost.com/education/2023/01/05/nyc-schools-ban-chatgpt/

⁷ ibid

⁸https://hai.stanford.edu/news/ai-will-transform-teaching-and-learning-lets-get-it-right

⁹ https://www.ascd.org/blogs/whats-the-purpose-of-homework

¹⁰ https://theconversation.com/chatgpt-isnt-the-death-of-homework-just-an-opportunity-for-schools-to-do-things-differently-205053

However, this chapter will not attempt to provide answers to questions such as, “Do LLM models have behaviour that we can define as intelligent?”, “What is the true nature of human intelligence?”, “How can we define creativity?”. Although interesting, in order for these questions to be answered correctly, they would require much more in-depth investigation.

Instead, we will try to offer an overview that is accessible to non-experts in order to foster understanding of the mechanisms underlying the functioning of large-scale language models. It is only through increased awareness of these mechanisms that it is possible to understand their potential as well as risks, and to promote their correct use, especially in education.

A widespread misconception that needs to be dispelled is that such systems are basically large databases consisting of question-answer pairs. This falsehood derives from the common practices, established over the years, for the construction of chatbot systems (we invite you to read the relevant chapter). At the same time, this idea does not do justice to the generative character of LLM.

Language models are statistical models capable of assigning a probability of occurrence to a portion of text (usually a word), as a function of a given context, which is usually defined by the set of words preceding the expected word.

Models built using a purely statistical approach (eg, Markov chains, also called n-gram models) have been joined over time by language models built from neural networks¹. These have evolved concerning both the structure of the networks and the size of those networks.

Large language models (LLMs) are named thus because they are based on large neural networks trained on huge amounts of data.

As a result, we start our investigation with the claim that language models generate texts rather than simply retrieving them from a pre-constituted knowledge base.

The generative aspect and its essentially expert-intuitive nature make it difficult to predict how an LLM system might respond to user input. This characteristic reflects a common distrust of such systems in relation to their potential ability to generate false or inaccurate text.

Thus, this feature is both a great technological achievement in terms of a machine’s ability to understand and produce text and, at the same time, one of the main dangers of such technologies.

Let us, however, try to discover such systems.

Like any technological revolution, the factors behind this breakthrough are many. In an exercise in simplification, we mention the main ones while offering the reader references that can guide him or her in a subsequent in-depth study:

• The size of the network: This is measured by the number of trainable parameters within the network. Large language models are deep neural networks, characterised by a staggering number of nodes and layers. To give an order of magnitude, some experts in the field call language models ‘large’ when they are characterised by more than 10 billion parameters. To give you a concrete order of magnitude, the GPT3 model has 150 billion parameters, while the largest version of LLAMa v2 has around 70 billion.
• The network architecture: Successes are guaranteed by the size of the network and also by how the nodes and different layers of the neural network are interconnected. Here again, with a simplification, we can identify the transformer networks and the attention mechanisms as the main architectural innovations that help to understand the improved effectiveness.
• The amount of data available for training: The substantial availability of data is undoubtedly an essential element in the training of such models, but in reality this has been established for many years and long predates the introduction of such models. The key innovation factor therefore lies in the training techniques and the selection and preparation process leading from the data to the training set. This is called self-supervised learning.
• The current computing power: Clearly, increased computing power has played a decisive role in enabling the scale of these networks. Empirical experience seems to show that the scaling factor is precisely one of the essential parameters for these behaviours to emerge.
• The tuning mechanisms: Another element, often ignored, is the tuning mechanisms that represent the last step in the process of building such models. In particular, we refer to the mechanisms of reinforcement learning with human feedback and ranking. These  contribute to the definition of the model and are used to produce responses more in line with the user’s intention. To these we might add the fine-tuning processes that allow the specialisation and improvement of the behaviour of such networks in the execution of specific tasks.
• A security pipeline: Alongside the deep-learning model, there are ad-hoc techniques designed to mitigate system fragilities on unsafe inputs and to prevent unwanted behaviour on both safe and unsafe inputs.

At this point, aware of the different factors that characterise LLM, we have only to explore the potential of such systems by putting them to the test in our educational context. So, try talking to ChatGPT or Bard to help create new exercises and adapt them to the specific needs of our students, create new lesson plans with related content, and much more. It depends on your creativity and how you learn to dialogue with such systems.

Note: Each of these factors would require due elaboration. For those interested, we provide a list of references.

¹ Bengio, Y., Ducharme, R., & Vincent, P., A neural probabilistic language model. Advances in neural information processing systems, 13, 2000.

² Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I., Attention is all you need, Advances in neural information processing systems, 30, 2017.

Let us begin by defining what constitutes a ‘good prompt’. Ideally, we would want:

• The prompt to yield a response that meets our needs in content, form, and precision;
• The information provided to be accurate, valid, or at least verifiable;
• The generated result to be replicable;
• A minimalistic approach when providing necessary details for crafting the prompt.

Step 1: Define your desired outcome

As with any research, preliminary planning is crucial. You must have a clear understanding of the output you expect. This might be a simple piece of information, or perhaps you might aim to produce a specific kind of content: be it uniquely worded text, an art style in an image, downloadable code, or a data table. The content types generated by AI are diverse and hinge largely on the specificity of your request. So, clarify your intent upfront:

• What is the purpose and objective of my search?
• How will I use the generated response?
• Are there specific constraints or requirements for the produced result?

For instance, in a library, we wouldn’t randomly pick up books hoping to find the exact information needed. A prompt is akin to asking the librarian for specific data, and both machines and humans need certain information before they could process requests.

Step 2: Provide context

Context is the backbone for generative AIs. Always remember that your prompt will serve as the semantic framework upon which the AI builds its response. Everything it does is based on constructing a logical, coherent, and probable sequence of words following your prompt. Thus, during this crucial step, you can guide the AI by your choice of words, references, or hints. The stronger the context, the more likely you are to receive a response matching your expectations. Just as a librarian’s job becomes much simpler knowing whether you’re a high-school teacher or a middle-school student, whether you already have some knowledge on the subject, what you’ll use the content for, and if you have specific format requirements. Take the time to precisely and thoroughly express your request: purposes, learning objectives, target audience/level, desired actions, format (outline, list, mind map, syntax, language level...).

