AI: more than one teacher in two knows it, but few use it in class. Between enthusiasm and caution, the challenge is clear: guarantee educational added value through true appropriation.
The school is not a simple place of techniques, but a space where transmission and the idea of justice take place. The introduction of artificial intelligence into classes falls under this responsibility: it requires us to specify what we expect from the instruments, and to better define what cannot and must not be delegated. A survey that we conducted among teachers, between January and October 2025, establishes, in partnership with the MAIF, an unambiguous inventory: the declared familiarity with generative AI is in the majority (53.4%), but the use is still mainly outside the classroom (61.3%) and integration in the teaching context strongly depends on the level and seniority. At the second level, 37% of familiar teachers declare use in class, compared to 18% at the first level; and among teachers with more than fifteen years of seniority, the “don’t know” mode becomes the majority (56%), a sign of informed caution and an expectation of governance. Michel Serres warned us that “we are changing the world”: but the school, to remain faithful to its mission, can only “change” by adding proof to promises, and frameworks to instruments.
These gaps do not reflect a moral hierarchy: they describe unequal regimes of exposure and conditions of appropriation. Teachers between 1 and 15 years old, more often familiar, experiment through self-training (54.7%) and through peer exchanges (22.9%), while more experienced colleagues set conditions: training, transparency, reliability. To take up Hannah Arendt, education is the meeting between the novelty of the world and the birth of beings. It is at this point of contact that adoption must be built, through differentiated mechanisms, to transform curiosity into appropriation without denial. In disciplines where written production, structured evaluation and technical autonomy of students are more present, integration into class is more easily achieved: logic confirmed by the 37% of uses in secondary school classes among familiars. Conversely, the centrality of first-degree human mediation and developmental constraints require careful engineering, where we adapt rather than transpose. Bergson would say that virtualities are not actualized at the same pace in all environments. It is up to the institution to organize these rhythms, so that they benefit everyone.
At the heart of the justifications lies a coherent architecture. Familiars value time saving and differentiation. The tool is envisaged as an instrument for freeing up useful time and precision of routes. Non-familiars and teachers with long experience formulate requirements for governance and reliability; this becomes a firm injunction for teachers who are sometimes exhausted by fluctuations and the effects of successive fashions. They are asking for transparency sheets, clauses on the non-commercial exploitation of school data, and periodic audits. This is not a firm refusal, but a conditioned caution.
When Karl Popper reminds us that all knowledge is fallible, he does not invite us to mistrust, but to write public protocols and put them to the test: this is how school transforms innovations into common goods. As such, prudence is not immobility, but a practical virtue: it is Aristotelian phronesis, or the ability to adjust the decision to reality.
The typology of four priority profiles derived from the data is particularly interesting for this purpose: educational explorers (documented classroom use), utilitarian adopters (outside the classroom, ready to switch with reproducible scenarios), experienced cautious (face-to-face training, demonstrators, contractual guarantees), conditional ambivalent (massification lever under conditions). It functions as a preliminary action map, not as a classification, but as an orientation. This would mean giving in to a mercantilist logic of customer analysis, which we refuse. It is by respecting the real dispositions of the actors that innovation ceases to be marginal.
The result is a roadmap that excludes improvisation and sets clear priorities. First, launch disciplinary pilots evaluated at the second level, where the educational instrumentation is the most accessible and where professional postures lend themselves more to experimentation; these pilots must produce replicable scenarios and public indicators measuring achievement, teaching workload and acceptability by students and families.
Then, build a modular training offer: introductory courses that combine technical mastery and ethical understanding, disciplinary workshops delivered with ready-to-use use cases, and advanced modules devoted to the design of prompts and the critical evaluation of productions. These pathways must rely on methods that have already proven successful in the survey, such as tutored self-study, just‑in‑time formats and peer learning. This is the condition for being immediately operational and useful in the field.
It is essential to prioritize, in particular, teachers with more than fifteen years of seniority: their high proportion of “don’t know” and their marked obstacles (lack of training, expectations of governance, question of reliability) make short face-to-face formats and contextualized demonstrators necessary to remove doubts through experience.
Finally, establish governance that is understandable to everyone and especially teachers! For what ? Because innovation must not reinforce inequalities. And this must be accompanied by concrete policy measures for material equity (equipment loans, offline access to resources, mediation for families and sustained attention to fragile contexts). How many teachers suffer in front of a mobile tablet that is not updated, needs to be revised or does not have suitable applications. It was said with humor that teachers are the only ones to steal personal equipment from home to take it to school; the same logic now applies to digital.
In addition, it is necessary to organize the circulation of collective intelligence: set milestones, publish protocols, document successes and failures, capitalize on the results so that teams can take ownership of them. Cross-analyses of the survey show that use in class is accompanied by a clearer perception of the educational benefits and less focus on the obstacles. This correlation does not establish causality, but suggests a lever: that of supervised introductory courses (observation, pairs, short courses) to lower the entry barrier for non-users. Francis Bacon is not being coquettish here: the school knows when things are valid because it has put them to the test and can report on them publicly.
There remains the ethical requirement, which is non-negotiable. The AI has no tact, and the master’s judgment cannot be delegated. Kant insisted that “pedagogical tact” is a practical faculty, which we must preserve and equip, but never replace. AI can reduce repetitive tasks, give time for differentiation and remediation, and improve the quality of individualized feedback. But this promise is conditional: it requires proof of effectiveness, contextualized training, robust governance and compensatory measures. In addition, families and students are not passive recipients: they must be involved in the evaluation, particularly with regard to equity and commitment. Without these precautions, innovation will accentuate the divides that the school is precisely working to reduce. Paul Ricoeur liked to say that the future is what we are going to do: AI is part of this active responsibility.
Transforming familiarity into appropriation is a collective task: it requires the courage of proof, the humility of corrections, the clarity of the rules and the desire to redistribute the benefits. The angles of action are known and supported by the data. They are instruments, but also commitments. At this price, AI could cease to be an object of polarization, and become what it should be: a support in the service of a demanding and fair school.
