Education and the transmission of knowledge

15.1The universal tutor: promise and reality

Under the hood

How do you build an AI tutor? You start from a large general-purpose model that you steer with a system prompt (chapter 4) imposing a Socratic stance: do not hand over the answer, but guide with questions, check understanding step by step, encourage. To this you often add the curriculum's knowledge base, supplied to the model at the moment of answering (retrieval-augmented generation, RAG, chapter 6) in order to limit hallucinations and stick to the curriculum; some systems continuously measure the student's answers to adjust the difficulty, heirs to the old "intelligent tutoring systems." Products such as Khanmigo (Khan Academy) or the "learning" modes of the major assistants illustrate this approach. But no instruction fully neutralizes the risk of error in the underlying model: human oversight remains necessary.

Under the hood

How a tutor adapts (spaced repetition and knowledge tracing)

Personalizing learning is not just a matter of a chatty language model; it rests on proven pedagogical techniques that AI automates and refines. Spaced repetition exploits a well-documented law of memory: you retain something lastingly by reviewing it at increasing intervals, just before you would forget it. Software thus calculates, for each student and each item, the best moment for the next review. Knowledge tracing goes further: from the student's answers, the system maintains a model of what they have mastered or not, item by item, and estimates the probability that they will get the next question right. It is this model that makes it possible to offer the right exercise at the right level, neither too easy nor discouraging (the educators' "zone of proximal development"). What generative AI adds is grafting onto this skeleton a natural-language explanation tailored to each sticking point. The promise of the personalized tutor (section 15.1) therefore rests as much on this machinery of modeling the student as on the conversation.

15.2The crisis of assessment

15.3Transforming school and university

In plain terms

Beyond the student, AI affects the entire education system.

The augmented teacher: preparing lessons, generating differentiated exercises, grading papers, handling administrative work. AI can give time back to the pedagogical relationship, provided it remains a tool in the service of the teacher.
What is taught: if AI can code, translate, and write, school must shift the emphasis toward what remains distinctly human and what AI does not master: critical thinking, creativity, collaboration, source-checking, and precisely the ability to work with AI (chapter 17). This AI literacy (understanding what a model is, its limits and its biases, knowing how to verify its outputs) is becoming a subject in its own right.
Lifelong learning: in an upended labor market (chapter 17), learning how to learn and retraining become central skills, ones that AI can in fact help to acquire.

In context

The French institutional response

France is taking a different path: regulating and equipping. Since the usage framework of June 2025, independent use of these tools by students has been authorized only from the 4e (the equivalent of eighth grade) onward, and unauthorized use of generative AI on an assignment constitutes cheating. Since September 8, 2026, a "sovereign" AI assistant has been available to all teachers through the apps.education.fr portal (help with lesson preparation, document analysis, differentiation), with no training on personal data. The "Pix IA" course is mandatory from the start of the 2026 school year for students in 4e, in the general 2de (tenth grade), and in the first year of the CAP vocational certificate. Finally, on August 25, 2026, the minister announced a dedicated AI course in 2de starting in 2027, whose curriculum is expected in December 2026.

15.4Risks: dependence, divide, and deskilling

Diagram 15.1. The two faces of educational AI. The same tool can reinforce learning or replace it. It all depends on use: thinking with AI, or letting it think in your place.

Deskilling: if you systematically outsource cognitive effort (writing, calculating, reasoning) to the machine, the underlying skills risk atrophying. Yet effort, difficulty, and even error are ingredients of deep learning. This risk affects adults too (chapter 19). It is now being measured (see the "In context" note below).

In context

Deskilling, now measured

Two working papers, not yet peer-reviewed, provide figures. The study by Strömberg, Lei, and Wu, presented by its authors in September 2026, follows about 27,000 Chinese students in grades 7 through 12 from 2023 to 2025: 80% started using generative AI and half shifted to fully outsourcing their homework; homework grades improve, but results on in-class exams fall by about 20% on average. Conversely, a more modest randomized experiment (Contractor and Reyes, July 2026) finds a positive effect, but only among students who use AI to have concepts explained to them. For its part, the report of the MIT committee on AI in education (September 2026) names the danger of "cognitive surrender" and of the "illusion of learning," notes that AI can respond credibly to almost any undergraduate written assignment, and recommends not banning it but redesigning assessment (oral exams, projects, proctored exams), in line with the approaches outlined in section 15.2. One piece of context: the PISA 2025 results, published by the OECD on September 8, 2026, show a marked decline in reading and mathematics across the OECD; neither the OECD nor the experts cited attribute this decline to generative AI.

The digital divide: unequal access to the best (paid) tools risks widening the gaps between students and between countries, instead of narrowing them. The promise of democratization can turn into a deepening of inequalities.
Minors' privacy: entrusting education to tools that collect data on children raises especially sensitive protection questions (chapter 21).

In context

Should one still pursue long studies?

AI has reignited a heated debate over the value of degrees. At one, provocative pole, the essay Stop Studying (Laurent Alexandre and Olivier Babeau, 2025, chapter 7) argues that, with intelligence becoming "free and infinitely available," long classical studies would become "counterproductive": one should accumulate less of the knowledge that AI provides on demand, and instead learn how to learn, continuously. At the other pole, many educators point out that the opposite arguments remain powerful: solid foundational knowledge is needed to use AI well and to judge its answers (without it, you cannot spot its errors); the degree retains a function of signaling and of social mobility; and school transmits far more than knowledge (critical thinking, method, socialization, autonomy). The risk, stressed on both sides, is that a poorly understood slogan of "stop studying" would widen inequalities, by steering the most vulnerable students away from study while the best-endowed keep training. The reasonable position is surely not to study less, but to study differently, with AI shifting the center of gravity from memorized knowledge toward skills of judgment, creation, and lifelong learning (chapter 17).

Key takeaways (chapter 15)

AI promises the personalized universal tutor, but a good human tutor motivates and builds a bond that the machine does not reproduce; AI helps above all the student who is already self-directed.
Traditional assessment (the homework assignment) is in crisis; detection is unreliable, and the real answer is to rethink assessment (oral, in-class, projects, AI used openly).
AI also transforms the teaching profession (augmentation) and the content taught, which must favor critical thinking and the ability to work with AI.
The major risks: deskilling (outsourcing cognitive effort), the digital divide, and minors' privacy; at the start of the 2026 school year, large American districts suspended generative AI for students, while France is equipping its teachers and making an awareness course (Pix IA) mandatory.

From learning to imagination: chapter 16 explores how AI is upending creation and culture.

15.1The universal tutor: promise and reality#

15.2The crisis of assessment#

15.3Transforming school and university#

15.4Risks: dependence, divide, and deskilling#

Key takeaways (chapter 15)

15.1The universal tutor: promise and reality

15.2The crisis of assessment

15.3Transforming school and university

15.4Risks: dependence, divide, and deskilling