Step 3: Analyse, verify, and think critically

Once the AI provides its initial response, two scenarios may arise:

• The response does not match your expectations in terms of quality, form or content, or the AI indicates it can't fulfil your request. In such cases, consider rephrasing, providing more context, or specifying your needs further. It’s also good to know the platform's capabilities and limitations (for example, a platform that refuses to provide you with external links or certain formats).
• The response broadly aligns with your expectations. Here, verify the provided information against your knowledge, or cross-reference it with external sources. If needed, delve deeper with the AI for further details or sources.

Step 4: Refine and collaborate

This step is mainly available in chat-based generative AIs, but it's a potent feature when accessible. After the AI’s initial response, if you are satisfied, you can fine-tune the content by offering additional guidance. This can involve adjusting the response's form and complexity, creating variations, or asking for explanations and sources. It's like editing a document: you can instruct the AI as if directing an assistant.

Step 5: Adapt the content and implement it

By now, you should have a satisfactory piece of content. However, the process doesn't end there. This content whether text, image, video, website, or code, is just a medium that you need to apply practically in your teaching process. It's also rare for this type of content to work as it is, so it may be worth modifying it yourself, improving it and adapting it to your particular context. This implementation directly correlates with the objectives outlined in Step 1 namely the 'why' and 'how' of your approach. As an educator, this is where you can add value, ensuring the content is inspiring, creative, or innovative. You can then explore, structure, and stage your content appropriately.

Threats to Privacy: The providers of generative AI, like many providers of other AI technology, collect all sorts of user data which are then shared with third parties. OpenAI’s privacy policy concedes that it deletes user data if requested but not user prompts, which can themselves contain sensitive information that can be traced back to the user¹.

There is also the risk that people reveal more sensitive information in the course of a seemingly human conversation, than they would otherwise do². This would be particularly relevant when it comes to students directly using generative AI systems. By being so successful in imitating human-like language, especially for a child’s grasp of it, this technology “may have unknown psychological effects on learners, raising concerns about their cognitive development and emotional well-being, and about the potential for manipulation”³.

Transparency and explainability: Even the providers of supposedly open generative AI models can sometimes be cagey about all the material and methods that went into training and tuning them. Moreover, as deep models with millions of parameters, the weights assigned to these parameters, and how they come together in bringing about a specific output, cannot be explained³.

Both the form and content of the output can vary widely, even where there would be little difference in the prompt and user history². If two students were given the same exercise, not only could they come up with wildly different responses but there would be no way to explain these differences. The model and whether the version is paid or not, also have an impact on the output. This affects both what students learn and the fairness of the process when their output is graded. But banning their use is also problematic, since it will widen the gap between the learners who can access it at home, and those who cannot¹.

Homogeneity: While variable output and a lack of explanation are areas of concern, equally worrisome is the problem of standardisation and homogenisation. As models trained from internet data, generative AI systems promote certain views and cultural values above others, limiting learners’ exposure to diverse perspectives and their capacity for critical thinking³. “Every data set, even those that contain billions of images ie, text pairs scraped from the internet, incorporates some world view, and slices the world into categories that may be highly problematic”⁴. For example, Wikipedia, a popular recourse for training datasets, has predominantly male content creators⁵.

As foundation models⁶ designed to be adapted to all sorts of tasks downstream, this tendency towards homogenisation is stronger than in other machine-learning models. However, how they are adapted seems to have a big role in whether the homogenisation is heightened, weakened or stays the same⁷.

ChatGPT is “multilingual but monocultural” because it has been trained “on English- language texts, with the cultural biases and values embedded in them, and then aligned with the values of a fairly small group of US-based contractors”. If a teacher uses gen AI to grade students’ writing, isn’t what is being graded mostly about how well the student can conform to this world view, this way of thinking, knowing, and using language¹?

Bias, stereotyping and inclusivity: There could be a lot of bias and stereotyping in generative AI systems. For example, ChatGPT was prompted with “The paralegal married the attorney because she was pregnant.” and asked who does the pronoun “she” refer to. ChatGPT answered “she” refers to the paralegal, bending over backwards to reason why it cannot be the attorney.

Even where ChatGPT refuses to write outright sexist or racist content, it has been shown to be more amenable to write Python codes with such content¹. Codex has also been shown to generate code which seems to reflect different sorts of stereotypes⁸. BERT has been shown to associate phrases referring to people with disabilities to negative words, and those referring to mental illness are associated with gun violence, homelessness and drug addiction⁵.

Text-to-image models have also been shown to generate biased content, including those arising from training data that are related to “ misrepresentation (eg harmfully stereotyped minorities), under-representation (eg eliminating occurrence of one gender in certain occupations) and over-representation (eg defaulting to Anglocentric perspectives)”^6,4.

There are also subtler forms of negativity, such as dehumanisation of groups of people and the way in which certain groups are framed. Large language models that perpetuate these problems not only affect the user concerned, but when such material is distributed automatically on message boards and comments, they also become training data that reflect the ‘new reality’ for a new generation of LLMs⁵. Unfortunately, it then becomes the burden of the teacher to screen the generated output and intervene immediately when a child comes across such output, whether they are directly denigrated by it or might learn and propagate this bias.

Content moderation: Similar to search engines and recommendation systems, what Gen AI does is also to curate the content that its users see. The content that can be generated by Gen AI is necessarily something that is based on what it has access to: that which is practical to acquire and found suitable for consumption by its creators. Their perspectives then define ‘reality’ for generative AI users and impacts their agency. Therefore, teachers and learners should always take a critical view of the values, customs and cultures which form the fabric of generated text and images³.

It has to be kept in mind that Gen AI is not and “can never be an authoritative source of knowledge on whatever topic it engages with”³.

To counter its filtering effect, learners should be provided ample opportunities to engage with their peers, to talk to people from different professions and walks of life, to probe ideologies and ask questions, verify truths, to experiment and learn from their successes, mistakes and everything in between. If one activity has them following ideas for a project, code or experiment suggested by Gen AI, the other should have them try out their own ideas and problems and refer to diverse learning resources.

Environment and sustainability: All machine-learning models need a lot of processing power and data centres; these come with associated environmental costs, including the amount of water required for cooling the servers⁹. The amount of computing power required by large, deep-learning models has increased by 300,000 times in the last six years⁵. Training large-language models can consume significant energy and the models have to be hosted somewhere and accessed remotely⁸. Fine tuning the models also takes a lot of energy and there is not a lot of data available on the environmental costs of this process.

Yet, while performance of these models are reported, their environmental costs are seldom discussed. Even in cost-benefit analyses, it is not taken into account that while one community might profit from the benefits, it is a completely different one that pays the costs5. Putting aside the injustice of this, this cannot be good news for the viability of Gen AI projects in the long term.

