Judgement Cannot Be Delegated · 12 of 12

It is often said that schools are split in two when it comes to artificial intelligence. On one side, those who ban it and police it; on the other, those who use it without restraint. And in between, according to the story, an unbridgeable generational gap: students who use it naturally and teachers who fear it. It is a tidy story, easy to tell and easy to headline. But it is also a story that disappears when somebody takes the trouble to measure it.

This is the last article in the series, and it is time to do what is asked of any learning process: to separate what we know from what we assume, and to decide what to do with it.

The map of the disagreement

In April 2026, the Observatory of Educational Innovation and Digital Culture published a report with content that is unusual in this kind of publication: a chapter devoted to validating its own conclusions statistically. It is by Neus Lorenzo Galés, vice-president of the Societat Catalana de Pedagogia. The procedure consisted in cross-checking three direct sources, a questionnaire to students (138 responses), another to teachers (63) and interviews with twelve experts, and comparing their answers with statistical tests, not with impressions (1).

Before examining the results it is worth establishing the size of what is being analysed. These are small, non-probabilistic samples with voluntary participation, and the authors themselves say so plainly: they do not claim to guarantee a representation that is scalable or transferable to other contexts (1). What they offer is not an X-ray of the Spanish-speaking education system. It is an honest snapshot of those who chose to answer. With that caution in place, the results are striking.

The first: on the question of whether AI improves academic results, students and teachers score almost identically (means of 3.36 and 3.37 out of 5), and the difference is so small that the statistical test rules it out altogether (1). It is the point of greatest consensus in the whole study. The overall perception also converges: around 70% of students and 79.4% of teachers rate the use of AI positively, and there was not a single teacher who rated it wholly negatively, while 6.5% of students did express explicit rejection (1). Read that slowly, because it inverts the cliché. Those who showed hostility towards AI were the students, not the teaching staff.

The second result: where there is real divergence, with statistical significance and a medium effect size, is in concern about critical thinking (means of 3.74 among students and 4.35 among teachers) and in the need to debate the risks in class (4.13 and 4.71). In both cases teachers score higher, and in the second the distance is the greatest in the whole study (1). The disagreement, therefore, is not about whether the tool is useful. It is about how much to worry and how much to talk about it.

The third is the most uncomfortable. The greatest polarisation in the whole report is not between sectors, but within one of them: the twelve experts interviewed disagree among themselves, more than on any other topic, about the credibility of the institutions responsible for governing AI in education (1). And at that point the report records a silence, because the students’ voice on governance does not yet exist (1). A complementary analysis also finds that these axes of divergence coincide with the three dilemmas that emerge from classroom experiences: automation versus cognitive effort, personalisation versus dependency, and efficiency versus pedagogical depth (11).

Listening to students qualifies the picture further. Gina María Ramírez Zöller, a systems engineer and head of the Unidad Educativa Lemas in Guayaquil, surveyed 327 students aged between 12 and 16 at her school, chosen by accessibility and consulted at a single point in the year. What worries them most about AI is not boredom or surveillance: it is dependency (42.70%), followed by the risk of receiving incorrect information (26.40%). Sixty-seven per cent consider that using it for homework is legitimate help if it is used well. And seven out of ten, 70.3%, believe they should be given guidance on responsible use (2). It sounds like a minor finding, doesn’t it? Well, it is the most demanding one in the report: students are asking in writing for what many teaching staffs have not yet decided to give.

Francesca Burriel Manzanares, doctor in Pedagogy and an education inspector, carried out a similar consultation at the Colegio Español María Moliner in Andorra before introducing AI in the school, with two very small groups, 19 students in the first year of ESO (Year 8) and 21 in the first year of Bachillerato (Year 12). The younger ones use AI little and almost never check what it answers, but they believe that nothing would happen to them without it. The older ones use it a lot, distrust it more and cross-check more, and at the same time acknowledge that working without it would be hard for them (3). Forty responses generalise nothing, and that is how it should be read. As an indication from one school, it points to something worth keeping in mind: the habit of checking does not arrive early, it arrives late and it arrives accompanied by dependency.

The map, then, looks like this. We agree that AI contributes something. We disagree about how much to worry. And we distrust, above all, whoever decides. That third axis deserves a conversation of its own, which is the one the previous article in this series takes up.

