When Artificial Intelligence Enters the French Classroom

A high school student opens his phone at midnight, on the eve of his philosophy exam. He doesn't call a classmate — he turns to a language model. A commonplace scene in 2026, yet one that concentrates all the tensions of an unprecedented pedagogical transformation: artificial intelligence has established itself in baccalauréat preparation and in French as a Foreign Language learning, not by ministerial decree, but through the force of silent adoption, student by student, word by word.

The Assisted Essay: Between Scaffold and Crutch

The French dissertation — that exercise in structured thinking that the French national education system defends as training in critical thinking — has met a formidable interlocutor. Generative AI tools now offer detailed outlines, argumentative transitions, even ready-made introductions. For a student who struggles to organize their thoughts, this can be a precious cognitive scaffold: they observe how one idea calls forth another, how a thesis is built in layers.

But a scaffold, if it remains in place, prevents the facade from setting. The risk is not that the student will cheat — it's that they will stop experiencing the resistance of the problem, that intellectual friction which forges precisely the reasoning the exam claims to assess. Romain Thibault, a literature teacher in the Lyon region, puts it in a terse phrase: "My students now know what a good outline looks like. They no longer know why it's good."

This distinction is crucial. Understanding the form of an argument without grasping its logical necessity is like learning to recognize a guitar chord without ever hearing the music. The most sophisticated AI tools — those that, instead of producing the outline, ask Socratic questions to help the student build it themselves — seem better equipped to preserve this formative dimension. The question remains whether students, under time pressure, will choose the longer path.

The Baccalauréat Oral and Conversational Simulation

The oral examination — the Grand Oral for the general baccalauréat, specialty orals, BTS interviews — may be the arena where AI offers its most indisputable benefits. Simulating an examination panel is something a pedagogical chatbot does with an availability and patience that no overworked teacher can reasonably match.

Specialized platforms allow students to present their exposition to an artificial interlocutor that rephrases, questions, and objects. The algorithm detects verbal tics, overly long silences, and evasive formulations. The feedback is immediate, granular, and reproducible — one can redo the simulation ten times without exhausting the patience of one's interlocutor.

What AI cannot replicate is human unpredictability: the examiner who frowns, the assessor who smiles at an unexpected moment, the physical tension of an examination room. AI training develops a real but partial competence — rather like training in boxing against a sophisticated punching bag: useful, necessary, insufficient.

French as a Foreign Language: The Learner Facing an Inexhaustible Interlocutor

For learners of French as a Foreign Language, the promise is of a different nature. The historic problem with FLE is not a lack of methods — it's a lack of authentic exposure time. A Korean student preparing for the DELF B2 in Seoul cannot immerse themselves in the language. They can, however, converse with a language model for several hours a day, with no time constraints, no social awkwardness, no fear of ridicule.

This accessibility fundamentally changes the equation. Recent-generation automatic correctors no longer merely flag errors: they explain why "je suis allé" is correct where "j'ai allé" is not, they offer stylistically more nuanced reformulations, they adapt register to the communicative situation. This is a form of individualized tutoring that public education systems — even the most generous — struggle to provide at scale.

Pedagogical Chatbots: A Didactics Yet to Be Built

Chatbots dedicated to FLE learning are multiplying, but their pedagogical quality remains uneven. Some reproduce a behaviorist correction model — error detected, rule stated, exercise proposed — that has barely evolved since the language laboratories of the 1970s. Others, more ambitious, draw on communicative approaches: they place the learner in situations, invite them to negotiate meaning, to rephrase, to infer.

Research in second language acquisition — notably Stephen Krashen's work on the comprehensible input hypothesis — suggests that this second approach is more fruitful. Understanding messages slightly above one's level, without seeking to consciously analyze everything: this is what the best tools attempt to reproduce. The question is therefore not "can AI help in FLE?" — it clearly can — but "what pedagogical designs make the best use of it?"

Pronunciation: The Poor Relation of the Digital Revolution

One area remains underdeveloped: phonetics. Speech recognition has improved considerably, but prosodic feedback systems — those that work on rhythm, intonation, and phrasing — remain rudimentary for French. An English-speaking learner struggling with the "u" sound or nasal vowels still receives insufficiently precise feedback. This is an open challenge, where future progress could radically transform the learning experience.

The Critical Issues We Prefer to Sidestep

Every pedagogical revolution carries its unexamined assumptions. Three deserve to be named plainly.

Dependency and the Atrophy of Working Memory

Studies in cognitive science — notably those from Stanislas Dehaene's laboratory at the Collège de France — remind us that the effort of memory retrieval is itself formative. Searching for a word, formulating a rule from memory, solving a problem without external help: these seemingly painful efforts consolidate learning. When an AI tool systematically eliminates difficulty, it may simultaneously eliminate the learning that difficulty induced. Convenience carries a cognitive cost that technological enthusiasm tends to minimize.

Assessment in a World Where the Draft Is Invisible

The baccalauréat assesses — in theory — a thought process as much as a result. Yet AI renders the draft invisible: the examiner sees a paper, not the tentative steps that preceded it. This opacity undermines the foundations of certifying assessment. Institutional responses — AI-generated text detectors, a return to handwritten exams, revision of examination formats — are still fumbling. None is fully satisfactory. This is less a technical problem than a problem of assessment philosophy: what are we trying to measure, and why?

Equity, or the Digital Divide in Reverse

Access to the most powerful AI tools remains unevenly distributed. A subscription to a high-performance language model represents a non-negligible monthly cost for a struggling family. Students who have recent equipment, a stable connection, and developed digital literacy enjoy an additional preparatory advantage — contrary to the equalizing promise that technology readily brandishes. Pedagogical equity is not decreed: it is built through the infrastructure, training, and access choices made by schools and public policy.

What Teachers Are Doing With It

Faced with this reality, teachers are not passive. Many are experimenting with thoughtful uses: employing AI to generate argumentative counterexamples that students must refute, producing deliberately erroneous texts to be corrected, proposing debate simulations on cultural topics. In these configurations, the tool is a material, not an oracle.

Others resist, and their resistance is not reactionary: it carries a legitimate pedagogical conviction, namely that certain forms of learning — spelling, memorizing conjugations, the intuitive structuring of an argument — require extended time, repeated friction, and integration that cannot be delegated. Both postures coexist in staff rooms, often in the same teacher depending on the day and the level.

What is lacking is less a doctrine than an institutional space for comparing, evaluating, and sharing these experiments. In-service training in France on the pedagogical use of AI remains thin — teachers learn primarily from one another, in informal networks, through trial and error.

Between the Tool and the Vision

Artificial intelligence is neither the savior of education nor its gravedigger. It is a tool of unusual power, whose effect depends entirely on the pedagogical vision within which it operates. A hammer can build a house or break a window — what matters is the intention and competence of the person holding it.

For FLE as well as baccalauréat preparation, the potential gains are real: accessibility, personalization, availability, granular feedback. The risks are equally real: cognitive atrophy, amplified inequity, distortion of assessment. Navigating between the two requires not choosing a side — pro-AI or anti-AI — but honestly asking a prior question: what do we want students to be able to do alone, without any help, on the day they close the screen?

The answer to that question — provided one accepts to formulate it — is the only valid compass for deciding what AI may do in their place, and what it must not touch.