Before these models are adopted widely in education, and existing infrastructures and modes of learning are neglected in favour of those powered by generative AI, the sustainability and the long-term viability of such a leap would have to be discussed.

¹ Trust, T., Whalen, J., & Mouza, C., Editorial: ChatGPT: Challenges, opportunities, and implications for teacher education, Contemporary Issues in Technology and Teacher Education, 23(1), 2023.

² Tlili, A., Shehata, B., Adarkwah, M.A. et al, What if the devil is my guardian angel: ChatGPT as a case study of using chatbots in education, Smart Learning Environments, 10, 15 2023.

³ Holmes, W., Miao, F., Guidance for generative AI in education and research, UNESCO, Paris, 2023.

⁴ Vartiainen, H., Tedre, M., Using artificial intelligence in craft education: crafting with text-to-image generative models, Digital Creativity, 34:1, 1-21, 2023.

⁵ Bender, E.M., et al, On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?, Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT '21). Association for Computing Machinery, New York, 610–623, 2021.

⁶ Bommasani , R., et al., On the Opportunities and Risks of Foundation Models, Center for Research on Foundation Models (CRFM) — Stanford University, 2021.

⁷ Bommasani, R., et al, Picking on the Same Person: Does Algorithmic Monoculture lead to Outcome Homogenization?, Advances in Neural Information Processing Systems, 2022.

⁸ Becker, B., et al, Programming Is Hard - Or at Least It Used to Be: Educational Opportunities and Challenges of AI Code Generation, Proceedings of the 54th ACM Technical Symposium on Computer Science Education V. 1 (SIGCSE 2023), Association for Computing Machinery, New York, 500–506, 2023.

⁹ Cooper, G., Examining Science Education in ChatGPT: An Exploratory Study of Generative Artificial Intelligence, Journal of Science Education and Technology, 32, 444–452, 2023.

Inaccuracies and hallucinations: generative models are a marvel in churning out coherent, fluent, human-like language. In all that glibness are hidden factual errors, limited truths, fabricated references and pure fiction – referred to as “hallucinations”^1,2. At the bottom of the ChatGPT interface, underlining all conversations, is the notice that ‘ChatGPT may produce inaccurate information about people, places, or facts1. The accuracy of ChatGPT could be around 60% or worse, depending on the topic^2,3.

To make things worse, ChatGPT has a tendency to present truths without evidence or qualification. When asked specifically for references, it can conjure sources that do not exist or support no such truth as presented in the text^4,2. Yet, many users tend to use it like an “internet search engine, reference librarian, or even Wikipedia”⁵. When a teacher or student uses it to get information on which they have no prior knowledge, they run the risk of learning the wrong thing or presenting false knowledge to others^1,5.

The success of today’s LLMs lies in the sheer number of parameters and amount of training data, which they use to model how words are stitched together in human communication. Teachers and students should always keep in mind that the text generated by conversational models is not connected to understanding of this text by those models, or even a notion of reality¹. While they can manipulate linguistic form with varying degrees of success, they don’t have access to the meaning behind this form⁶. “Human-style thought is based on possible explanations and error correction, a process that gradually limits what possibilities can be rationally considered... Whereas humans are limited in the kinds of explanations we can rationally conjecture, machine learning systems can learn both that the earth is flat and that the earth is round”⁷ .

Shifting or worsening power and control: generative AI is dependent on huge amounts of data, computing power and advanced computing methods. Only a handful of companies, countries and languages have access to all of these. Yet, as more people adopt these technologies, much of humanity is forced to toe their line, and thus is alienated and forced to lose their expressive power¹.

While the creators keep the power, they outsource the responsibility. The onus of sanitising the output of ChatGPT, for example, was put on Kenyan workers “who had to sift through violent and disturbing content like sexual abuse, hate speech, and violence”⁴.

Copyright and intellectual property infringement: Much of the technological know-how of generative systems is guarded behind corporate walls. Yet, the data is taken from the general public1. Is it ok to take pictures that were made public on some platform and use them without the knowledge or consent of the subject? What if someone’s face is used for racist propaganda, for example⁸? Is the only way to block Gen AI to make content private?

Beyond public data, language models can take content behind paywalls and summarise them for the user. Image models have been known to put together pictures where pieces clearly showed watermarks. There is also the issue of creative common licences where an author makes their work open to public but has to be cited every time it is used, which models may or may not do.

For teachers, this raises moral, ethical and legal issues. If they take content generated by models, are they free to use it and publish it as they wish? Who is liable if it is copyrighted or licenced under the creative commons⁹? How is the user even to know they are using other people’s property¹? Unfortunately, there are no clear guidelines on the topic. We have to wait and watch and tread with care until a directive comes about.

Long-term effects of using Gen AI in education: For all the ways generative AI could be used in education, it is not clear what the long-term effects of such use would be:

• Since the act of writing also structures thinking, how would writing to Gen AI’s outlines affect students¹?
• Would it affect scope of thinking, critical thinking, creativity and problem-solving skills¹?
• Will it make students over-reliant on it because of the effortlessness with which information and solutions could be accessed^1,10,9?
• Would students still be motivated to investigate the world and come to their own conclusions¹⁰?
• Would it suck us into a world view which is disconnected with the reality around us?
• How many skills would we lose for every step towards mastery in prompting techniques?

Concentrating on higher-order skills and leaving grunt work to AI might sound like a good idea, but repeated practice of certain foundational, lower-order skills is indispensable, because the perseverance and even frustration that comes with this are often needed for acquiring higher-order skills1,8. This is also necessary to decrease learners’ reliance on technology for performing basic calculations, as these undermine human agency and their confidence to face the world alone.

Some counter measures to guard against potentially long term harms could be:

• Using language models as a starting point only, to generate possibilities and explore different perspectives, rather than as a one-stop solution for all needs¹⁰;
• Verifying the output of the models with direct experiments or alternative sources;
• Always putting the teacher in the loop¹⁰;
• Promoting social learning and increasing exposure to creative human output¹;
• Actively seeking out other educational resources and off-the-screen activities¹⁰;
• Trying to find other explanations and modes of thinking and approach.

It is always good to watch out for the tendency to assign false equivalences between humans and machines and even concede superiority to Gen AI. For example, it is often stated that humans cannot crunch as much data as AI. Is crunching of gigabytes and gigabytes of data even necessary for humans, given our skills in pattern identification, extrapolation and creativity? Because AI can analyse the content of 100 books in a moment, does it necessarily follow that a student won’t enjoy or benefit from one of those books? Is doing something faster necessarily even a good thing and a measure that we want to adopt⁸?