A future without a crystal ball

If the disagreement is not about whether AI is useful, the question shifts: for what world are we preparing these students? This is where caution becomes obligatory, because most of what is published about the future of work consists of projections, and many of them come from consultancies and investment banks, not from peer-reviewed research.

José Antonio Bowen and C. Edward Watson deal with this in the chapter of their book on the new era of work, and they are transparent about where their figures come from. The percentages in circulation about each sector’s exposure to AI (6% of tasks in construction, 44% in the legal field, 46% in administrative support) come from a Goldman Sachs report, as does the forecast of a 7% increase in global GDP (4, 13). They are estimates by an investment bank, and that is how they should always be cited.

What the chapter does support with studies behind it is more interesting than any percentage. Bowen and Watson sum it up in one sentence: AI will eliminate some jobs, but it is going to change them all (4). And in the studies they review a constant pattern appears. Both in a study of five thousand customer support agents and in an MIT experiment with 453 college-educated professionals, the one who benefits most from AI assistance is always the novice or less qualified worker. In the MIT case, those who used ChatGPT completed the writing tasks 40% faster and with 18% higher quality, and the weakest writers were helped the most, so that inequality between workers decreased (4). Bowen’s explanation is sober and useful for the classroom: AI does C-grade work, but it is consistent C work, and for someone starting out, that average is already an improvement.

There lies the problem, and it is a pedagogical problem before it is an economic one. If AI takes on the routine tasks that until now were done by entry-level profiles, interns, junior staff and our beginners lose precisely the opportunities in which the craft is learned (4). Alejandro Espeso-García, of the Catholic University of Murcia, develops this in a conceptual review published in 2025 with two concepts worth adding to the vocabulary of leadership teams. The first is deskilling, the progressive atrophy of professional skills resulting from automation, which hits intermediate profiles hardest. The second is the paradox of automation: the more competent the system becomes, the fewer opportunities the person has to practise, and when the system fails or hallucinates, there is no longer the experience to correct it (5). It should be stressed that this is a review article, with no data of its own, synthesising recent literature.

Against that scenario, Espeso-García proposes hybrid intelligence, understood not as substitution but as a complementary system, in which the person guides and refines what the algorithm produces, with evidence that this collaboration generates more applicable solutions than either the algorithm or the person separately (5). As for the competences that gain in value, the article draws on the World Economic Forum’s 2025 report: analytical thinking and innovation in first place, and alongside them leadership, resilience, flexibility and, very prominently, learning to learn (5, 12).

It is not a surprising list for anyone who has spent years in a school. It is, almost exactly, what schools have spent decades saying they want to develop. And that is the most useful result of looking at the future of work without a crystal ball: the pedagogical answer does not depend on getting the prediction right.

Who is holding the wheel

Espeso-García closes his article with an image that also serves to close this series. Human cognition, he argues, is already hybrid and will irreversibly be so, so the question is not whether the mind will merge with these systems, but whether that hybridisation will be evolution or involution. And everything depends on one condition: that the mind retains its sovereignty as a cognitive agent, that is, that it defines the ends, judges the means and exercises final judgement. The outcome, he writes, will be decided by whether we act as pilots guided by technology or as passengers confusing the illusion of understanding with the act of thinking (5).

The distinction is not rhetorical. At another point in the same article he translates it into a change of model: moving from human in the loop, where the person merely supervises what the system proposes, to human in charge, which restores critical control and final authority to them. And he adds the condition that underpins everything above: without a solid foundation of knowledge that allows the algorithm to be audited and corrected, the student is not an assisted pilot, but a blind passenger (5). It is the same thesis that has run through this series since the first instalment. You cannot supervise what you do not understand.

The AILit Framework, published in June 2026 by the OECD and the European Union, organises AI literacy into four domains: engaging, creating, managing and shaping. The last two are the ones that concern a leadership team that has to decide. Managing AI is defined as deliberately dividing the work between people and AI, and Shaping AI leads students to regard these systems as something deliberate and human-directed, rather than something inevitable (6). Cathy Adams, of the University of Alberta, puts it in the framework itself with a precision that any teaching staff without a technical profile will appreciate: students do not need to be AI engineers to understand that they can shape technology, not just be shaped by it (6).

