Judgement Cannot Be Delegated · 7 of 12

It is often said that schools always lag behind technology, and this time we have the figure that proves it. According to a survey carried out in 2025 among European teenagers, 88% of thirteen- to fifteen-year-olds and 96% of sixteen- to eighteen-year-olds use artificial intelligence tools for learning and creative tasks at least a couple of times a week. It should be said that the survey was commissioned by Google. The Eurostat figure points in the same direction, with young people aged sixteen to twenty-four using generative AI almost twice as much as the general population. (1)

The official map of what they ought to be taught arrived on 18 June 2026. It is called the AILit Framework, it is signed by the European Commission and the OECD, and it answers in detail the question leadership teams have been asking for three school years: what students need to learn about artificial intelligence, and in what order. It arrives late, as almost all maps do. And it arrives with a surprise inside: the scientific consensus on what needs to be taught is considerably smaller than the map.

The map was drawn twice

The AILit Framework is a joint initiative of the OECD, produced by the CodeAI team and backed by a group of twelve international experts, among them Pati Ruiz, of Digital Promise, and Victor R. Lee, of Stanford University. The PISA Governing Board approved it on 8 April 2026. (1)

It organises AI literacy into four domains, presented in sequence. Engaging with AI is the foundational domain: recognising where AI is, evaluating its outputs and deciding on its role in everyday life. On that basis two domains develop in parallel, Creating with AI, which is using it as a creative collaborator while still making one’s own decisions, and Managing AI, which is deliberately dividing the work between the person and the machine. And it culminates in Shaping AI, improving systems so that they reflect human values. The framework warns that this last domain demands deeper technical knowledge, that it may require the collaboration of computer science teachers or community partners, and that it is reached only once the previous three have been mastered. (1)

The four domains are set out in nineteen competences, each unfolded across three proficiency levels, basic, intermediate and advanced, which are deliberately not tied to school years or ages. Each level comes with an ‘In the classroom’ scenario ready to be adapted. (1)

And here is the detail that says most about this document. The May 2025 draft had twenty-two competences, and its fourth domain was not called shaping but designing AI. Between May and October of that year more than two thousand people from more than a hundred countries gave their views on it, most of them teachers, who accounted for 41% of the survey responses. What they said is recorded in the annex to the final framework: the draft offered too positive a view of AI, and they asked for more scepticism, more tools for questioning the design of the systems, more metacognition and more attention to the environmental cost. The drafting team took it on board. (1)

A map that is corrected when read by those who are going to walk it deserves rather more confidence than a map that is not corrected. And its most useful sentence for a teaching staff fits on one line: ‘AI literacy is not the same as AI tool use’. (1)

What survives three rounds

While the map was being drawn, a team at Linköping University led by Jonas Hallström, professor of technology education, was asking the same question by another route. They convened an international panel of experts and asked them to say what primary and secondary students need to learn about artificial intelligence. It is the Delphi method: several anonymous rounds in which each participant sees the agreement reached by the group and gives their view again. Thirty-three experts responded in the first round, twenty-seven in the second and twenty-five in the third. (2)

Twenty-three statements came out of the first round. Eight remained. To survive, each one had to meet three conditions: that at least 60% of the panel rated it highly, that the spread of responses was small, and that the mean did not move more than two tenths between rounds. The eight survivors are grouped into two themes, and only two: foundational AI knowledge and critical perspectives. (2)

What is interesting is not the eight that remained. It is the fifteen that fell away, because they are precisely the ones that tend to head up training plans.
The statement that learning about AI is primarily about using text-based generative applications such as ChatGPT obtained a mean of 2.04 out of 5, the lowest score of the twenty-three. Prompt engineering came out at 3.22, and with the highest dispersion in the questionnaire, which means there was not even agreement about the disagreement. Programming for designing AI solutions came out at 3.08, and the most frequent answer was ‘disagree’. Computer science as the most important subject for understanding AI, at 3.04. (2)