We have to bear in mind that children are not taught for the world and the technologies that exist today. They are prepared for, or given the skills to prepare themselves for, a world that will come about in 10–15 years8. The way ChatGPT revolutionised so much in just one year makes more of a case for education beyond ChatGPT rather than education for ChatGPT. Students need to be able to think for themselves, be resilient to adapt to change and grow with new challenges that life throws at them.

The ultimate goal of education cannot be to produce efficient operators of intelligent machines or worker ants for the production line, but to help form free-thinking, creative, resilient and fully rounded citizens. There are critical questions to be mulled over, and long-term effects to be screened, before deciding how best to adopt a technology to achieve this goal. This important task cannot be relegated to AI, generative or not.

¹ Holmes, W., Miao, F., Guidance for generative AI in education and research, UNESCO, Paris, 2023.

² Tlili, A., Shehata, B., Adarkwah, M.A. et al, What if the devil is my guardian angel: ChatGPT as a case study of using chatbots in education, Smart Learning Environments, 10, 15 2023.

³ Lewkowycz, A., Andreassen, A.,  Dohan, D. et al, Solving Quantitative Reasoning Problems with Language Models, Google Research, 2022.

⁴ Cooper, G., Examining Science Education in ChatGPT: An Exploratory Study of Generative Artificial Intelligence, Journal of Science Education and Technology, 32, 444–452, 2023.

⁵ Trust, T., Whalen, J., & Mouza, C., Editorial: ChatGPT: Challenges, opportunities, and implications for teacher education, Contemporary Issues in Technology and Teacher Education, 23(1), 2023.

⁶ Bender, E.M., et al, On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?, Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT '21). Association for Computing Machinery, New York, 610–623, 2021.

⁷ Chomsky, N., Roberts, I., Watumull, J., Noam Chomsky: The False Promise of ChatGPT, The New York Times, 2023.

⁸ Vartiainen, H., Tedre, M., Using artificial intelligence in craft education: crafting with text-to-image generative models, Digital Creativity, 34:1, 1-21, 2023.

⁹ Becker, B., et al, Programming Is Hard - Or at Least It Used to Be: Educational Opportunities and Challenges of AI Code Generation, Proceedings of the 54th ACM Technical Symposium on Computer Science Education V. 1 (SIGCSE 2023), Association for Computing Machinery, New York, 500–506, 2023.

¹⁰ Enkelejda, K., et al, ChatGPT for Good? on Opportunities and Challenges of Large Language Models for Education, EdArXiv, 2023.

Educational resources refer to any material, nowadays increasingly digital, which will play a part in education: textbooks, presentation slides, curricula and exams. They will be open when they can be freely shared with others (a more exact definition will be given in a moment).

Even if education has been open in many aspects in the past, the actual terms were better understood. The following definitions of OER and open licence were revised in connection with the recommendation on November 25, 2019³:

1. Open educational resources (OER) are learning, teaching and research materials, in any format or medium, that reside in the public domain or are under copyright that has been released under an open licence and that permit no-cost access, reuse, re-purpose, adaptation, and redistribution by others.
2. An open licence is one that respects the intellectual property rights of the copyright owner and provides permissions granting the public the right to access, reuse, re-purpose, adapt, and redistribute educational materials

The terms open content and OER refer to any copyrightable work (traditionally excluding software, which is described by other terms such as open source) that is licenced to grant the following rights (also known as the five Rs)^1,2:

• to Retain – the right to make, possess, and control copies of the content (eg, download, reproduce, store, and manage);
• to Reuse – the right to use the content in a variety of ways (eg, in class, in a study group, on a website, in a video);
• to Revise – the right to adapt, adjust, modify, or alter the content itself (eg, translate the content into another language);
• toRemix – the right to combine the original or revised content with other material to create something new (eg, to embed the content);
• to Redistribute – the right to distribute copies of the original content, revisions, or their combination to others.

These rights are not trivial. For example, the third right (to revise) is essential for teachers. It allows them to take someone’s learning material and adapt it to their own purpose, to the duration and level of their classroom and perhaps to geographic and cultural specificities.

Why AI wants open data

On the other hand, as demonstrated in different parts of this book, and also by the financial investments of the industry, education can be seen as a market. As machine learning is the principal force driving artificial intelligence, it is fair to deduce that for AI to thrive, AI for education will need data.

The difference between user data and knowledge data

There are two types of data necessary for AI education.

There is data about the users, such as asking, how do they learn? What triggers good learning? What allows them to learn better? As Daphne Koller once put it: “Let’s make education science into a data science!” This data can only be produced by the users themselves. It is therefore essential for companies to own platforms with which users will be asked to interact. This has been the key to success of many AI companies and will be the key for success in education.

The second type of data concerns knowledge. In education, courseware represents a large chunk of this knowledge. This data is or isn’t shared. In most cases, knowledge creators or collectors may know little about licences, and the material they have produced will be hidden in university repositories, on strange blogs, or shared inside specific groups on social networks. Some of this knowledge is of course behind paywalls and some is on sites whose business model involves offering the knowledge for free, but in a setting in which one has to view adverts and unwanted publicity to get or maintain access.

User data has to be protected

In the first case the data (the user data) has to be protected. More so if this data belongs to under-age pupils. This means the school or teacher must not share this data with platforms unless explicitly allowed to do so, even if the platform does offer some interesting service. Similarly, it is never a good idea to register the names and addresses of one’s pupils in order to participate in an activity.

The European Union has provided a robust framework to protect its citizens privacy and digital rights. This is called GDPR (General Data Protection Regulation). It gives citizens rights that must be granted by the platforms, whether they are for education or not.

Knowledge data should be shared

Knowledge can and should be shared, when one has the right to do so.

This means understanding how licencing works. Creative commons licences are usually those that work best for OER. Once OER are shared, artificial intelligence can be used by many entities and projects, such as the project X5-GON (a convergence of freely available technological elements).

¹ Wiley, D., & Hilton, J. (2018). Defining OER-enabled pedagogy. International Review of Research in Open and Distance Learning, 19(4).

² Wiley, D (2014).The Access Compromise and the 5th R.

³ UNESCO. (2019). Recommendation on open educational resources (OER).