With those two sources a simple decision framework can be built, which I offer as my own synthesis and not as a recommendation from either of them. Three verbs: protect, delegate, shape.

Protecting is deciding which effort is not outsourced, because it is precisely the effort that teaches. It is the cognitive debt we discussed in the fourth article and the desirable difficulty of the fifth. Delegating is the managing domain of the AILit Framework: dividing the work deliberately, which requires knowing beforehand what is being divided. Shaping does not mean programming models. It means that a school understands that these systems respond to human decisions about data, criteria and purposes, and that it can hold those decisions to account.

And the word that best sums up this closing is not a technological one. In 2025, UNESCO devoted International Day of Education to artificial intelligence technologies under the motto ‘Preserving human agency in a world of automation’, which Rosa Maria de la Fuente Castelló and Neus Lorenzo take up as the axis of their teacher training work (10). Human agency is the capacity to decide and to answer for what has been decided. It is not preserved with a circular. It is preserved with agreed, reviewable criteria.

From analysis to decision

For the leadership team: the school strategy on one page. Three decisions a year, one per verb. What we protect (which learning tasks or processes are done without AI and why), what we delegate (which repetitive work, teaching or administrative, passes to AI with defined oversight) and what we shape (which criteria we require of the tools that come into the school). Each decision with an indicator and a review date. It is a proposal from this blog, not a validated instrument, and its value lies in the fact that it fits on one page and forces a choice.

For the leadership team: the staff debate with its own data. The ODITE Report concludes that the mismatch between the students’ perspective, more attentive to immediate benefit, and the teachers’, more attentive to structural conditions, makes it necessary to create spaces for deliberate pedagogical dialogue between them (1). The proposal is simple: replicate those four questions in the school, compare the staff’s means with the students’ and argue about what comes out rather than about headlines. It fits in a ten-minute form.

For the classroom: three documented experiences, with their real scope.

Javi Badia, teaching technician at the Institut de Ciències de l’Educació of the University of Lleida, worked with students in the fourth year of ESO (Year 10) at the Institut Torre Vicens on programming in Python a chatbot about the biochemist Joan Oró. The system answers by keyword matching, and the aim was not to learn to program but to open the black box. By deliberately entering incorrect information, the students found that the system does not understand, it compares. The assessment was done with a rubric, observation and a learning journal, without before-and-after measures, so we are talking about formative evidence and not an experiment (7).

Maite Carbonell Cajal, a Drawing teacher at the INS Leonardo da Vinci in Sant Cugat del Vallès, took 120 students in the second year of ESO (Year 9) to draw with Quick, Draw!, Google’s tool that recognises drawings on the basis of millions of strokes made by people. The activity contrasts two things: how we humans represent the world and how a system trained on those same representations classifies it (8).

Marc Alier, a lecturer at the Universitat Politècnica de Catalunya, tried out with 47 Computer Engineering students an open platform that allows the teacher to create assistants restricted to selected sources and integrated into Moodle, with no external registration of students (9). It is a university experience, and what it brings to schools is not the activity but the criterion. An assistant defined by the teacher and anchored to verifiable sources is not the same thing as a generic chatbot.

What to read after this series

A series is of little use if it does not leave a path ahead. This is the complete corpus it has drawn on, ordered by reader profile.

• To understand the machine: Bowen and Watson, Teaching with AI (2024), chapters 1 to 3; Wulff and Mata in Trends in Cognitive Sciences (2026); and the open applications of the Finnish project Generation AI, winner of the 2026 European Digital Skills Award, which allow a classifier or a language model to be manipulated in the browser without programming.
• To understand what it does to the learning mind: Bjork and Bjork on desirable difficulties (2020); Kosmyna and her team at the MIT Media Lab (2025), remembering that it is a preprint; Lee and colleagues at the CHI conference (2025); Gerlich in Societies (2025), of correlational design; Doshi and Hauser in Science Advances (2024); and the review by Espeso-García (2025).
• To design AI literacy: the AILit Framework of the OECD and the European Union (2026); the Digital Promise framework by Mills and colleagues (2024); the international Delphi study by Hallström and his team (2026); and, for the critical perspective, the ‘digital plastic’ of Roe, Furze and Perkins (2026).
• For classroom practice and teacher training: the full ODITE Report 2026; the PRISMA review by Guaranda-Arias and colleagues (2026) on ethical literacy and teachers’ critical thinking; and the open materials of CRAFT, the Stanford Graduate School of Education initiative that produces AI literacy resources for secondary education across all subjects.