It is worth pausing on one detail, because it teaches you how to read studies. The statement that AI ethics can be taught from an early age obtained high support, and in the final round it rose further still, from 4.16 to 4.42. Even so, it fell outside the final consensus. Not for lack of agreement, but because it rose too much: it failed the stability criterion. What did make it in, and what can be quoted without qualification, is that basic AI knowledge can be taught from the early years and primary school. (2)

And so we come to the clash. The article itself compares its results with the framework of the European Commission and the OECD, and observes that the managing and designing domains include understanding AI design principles and being able to shape AI, and that this is precisely what its panel did not reach consensus on. The comparison was made with the 2025 draft, however, not with the final version. But the substance holds: what the official map places at the summit is what expert consensus does not endorse as necessary for everyone. (2)

Despite the panel’s limits, the result is both a relief and a warning. A relief because the list is short and fits in a real school. A warning because the order is the opposite of the one we usually follow: first know, then criticise. Hallström and his team say it plainly: to be meaningfully critical towards AI, students need knowledge about AI (2). Just as we argued in the previous article in this series, and now it has an international panel behind it.

In the Spanish-speaking world a similar place is reached by another route. The analysis of expert interviews in the ODITE Report 2026 identifies digital and ethical literacy in the use of AI as one of the six categories running through the whole expert discourse, and proposes moving towards models in which students learn not only to use the systems, but to understand their limits, their biases and their effects on knowledge. (5)

From competence to behaviour

A competence framework is only useful if you can establish whether it is happening in the classroom or not. That is the work Digital Promise did in 2024, in a framework led by Kelly Mills. It defines six AI literacy practices and accompanies them with indicators written as observable student behaviours: algorithmic thinking, abstraction and decomposition; data analysis and inference; data privacy and security; communication and expression; ethics and impact; and mis/disinformation. Underneath all of them, two guiding values: centering human judgment and centering justice. (3)

The two frameworks do not ignore each other. The annex to the AILit Framework acknowledges that Digital Promise’s modes of engagement served as a basis for defining its own competences, and Pati Ruiz, co-author of the American framework, is a member of its expert group. (1) It is the same work at two scales.
Digital Promise adds a recommendation that sounds odd the first time you read it and very sensible the second: in the early grades, these practices should be worked on without AI tools. Analysing a dataset, discussing what a sample captures and what it leaves out, describing a procedure step by step. All of that is learned earlier and better without a screen, and it is what sustains the critique afterwards. And the same framework recalls why that critique is needed: systems that work as a black box ‘should not be used for high-stakes applications’. (3)

There remains the part that turns all of this into a school decision and not an interesting read. In 2029, PISA will incorporate an innovative domain called media and artificial intelligence literacy, MAIL for short, which will be assessed by combining traditional questions with simulated environments and whose results are expected in December 2031. (4) It is worth being precise, because the confusion is already circulating: PISA is not going to measure the AILit Framework, but MAIL, a broader domain, with its own assessment framework and still in first draft, to which the AILit Framework contributes. (1) Anyone who prepares their students with the AI map alone will leave out the media half of the exam.

And one figure to bear in mind before planning anything: according to the OECD’s TALIS 2024 study, only one in three teachers uses AI and three in four say they lack the knowledge needed to teach with it. (1) The article ‘What No One Can Learn on Their Own’ will deal with that, and it is the piece that decides whether everything else happens.

The three frameworks, in one table

Practical application

For the school: a self-diagnosis by domain. Before buying in training it is worth knowing where the school stands. The four domains of the AILit Framework work as four columns, and each department in the school answers three questions per column: what is already being done, in which subject and year, and on what evidence the claim that it is being done rests. What usually emerges is that a fair amount of the first domain is covered without ever having been called that, and that what there is rests in the hands of two or three teachers. The instrument is our own proposal based on the framework, not a tool taken from it.