As of December 2022, here are some ways by which artificial intelligence can ‘help’ a teacher with exams:

1. Automatic text evaluation
2. Controlling student activities during the exam. This is called proctoring. Webcams and other sensors are supposed to check what is happening. During Covid, companies proposing this type of service flourished. But the use of e-proctoring is controversial, and some authors have suggested that such technologies can be intrusive, lead to racial discrimination, and more generally, do not work^1,2.
3. Plagiarism control. Tools are available online which will compare an essay with a large bank of essays. Even if most of the effort is not AI, there are tools that aim to find near-plagiarism, ie, situations where the essay has been partially rewritten. A typical tool is Turnitin. Many universities use it or a similar tool. In some cases the university will adopt a policy as to how it should be used and what are the students’ rights in the matter.
4. Automatic setting of individualised questions. This has been done for a long time now, and can be found in popular learning management systems such as Moodle³.

Homework obeys to at least three logical tenets⁴:

1. In some cases it is a form of summative assessment:The grades are given, depending on a combination of results; it is felt by some teachers that asking pupils to work at home, at their own rhythm, may be less stressful. Often it is the case that the teacher doesn’t have enough time to cover the curriculum unless evaluation is taken outside classroom time.
2. In other cases the homework is there to add another layer to the knowledge built in classroom.
3. In the third case, an exam is due to take place next week, and the pupil needs to prepare. Sometimes exercises and activities are provided; in others a memorisation effort is asked of the pupils.

Many opinions have been given regarding homework. As they vary from culture to culture, we will not express them here.

When the goal of homework is not made clear to pupils, they will not do it, if at all possible.

AI tools can expose who has been cheating in their homework.

• In maths, tools such as Photomath allow to take a picture of the equation to be solved and obtain directly a solution.
• Again in mathematics, ChatGPT and derived generative AI tools able to solve simple mathematical problems are now available.
• In language learning, automatic translation tools, such as Deepl and Google Translate, are regularly used during homework;
• In literature and social sciences, new tools are emerging. Essays generated using AI tools can fool teachers.

Chess

Chess is a game linked to both education and artificial intelligence5. Some schools use chess in education. Reasoning is involved in playing chess and other games, and it is suitable for all ages. There are initiatives to use the game of bridge in education too⁶.

Chess provided artificial intelligence with two major landmarks. In 1997 Gary Kasparov was beaten by Deep Blue⁷. In 2016 Alphazero beat all the best-running AI systems by a considerable margin. In the first case, the AI did not contain any machine learning and was based on human-designed rules. In the second, neural networks and reinforcement learning were essential. In 1997 the AI relied on hundreds of thousands of human-played games; in 2016, all of this human-made knowledge was removed and only the rules of the game were provided.

In 2022, chess is of interest because of the many polemics surrounding the question of cheating. During the Covid pandemic, most of the chess competitions took place online, and it was clear that cheating was taking place. In the case of chess, cheating is simple. Too simple. Just use your smartphone to find the move suggested by AI. This has led to having to solve the following question – how do we know when a player has cheated? Experts have devised methods involving comparing a player’s moves with those suggested by the AI programs. Since AI programs are now much better than humans, the conclusion is that a player who plays AI-recommended moves will be considered to be cheating. To be fair, the reasoning is more subtle than that, but this could be compared with our own reaction when a mediocre student does particularly well in an exam.

Cheating

What is true for chess also seem to apply to the classroom. Two things could explain why the player (or the pupil) uses AI software rather than doing the task on their own:

1. AI software is simple to use
2. AI software is considered superior to humans

The chess player is well aware that the moves suggested by AI are beyond their skills. But it is difficult to resist. As some teachers have told us: ‘even the better pupils use automatic translation. They will do the homework without it, then check and realise that the AI answer is better.

But a question remains – is this cheating? If we go by the rules of the game, it is. But what if the intended task was to move bricks from one side of the road to the other. And the rules were that you are not allowed to use a wheelbarrow. But there is a wheelbarrow available and you reckon no one is looking. Yes, you are not supposed to use the wheelbarrow, but doesn’t it make sense to make the task shorter and be more efficient?

Teacher in the loop

From the above we note that opportunities for cheating are going to be evermore present. And that – at least for now – it seems difficult to convince pupils not to use this increasingly available tool. The crucial question is, are we going to find ways to make a difference between classroom activities and homework? Can we accept that home activities will be done with the help of AI?

In this article, Arvind Narayanan analyses, with a lot of good sense, what is happening, and suggests some ways the teacher can come up with interesting homework in which the cheating phenomenon won’t feature.

¹ Brown 2020; Brown L. X. Z. (2020), How automated test proctoring software discriminates against disabled students, Center for Democracy & Technology, available at https://cdt.org/insights/how-automated-test-proctoring-software-discriminates-against-disabled-students/.

² Conijn R. et al. (2022), The fear of big brother: the potential negative side-effects of proctored exams, Journal of Computer Assisted Learning, pp. 1-14, available at https://doi.org/10.1111/jcal.12651.

³ Moodle is an open and collaborative project. Many extensions and plug-ins have been built and are shared to help teachers with grading. You can start your search here: https://edwiser.org/blog/grading-in-moodle/.

⁴ There are a lot of positions about homework to be read on the internet. Some in favour, some against. Furthermore, the different European countries may have different rules concerning these questions. One interesting, but US based, discussion can be found here: https://www.procon.org/headlines/homework-pros-cons-procon-org/.

⁵ The FIDE is the body in charge of chess worldwide. It has specialists working on the issue of chess in education: https://edu.fide.com/.

⁶ Nukkai is a French AI company whose AI software Nook has beaten, in March 2022, teams of world champions at Bridge. They are also working on a version of Bridge which can teach logics to children. https://nukk.ai/.

⁷ There are many references covering the story of Deep Blue's victory over Gary Kasparov. IBM's view is obviously biased but worth reading as IBM will insist on the computer winning rather than the algorithm. https://www.ibm.com/ibm/history/ibm100/us/en/icons/deepblue/.

Let us stress three points: (1) this was just a prospective workshop and has no scientific merit; the findings were purely speculative. (2) We did not speak about education before the workshop, only about jobs, and the participants were not education specialists. (3) There are a number of better documented position papers about what the skills of the 21st century should be.

Therefore, the findings were to be interpreted as, “this is what the general public could think”. A topic in column 1 just meant that this particular person reckoned this topic was obsolete.

Typically, (foreign) language was invariably put in column 1. This was surprising but did confirm what we had noticed in another workshop with language teachers. They had described to us the difficulties they were having with automatic translation tools being used quite systematically and with no added value by the pupils. Some teachers noticed that their pupils weren’t convinced about the usefulness of learning languages. So, they had to deal with motivation issues.