To conclude

Twelve instalments on, the balance is more modest and more demanding than the headlines promised. We know that these systems predict and do not understand, that hallucination is how they work and not how they break down, that bias is a property of the corpus, that cognitive saving is paid for and that the international consensus on what to teach is small. We do not know what work will be like fifteen years from now, and whoever claims otherwise is selling something.

What we do know is that the future of schools with AI is not predicted. It is decided, in specific meetings, with written and reviewable criteria.

And the question this series closes with is not about technology: which effort is your school going to protect this year, and how will you know in a year’s time whether it has really protected it?

Glossary

• Deskilling. The progressive loss of professional skills through sustained delegation of tasks to automatic systems. It affects above all those who are starting out, because the tasks that are automated are the ones that used to serve to learn the craft.
• Hybrid intelligence. A model of collaboration in which the person orients, corrects and validates what the system produces, instead of replacing their own work with it.
• Human agency. The capacity of a person or an institution to decide with judgement of their own and to answer for the consequences of that decision.
• Cognitive sovereignty. A situation in which the mind retains the definition of the ends, the judgement about the means and the final decision, even when it delegates the execution.
• Shaping AI. The fourth domain of the AILit Framework. Understanding that the behaviour of these systems responds to human decisions about data and criteria, and being able to propose improvements. It does not consist in programming models.

References

1. Lorenzo Galés, N. (2026). La validación científica del informe ODITE 2026: un ejemplo de uso de IA en investigación educativa [The scientific validation of the ODITE Report 2026: an example of the use of AI in educational research]. In J. M. Muñoz, N. Lorenzo, X. Suñé, M. À. Prats and C. López (Coords.), Informe ODITE 2026. Claves para una nueva educación [ODITE Report 2026. Keys for a new education]. ODITE, Asociación Espiral.
2. Ramírez Zöller, G. M. (2026). La IA bajo la lupa de los estudiantes [AI under the students’ microscope]. In Informe ODITE 2026, Praxis section.
3. Burriel Manzanares, F. (2026). La voz de los protagonistas: expectativas del alumnado sobre la IA [The voice of those involved: students’ expectations about AI]. In Informe ODITE 2026, Praxis section.
4. Bowen, J. A., and Watson, C. E. (2024). Teaching with AI: A Practical Guide to a New Era of Human Learning. Johns Hopkins University Press.
5. Espeso-García, A. (2025). Generative artificial intelligence: Between enhancement and cognitive offloading. Cultura, Ciencia y Deporte, 20(66). https://doi.org/10.12800/ccd.v20i66.2698
6. 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. https://doi.org/10.1787/65cd27d4-en
7. Badia, J. (2026). Programar para desmitificar la IA: el chatbot de Joan Oró en la educación secundaria [Programming to demystify AI: the Joan Oró chatbot in secondary education]. In Informe ODITE 2026.
8. Carbonell Cajal, M. (2026). La IA en el arte: conciencia digital y lenguaje visual [AI in art: digital awareness and visual language]. In Informe ODITE 2026.
9. Alier, M. (2026). Una experiencia de integración segura de inteligencia artificial generativa en educación superior mediante asistentes de aprendizaje [An experience of safe integration of generative artificial intelligence in higher education through learning assistants]. In Informe ODITE 2026.
10. De la Fuente Castelló, R. M., and Lorenzo Galés, N. (2026). Agencia humana en la IA educativa: APLICA-Soluciones para pasar a la acción [Human agency in educational AI: APLICA-Solutions for moving into action]. In Informe ODITE 2026.
11. Lorenzo Galés, N. (2026). Justificación estadística: correspondencia entre los dilemas pedagógicos de la praxis y las divergencias entre expertos [Statistical justification: correspondence between the pedagogical dilemmas of practice and the divergences among experts]. In Informe ODITE 2026.
12. World Economic Forum (2025). The Future of Jobs Report 2025. Cited in Espeso-García (2025).
13. Briggs, J., and Kodnani, D. (2023). The Potentially Large Effects of Artificial Intelligence on Economic Growth. Goldman Sachs. Cited in Bowen and Watson (2024).