For the leadership team: a decision about equity, not about organisation. Digital Promise recommends integrating AI literacy into core subjects, and the reason it gives is not pedagogical but a matter of justice: computer science and library media are usually elective courses, not taken by all students and, according to the data it cites, they disproportionately enrol white male students. Leaving AI literacy there turns it into a privilege. (3) The Delphi study points to the symmetrical risk: when responsibility belongs to all teachers and to no subject, it ends up belonging to nobody. (2) This decision cannot be taken by a single department.

In the classroom: the ‘for you’ scenario. It comes from the AILit Framework’s own scenarios, with the progression already worked out. At the basic level, students classify everyday objects and applications according to whether or not they use AI. At the intermediate level, they compare the ‘for you’ page of a social network with a classmate’s and explain why what appears on it influences them. At the advanced level, they analyse real cases of recommender systems that have shaped public perception of an issue, and discuss what effect that can have on voting behaviour. (1)

To conclude

We finally have an official map, and we have a scientific consensus that is smaller than the map. There is no need to choose between the two. The map tells us where we are walking; the consensus, what to do this school year and what can wait. And what the consensus says cannot wait is the most modest thing of all: that students understand how this works and that they learn to look at it with judgement. The rest is itinerary.

Food for thought. If someone at your school had to demonstrate tomorrow that students in the fourth year of ESO (Year 10) are literate in artificial intelligence, what would they show as evidence, and who would have gathered it?

Glossary

• AI literacy. The set of knowledge, skills and attitudes that make it possible to understand how AI systems work, evaluate their outputs critically and use them ethically. It is not the same as knowing how to use tools.
• AILit Framework. The AI competence framework for primary and secondary education published in June 2026 by the OECD and the European Commission. It is not binding.
• Domain. In the AILit Framework, each of the four large blocks into which the competences are grouped: engaging, creating, managing and shaping.
• Learning expectation. A description of what a student is able to do at a given proficiency level. In the AILit Framework there are three per competence, and they are not tied to a school year.
• Observable indicator. A specific behaviour that a teacher can see and record in class, and that serves as evidence that a competence is developing. It is what distinguishes a usable framework from a statement of intent.
• Delphi method. A research procedure in several anonymous rounds in which a panel of experts rates a series of statements, sees the group’s result and rates again, until opinions stabilise. It measures consensus, not effectiveness.
• Innovative domain (PISA). An assessment area that PISA adds to reading, mathematics and science in each cycle to measure cross-cutting competences. In 2029 it will be media and artificial intelligence literacy, known as MAIL.

References

1. OECD and European Union (2026). Cómo preparar a los alumnos para la era de la IA. Marco de competencias sobre IA para la educación primaria y secundaria (Marco AILit) [Empowering learners for the age of AI. AI literacy framework for primary and secondary education (AILit Framework)]. OECD Publishing, Paris. https://doi.org/10.1787/65cd27d4-en
2. Hallström, J., Mannila, L., Nordlöf, C., Heintz, F., Sperling, K., and Stenliden, L. (2026). AI literacy for K-12 education: an international Delphi study. Interactive Learning Environments. https://doi.org/10.1080/10494820.2026.2649553
3. Mills, K., Ruiz, P., Lee, K., Coenraad, M., Fusco, J., Roschelle, J., and Weisgrau, J. (2024). AI Literacy: A Framework to Understand, Evaluate, and Use Emerging Technology. Digital Promise. https://doi.org/10.51388/20.500.12265/218
4. OECD (2026). PISA 2029 Media and Artificial Intelligence Literacy. Project page. https://www.oecd.org/en/about/projects/pisa-2029-media-and-artificial-intelligence-literacy.html
5. Informe ODITE 2026. Claves para una nueva educación. Tendencias, retos y propuestas en la era de la IA [ODITE Report 2026. Keys for a new education. Trends, challenges and proposals in the age of AI], analysis of expert interviews, emerging categories and proposals for debate. Observatorio de Innovación Educativa y Cultura Digital, Asociación Espiral, Educación y Tecnología.