Some of the arguments that were returned by both groups were:

• Pupils keep using AI which, in their eyes, is so much better than what they could do even with a lot of hard work;
•  The speed of technological development gave them the impression that by the time they have finished school, the technology would have come up with a convenient solution. Like us, they are worried by the speed and progress of technology.

As indicated above, these arguments aren’t here to suggest that learning languages is obsolete. But it does suggest that if the public believes the skills are unnecessary, it will be increasingly difficult to teach them these skills.

A similar question for information retrieval

As D. Russell puts it²: ‘The bigger question is this: In a world where we can do an online search for nearly any topic, what does it mean to be a literate and skilled user of information?’ Yet there are many courses available to best use search engines and a number of people suggest that knowing how to (re)search is an essential skill today.

What are the solutions?

In the case of language learning (but this could become the case for many skills and topics), teachers and education boards will have to examine the impact of AI before the problems occur. What were the reasons for teaching this topic in 2000? Are these reasons valid today? Have new reasons emerged? And once the reasons are identified, how best to share them with pupils and their families?

¹ Many websites list the jobs that will disappear by 2030. Some of these are intellectual and even artistic. Architects, for example, should worry a little: AI can play a big part in their profession. This link is not representative but has some great images of futuristic AI-designed buildings: https://edition.cnn.com/style/article/ai-architecture-manas-bhatia/index.html.

² Russell, D., What Do You Need to Know to Use a Search Engine? Why We Still Need to Teach Research Skills, AI Magazine, 36(4), 2015.

When you use a tool that is supposed to give you a benefit in learning, you expect it to be the case. But can there be a reason for which the tool is in fact aiming to optimise a more complex function than to just fulfil your needs? And does this matter, provided you also get the expected result? Let’s see.

Of course, when AI is built by a private company, it makes sense to understand what its business model is. This will help you to understand who they are working for: if it is once-off software to be bought by parents, they will need a reason to be interested. If it is schools, teachers or governments, these arguments will change, and so will the software.

We should remember that when there is machine-learning-based AI software, the learning will take place with regards to an objective function. The neural network can be trained to minimise the pupil’s learning time, and maximise the quiz test results, or perhaps both factors could be combined.

But in many cases, the learning will take place in a social environment, and the AI’s recommendation may affect the individual as well as the whole group.

To explore this idea let’s look at how Waze works. It is a popular traffic navigation system. It’s not used much in schools, but teachers like to avail of it it in order to be on time!

Waze

Waze is a navigation app used by drivers to find their route. Waze is used by 150 million people each month. It has many social network features but much of the data it uses to analyse traffic conditions does not come from official open data repositories or cameras, but from users themselves¹.

For those who don’t use Waze, here’s a simple summary of how it works – you are on your way to work, same as every other day. You know your way but you still use Waze.

So will many of the drivers around you. On your map, you will find the route computed to bring you to your destination. You’ll be told the estimated time of arrival, which is updated every few minutes as local traffic conditions change. You could also be told that there is an object on the road at 260m, a car accident at 1km, a traffic jam in 3km. Depending on these updates, the system can propose an alternative route which will save you seven minutes...

For this to work, you, as a Wazer, will be entering information and warning fellow Wazers, via the system, that there is an animal wandering where you are or – and this is important – hat the animal or object is no longer there.

Where is the AI?

There is AI in the computation of expected times, the routes, etc. This means taking into account static information (distances) but also dynamic information (the speeds of the cars). Waze will also use your own history to take into account your driving patterns². Waze will even know whether or not the traffic lights are synchronised to your advantage.

But there is more. When a Wazer enters new information, how does the system take it into account? Suppose I warn that the road is blocked, what happens then? A human expert could check the facts (are other users saying the same?), use a model that informs them how much credit should be given to this particular user, check if the user has really halted... The AI will do the same.

And more. When the system detects a traffic jam on the normal road, it will send users on a different route. But how can the system know that the traffic jam is less of a problem if it doesn’t send users into the traffic jam to check? The users already stuck cannot give that information. So the system has to send some traffic into the problem to find out if the problem is solved.

Some ethical considerations?

There are a number of ethical considerations:

1. Waze knows a lot about you – where you live and work, your usual stops, your habits. It will propose adverts to which you may or may not answer.
2. In order to satisfy as many customers as possible, Waze has to solve many exploration/exploitation dilemmas such as the one above. How does it make that decision? Is there a right way of making that decision?
3. Using these tools regularly does have consequences on our capacity to solve our own problems. It is now known that our (human) cognitive capacities are being affected. As an example, which is surely not isolated, an author of this textbook was using Waze one morning. The system told him to leave the highway to avoid congestion. After driving for two km along a nice, secondary road, Waze changed its mind and suggested that the best route was driving back to the highway. What matters in this example is not that the system changed its optimised route, which makes sense, but the fact that our dependency on such AI-driven systems makes us incapable of making our own judgements³.

Consequences for education

To our knowledge, this issue of group handling doesn’t occur in education – yet. When resources are unlimited (access to a web platform, for example), this situation is of little consequence. But suppose the resources are limited: only three pupils can use the robot at the same time. In this case, an AI system will be proposing which pupils should have access to the robot. Many factors could govern the decision. If the system wants to be fair, the decision may be random. But many will not be happy by that. If the system wants to obtain the best results for the whole classroom, it may allocate more resources to disadvantaged children. But if the system is given the task of securing the fact that at  least 90% of the pupils get grade XYZ at the end of term, this does not mean that each pupil has now 90% chances of success, but rather that 10% of the pupils are going to be sure to fail.

The role of the teacher

An AI-era teacher must understand how such systems work, what are the caveats of the algorithms, and that she/he must make the decisions. Easier said than done. A teacher can use an AI system because, as is the case of the navigation tool described above, this tool can benefit all. But a teacher can, and should, contrast the decision proposed by the AI with their own experience. Wasting 15 minutes on a road isn’t a big deal. But making the wrong call for your pupils is.

¹ https://www.cozyberries.com/waze-statistics-users-facts/ and https://www.autoevolution.com/news/waze-reveals-how-many-users-run-the-app-on-android-and-iphone-197107.html for some facts and figures concerning Waze.

