Judgement Cannot Be Delegated · 9 of 12

It is often said that teaching staff need training in artificial intelligence, and that this training consists in learning to use the tools. An afternoon course, four applications, a few examples of a well-written instruction. The idea is reasonable and very widespread.

The data in the ODITE Report 2026 point towards a different recommendation. When the responses of 138 students and 63 teachers about AI in schools are compared, the largest difference in the whole study does not appear in critical thinking or in learner autonomy. It appears in the need to talk about the risks in class and to be trained to do so: teachers score it 4.71 out of 5 and students 4.13 (1).

And there is a second part to the finding, which is the uncomfortable one. When the same report analyses the views of the twelve experts interviewed, the point where they diverge most from one another is how teachers should support the use of AI (1).

Teachers are clearly asking for precisely the thing there is least agreement about. That is where this article begins.

Those who design see what those who learn do not yet see

The ODITE Report’s triangulation study, by Dr Neus Lorenzo, does not stop at establishing that teachers and students hold different views. It does something more demanding: it predicts where they should differ and where they should not.

The prediction is that teachers will score higher on the three central dilemmas of practice, because they perceive them more sharply, and that on a topic with no dilemma there will be no appreciable difference. The three differences are confirmed, with medium effect sizes. And the control case, the perception that AI has improved academic performance, behaves as expected: 3.36 for students, 3.37 for teachers, a difference of one hundredth (1).

That last figure is what turns the finding into an argument. A framework that is also right where it announces there will be no difference has explanatory power; one that finds divergence in everything does not.

It is worth saying how far the evidence goes. The samples are small and non-probabilistic, the predictions were formulated on the basis of the framework but were not registered before the data were seen, and the author herself warns: correspondence does not amount to causality (1). What the data allow us to assert is that the asymmetry exists, is systematic and always runs in the same direction.

The explanation the report offers is sober and fits the experience of any teaching staff. Teachers perceive greater risk because they design the learning and gain access to a dimension of the problem that the learner has not yet needed to resolve (1). It is not resistance to change. It is the point of view of someone who answers for the consequences.

The same asymmetry appears outside our own context. José Antonio Bowen and C. Edward Watson report a US survey in which 47% of university students consider that AI has a positive impact on learning, compared with 22% of faculty, while 50% of faculty consider the impact negative and only 22% of students agree (2). The gap is not a Spanish one; it crosses borders.

A syllabus that is not about tools

If teacher training is not a course on applications, what is it a course on?

The systematic review by Iliana Guaranda-Arias and her team, published in 2026, analysed 52 studies on ethical AI literacy and teachers’ critical thinking. The answer it synthesises is a list of seven components: privacy, transparency, bias, source verification, academic integrity, authorship and well-grounded pedagogical decision-making (3).

In practical terms they translate this into something that looks very much like a syllabus for a teaching staff: case analysis, source verification, identification of bias, data protection, discussion of authorship and evaluation of AI-generated answers (3). Not a single tool on the list.

The same review documents three risks that go together: technological dependency, uncritical use and insufficient teacher training (3).

This study has to be read knowing what it is. Its four databases are Ibero-American, with the exception of ERIC, and it excludes books, chapters and theses. Documentary reviews, theoretical analyses and cross-sectional studies predominate. The authors themselves acknowledge that there is no record of the protocol registration or of the exact date of the last search. That is why they conclude that ethical literacy and teachers’ critical thinking are interrelated competences at the conceptual level, and not that one causes the other (3). It is a review that is honest with itself, and that is how it should be cited.

There remains the question that matters most to a teaching staff: what is delegated and what is not. Miquel Àngel Prats, senior lecturer in educational technology at Blanquerna-URL, formulates it in the ODITE Report with a precision that saves a great deal of argument. AI reduces the time taken by administrative tasks, generates materials and produces preliminary feedback. The question is not how much time it saves, but what is done with the time freed up (4).

If that time is reinvested in support, in meaningful interaction and in instructional design, efficiency works in favour of quality. If it is used to produce more of the same, it erodes the humanising dimension of teaching. Time freed up is not a saving. It is a decision.

From this comes the professional shift that this article wants to make clear: from the teacher who transmits to the teacher who exercises critical mediation. They do not compete with the machine in speed or in volume of information, because that competition is lost. They accompany interpretation, contextualisation and the formation of judgement (4). Sarah Elaine Eaton, professor at the University of Calgary, sums it up in one of her six tenets of post-plagiarism, which Bowen and Watson report: ‘Humans can relinquish control, but not responsibility’ (2).

Criteria are not handed out, they are built

Digital Promise says it plainly in its AI literacy framework: safe and effective integration of these tools depends on AI literate teachers, and it is not reasonable to assume that teachers will learn it on their own (5). It is an institutional framework that argues a case, not a study that measures, and that is worth bearing in mind. But it is the clearest formulation in the corpus.

What training meets that condition? The Guaranda-Arias review takes up the thesis of Kelley and Wenzel: situated professional experiences are needed, not just technical training (3). The words in context are the key. Training that happens inside the practice of a particular team, with its subjects, its students and its dilemmas.

The CRAFT initiative, from Stanford’s Graduate School of Education together with its institute for human-centred AI, works that way by design. Its materials are not produced at the university and distributed afterwards: they are co-designed with annual cohorts of teachers from different disciplines, and what happens in their classrooms modifies the content, the structure and the style of the material. Its training is framed as the development of a well-grounded position that guides responsible decisions (6). It is a clear model, with a specific scope: its resources and its cohorts are secondary-level and from the United States.