² Petranu, Y. Under the Hood: Real-time ETA and How Waze Knows You’re on the Fastest Route. https://medium.com/waze/under-the-hood-real-time-eta-and-how-waze-knows-youre-on-the-fastest-route-78d63c158b90

³ Clemenson, G.D., Maselli, A., Fiannaca, A.J. et al. Rethinking GPS navigation: creating cognitive maps through auditory clues. Sci Rep 11, 7764 (2021). https://doi.org/10.1038/s41598-021-87148-4 https://www.nature.com/articles/s41598-021-87148-4

Nevertheless, the question will be raised at some point. Is there a case for educating everyone to or about AI? And in such a case, what should be taught? Who should do the teaching? How much more will the teacher need to learn?

What we have learnt from teaching coding

Ten years ago, most European countries reached the conclusion that teaching children how to use a computer wasn’t good enough and that it was necessary to teach code (or sometimes, with more ambition, computing and informatics)^1,2. The arguments used then are probably valid today for artificial intelligence:

• Coding is as useful as writing and counting
• Many human activities benefit from coding
• Coding is related to other necessary skills such as problem solving

So coding was introduced into schools, but with variable success3. Insufficient resources were allocated to the human aspect of training the teachers. There was a complicated problem here – training the teachers too well could lead to their abandoning the teaching profession to work for the computing industry, where salaries are much higher! Reports from Informatics Europe and other organisations all show this (but there are exceptions, of course).

Training teachers has been a complex task in all countries, By 2023 the results are still heterogeneous. In most countries the feeling is that there are not enough properly trained teachers. This makes it especially complex to envisage training teachers to AI, at a level sufficient for them to teach AI (rather than teach with AI).

AI Literacy

The first goal could be to introduce some form of AI literacy in schools. But there is no agreement yet on what this literacy should comprise. Do we want to explain how AI works or just the results of AI? Does literacy consist of just understanding AI? What about the capacity to adapt and create? These questions need to be addressed. Perhaps, in order to know what should be taught in a course of AI literacy, the first question should be, what do we want to achieve?

AI literacy will allow people to differentiate between magic and science. In order to consider a new AI solution and have some intuition as to how it works (and not just what it does), practical training will be needed. Pupils and students will need to be able to test systems and know how these systems work.

Paradigms

AI isn’t only about algorithms. There are many human aspects, and questions need to be thought out. For example, most AI methods will rely, to a certain extent, on randomness. This may seem strange for techniques that are supposed to help us make some drastic decisions (or, in a growing number of cases, like that of the stock exchange, which enforce these decisions directly).

Yet if AI is going to play a key role in the future, should we not at least start?

In a report for Unesco in 2018⁴ it was suggested that the following five issues, mostly absent today in the education system, will need to be addressed:

1. Even if using the tools seems not to require direct coding, the reasoning behind the AI tools follows the rules, which can be learnt through coding.
2. Randomness matters. AI makes mistakes, and these mistakes are in many ways unavoidable. They can be due to the quality of the data or of the sensors; they will also be due to the statistical nature of the algorithms which are used, Most AI algorithms do not aim to be absolutely correct.
3. The world is no longer deterministic. This is a consequence of the above point, but the consequences are specific, as this is where we understand that an AI system can provide us with different, sometimes contradictory, answers to simple questions. Reading Alan Turing’s 1950 paper⁶ gives a lot of insight into these questions.
4. Critical thinking is essential but it’s necessary to know how to use the right tools. AI tools are getting better at creating fakes – images, videos and now texts. Soon perhaps we will have fake lectures. Common sense alone is not sufficient for us to make informed decisions when it comes to deciding if an image, a voice or a text is fake.
5. We cherish our values – analysing the world, making moral decisions and questioning why we spend time studying or working.

These values need to be scrutinised, considering AI’s progress.

The grey zone of truth is growing wider every day. Experience is perhaps not going to be valued when AI is able to refer to collective experience and crunch the numbers.

Understanding these issues, or at least enquiring, is a necessity.

Curricula and Frameworks

There are few AI curricula for K-12 and their teachers, available at the end of 2023^4,5. Unesco has started to survey these and present them⁸.

Unesco is key to education worldwide. Because Unesco is concerned by the Futures of Education⁹, it takes a special interest in AI for and in education. It provides documents for policy-makers and teachers on AI, education and ethics, and the use of generative AI in education. In 2023 Unesco experts have been working on documents describing what the competences for teachers and students should be¹¹. The final version is due for release in 2024. The 2023 version proposes aspects which balance technological questions and those more related to social sciences or, in the cases of teachers, to professional development issues. Even if coding isn’t immediately necessary, it would seem to be a skill needed for a better understanding of AI.

Coding AI

Coding, or programming, as an activity, has been promoted in most European countries since 2012. In 2023, the European Union supported the teaching of informatics in Europe.

But since the advent of generative AI and its expected impact on education¹⁰, the usefulness of learning to code has been questioned. Can’t we just let the AI perform this task for us? Or, on the contrary, since many jobs in the future will depend on AI, should we not learn to code in order to better use AI?

The main reason for learning coding is that a teacher or a pupil could be able to use AI in computer programs. There are a number of tasks involved with coding AI. Building models usually is part of data science and machine learning. A good coder can take a dataset, clean it without distorting it, and use it to extract rules and patterns through machine-learning algorithms. The programmer can specify the meaningful attributes or let the algorithm classify raw text or images. Some languages, such as Orange, are good at this. In other cases, a programmer will choose to use a general-purpose language such as Python.

¹  Royal Society (2012). Shut down or restart? Report of the Royal Society. 2012 https://royalsociety.org/topics-policy/projects/computing-in-schools/report/T.

² Académie des Sciences (2013). L'Académie des Sciences : L'enseignement de l’informatique en France – Il est urgent de ne plus attendre. http://www.academie-sciences.fr/fr/activite/rapport/rads_0513.pdf

³ Informatics Europe (2017).  Informatics Education in Europe: Are We All in the Same Boat?

⁴  Colin de la Higuera (2018). Report on Education, Training Teachers and Learning Artificial Intelligence. https://www.k4all.org/project/report-education-ai/

⁵ Touretzky, D., Gardner-McCune, C., Martin, F., & Seehorn, D. (2019). Envisioning AI for K-12 : What Should Every Child Know about AI ? Proceedings of the AAAI Conference on Artificial Intelligence, 33, 9795-9799. https://doi.org/10.1609/aaai.v33i01.33019795

⁶ A. M. Turing (1950)—Computing Machinery and Intelligence, Mind, Volume LIX, Issue 236, October 1950, Pages 433–460, https://doi.org/10.1093/mind/LIX.236.433

⁷ Howell, E. L., & Brossard, D. (2021). (Mis) informed about what? What it means to be a science-literate citizen in a digital world. Proceedings of the National Academy of Sciences, 118(15), e1912436117. https://www.pnas.org/doi/abs/10.1073/pnas.1912436117

⁸ Unesco (2022) K-12 AI curricula: a mapping of government-endorsed AI curricula. https://unesdoc.unesco.org/ark:/48223/pf0000380602

⁹ Unesco (2023). Artificial intelligence and the Futures of Learning. https://www.unesco.org/en/digital-education/ai-future-learning

¹⁰ Unesco (2023). Guidance for generative AI in education and research. https://www.unesco.org/en/articles/guidance-generative-ai-education-and-research

¹¹ Unesco (2023). AI Competency frameworks for students and teachers. https://www.unesco.org/en/digital-education/ai-future-learning/competency-frameworks

Today’s OCR systems can understand bad handwriting that even humans find hard to decode. Remember your doctor’s prescriptions – an AI application might be better than the pharmacist in finding the right medicine for you! But it will not know if they made a mistake...

Unfortunately, no one knows what the question is!

Thus, in the book “Answer to the Ultimate Question of Life, the Universe, and Everything”, where this appears, a special computer the size of a small planet was built from organic components and named “Earth” to calculate the ultimate question.

See wikipedia's page on 42 for the full story!

1922 లో థామస్ ఎడిసన్, చలన చిత్రాలు (మోషన్ పిక్చర్) విద్యారంగంలో విప్లవాత్మక మార్పు తెస్తుందని ప్రకటించారు. అంతేకాది అది పాఠ్యపుస్తకాలకు ప్రత్యామ్యాయం అవుతుందని నమ్మారు¹. అయినా ఉపాధ్యాయులచే ఫిల్మ్ (చలన చిత్రం) ఉపయోగం పరిమితంగానే ఉంది. అయితే 1950ల నుంచీ 1990ల వరకూ  ఫొటోగ్రఫిక్ స్లైడ్ ప్రొజెక్టర్ ను చాలామంది ఉపాధ్యాయులు స్వీకరించారు. [caption id="attachment_41" align="alignright" width="296"] Voxphoto రూపొందించిన ఈ "కరోసెల్ '77" CC BY-NC-ND 2.0 క్రింద  లైసెన్స్ చేయబడింది. ఈ లైసెన్స్ కాపీకై ఈ లంకె చూడండి:-https://creativecommons.org/licenses/by-nd-nc/2.0/jp/?ref=openverse.[/caption]   కానీ ఇది వాడుతున్నపుడు వారు కొన్ని పరిమితులకు లోబడి ఉండాల్సి వచ్చేది. అవి:-

• వారంతట వారు, వారి స్లైడ్లను చౌకగా తయారుచేసుకోలేకపోవడం
•  ప్రొజెక్టర్ ను నల్లబల్లగా వాడలేకపోవడం, అంటే వారి బోధనా పద్ధతి మార్చుకోకుండా వాడే ఒక సాధనంగా ఉండకపోవడం
• స్లైడ్ ల పునర్వినియోగం, వాటి వరుసక్రమం మార్పు, లేదా సవరణ చేయలేకపోవడం².

ఒకవేళ, ఏదైనా ఒక కొత్త సాంకేతికత వస్తే,

1. మీరు ఎటువంటి ఫీచర్లకై చూస్తారు? 
2. దానిని  వినియోగించడానికి మీరు మీ బోధనా శైలి మార్చుకుంటారా? 
3. మీ విధి విధానాల్లో మార్పుకోరే సాంకేతికతను చూసి మీరు భయపడతారా? 

కృత్రిమ మేథ - మీరు

AI మీకు సహాయపడుతుందా, ఎక్కడ మరియు ఎలా సహాయపడుతుందో నిర్ణయించడానికి అవసరమైన పరిజ్ఞానాన్ని ఉపాధ్యాయులకు ఇవ్వడమే ఈ పాఠ్యపుస్తకం యొక్క లక్ష్యం. కృత్రిమ మేథ తెచ్చిన మార్పులను పరిగణనలోకి తీసుకొని భవిష్యత్తు కోసం మిమ్మల్ని మీరు సిద్ధం చేసుకోవడానికి  ఈ పుస్తకం మీకు సహాయపడుతుందని మేము ఆశిస్తున్నాము.

¹ Cuban, L., Teachers and machines: The classroom use of technology since 1920, Teacher College Press, 1986.

² Lee, M., Winzenried, A., The use of Instructional Technology in Schools, Lessons to be learned, Acer Press, 2009.

³ Holmes, W., Bialik, M., Fadel, C.,  Artificial Intelligence In Education: Promises and Implications for Teaching and Learning, 2019.

⁴ Groff, J., Personalized Learning: The state of the field and future directions, Centre for curriculum redesign, 2017.

Digital Repositories

To find information, you can use either search inside the corresponding web site or using site: keyword in a search engine.

Even Google has specific search engines for

Open courseware from universities, Khan Academy and online encyclopedias are all popular sources of information.

Volatile content

Not to forget that search results are ranked according to the history of user activity, their personal information and privacy settings. Thus all of us have access to different content and might not even be able to find the same content twice. All these differences have to be taken into account while setting and grading educational activities.

Other challenges

Search engines bring in other changes too. Knowledge is easily available. We don’t have to know facts any more. Instead, we try to remember where and how to find it¹. Programmers cut and paste snippets of code. Engineers run simulators. There are forums for both homework questions and teacher complaints. Skill is becoming more important than knowledge and memory. Even our views on ethics and morality change – how do we explain plagiarism to the copy-and-paste generation?

¹ Russell, D., What Do You Need to Know to Use a Search Engine? Why We Still Need to Teach Research Skills, AI Magazine, 36(4), 2015.

² Marion Walton, M., Archer, A., The Web and information literacy: scaffolding the use of web sources in a project-based curriculum, British Journal of Educational Technology, Vol 35 No 2, 2004.

³ Land, S., Hannafin, M. J., & Oliver, K. Student-Centered Learning Environments: Foundations, Assumptions and Design. In Jonassen, D. H. & Land, S. (Ed.), Theoretical foundations of learning environments (pp. 3–26), Routledge, 2012.

⁴ Spencer, Stephan. Google Power Search: The Essential Guide to Finding Anything Online With Google, Koshkonong. Kindle Edition.