Closer to home there is an experience with data. The APLICA-IA project, of the Societat Catalana de Pedagogia, trained 170 direct participants between 2023 and 2025, from thirty-four schools, two universities and two teacher training centres, in Catalonia, the Balearic Islands and Andorra. Its approach is collaborative action research: teachers do not receive practices, they co-create them. The before-and-after measures improve in knowledge about AI, in self-efficacy, in attitudes and in critical digital competence (7).

Those data describe an improvement, although they do not demonstrate that it was caused by the training: the design compares the same group before and after, with no control group, and much of what is measured is each teacher’s perception of their own capability. With that caution, there is one result that deserves attention for how honest it is: after the training, teachers’ wariness towards AI remains (7). It was not about convincing anyone.

The underlying argument is supplied by Miquel F. Oliver, Begoña De la Iglesia, Lucía Buil and Carme Pinya, who document in the ODITE Report a joint training programme for the teaching staff of the faculties of Education and of Law in the Balearic Islands. They worked in interdisciplinary groups, piloted tools and collectively drew up a guiding code of practice and a rubric. What they learned along the way is what gives this section its title: the principles that guide the use of AI cannot be imposed from outside, they have to be built from practice, dialogue and professional consensus (8).

What to do with this

For the leadership team. Prats identifies four elements a school needs in order to sustain critical teaching practices, and they work well as a self-evaluation for the teaching staff: clarity of rules about permitted use in tasks and assessments, continuing training in the technical, ethical and methodological dimensions, data protection and equity as operating principles, and coherence with the school’s educational project (4). Without that architecture, he warns, practice remains fragmented and dependent on individual initiatives.

For the leadership team. The itinerary of the University of the Balearic Islands is transferable to a secondary school staff: mixed groups by department, real piloting of one tool on one specific task, deliberation on what has been seen and collective drafting of a guiding, non-prescriptive code of practice, with a rubric to accompany it (8). The document matters less than the process. The process is the training.

For the teaching staff. Declaring one’s own use. The ODITE Report includes chapters that set out at the end what AI did in their preparation and what it did not: support with the drafting and the clarity of the exposition, without in any case replacing professional judgement or generating the conclusions (1). Adopting that practice in the school’s schemes of work and materials teaches more about integrity than any rule, because it models the behaviour being asked of students.

To conclude

Prats’s conclusion is the one that best closes this journey: what is decisive is not the power of the algorithm, but the quality of the pedagogical judgement that guides its integration (4). A participant in the APLICA training put it another way, and it is worth recording: the real potential of AI is not technological, but pedagogical and social (7).

Far from diluting the figure of the teacher, the presence of these systems can reinforce it. In a world where producing information is trivial, professional value shifts towards mediation, interpretation and judgement. That is not learned in an afternoon, and it is not learned alone.

Food for thought. Of last year’s training hours, how many served to learn a tool and how many to build a judgement?

Glossary

• Ethical AI literacy. The ability to understand and critically evaluate the workings, the limits and the educational consequences of AI systems, with attention to privacy, transparency, equity, authorship, responsibility and academic integrity.
• Contextualised or situated training. Training that takes place within the real professional practice of a team, on its tasks and its dilemmas, rather than transmitting decontextualised content.
• Co-design. A method by which materials or criteria are drawn up together with the people who are going to use them, so that their experience modifies the result and does not merely validate it at the end.
• Critical mediation. The teaching role that consists in accompanying the interpretation, contextualisation and problematisation of generated information, rather than competing with the machine in producing it.
• Self-efficacy. The confidence a person has in their own capacity to carry out a task. It is measured by asking, not by observing, and so it is worth distinguishing it from actual performance.
• Pre-post design. A study that measures the same group before and after an intervention. It shows whether something changed, but without a comparison group it does not allow the change to be attributed to the intervention.

References

1. 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 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]. Observatorio de Innovación Educativa y Cultura Digital, Asociación Espiral. Sections 3, 4 and 5, and declaration of AI use.
2. 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.
3. Guaranda-Arias, I. C., García-Ortiz, L. R. and Guerrero-Zambrano, M. (2026). Alfabetización ética en inteligencia artificial y pensamiento crítico del profesorado: una revisión sistemática reportada conforme a PRISMA 2020 [Ethical literacy in artificial intelligence and teachers’ critical thinking: a systematic review reported in accordance with PRISMA 2020]. Innova Science Journal, 4(3), 964–974.
4. Prats, M. À. (2026). Enfoques metodológicos y prácticas docentes críticas en la era de la inteligencia artificial [Methodological approaches and critical teaching practices in the age of artificial intelligence]. In Informe ODITE 2026.
5. 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.
6. CRAFT: Classroom-Ready Resources About AI for Teaching. Stanford Graduate School of Education and Institute for Human-Centered AI. Core principles, professional development and co-design teacher programme. https://craft.stanford.edu/
7. 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.
8. Oliver Trobat, M. F., De la Iglesia Mayol, B., Buil Legaz, L. and Pinya Medina, C. (2026). Coconstruyendo criterios para el uso pedagógico y ético de la inteligencia artificial en la educación universitaria [Co-constructing criteria for the pedagogical and ethical use of artificial intelligence in university education]. In Informe ODITE 2026.