Education · scaffolding · autonomy · verification

AI & Learning

Using AI as a pedagogical aid without confusing assisted production with a capacity genuinely acquired.

Central thesis

A produced answer is not an acquired capacity.

An AI can provide a correct solution, explanation or reformulation without the learner later being able to reconstruct the reasoning, recognise the relevant situation or act without assistance.

What the tool produces and what the learner can do again alone are two different facts.

01 · Questions and examples

AI can vary examples and questions.

The educational report allows AI to propose questions, vary examples and generate graduated exercises. These uses can support learning when they serve a pedagogical capacity that has already been defined.

Multiplying examples is useful only if it helps locate a point of breakdown, progressively withdraw assistance or verify transfer.

02 · Error reformulation

Reformulating an error can help without explaining its cause.

AI can help reformulate an error or compare learning traces. It can make a structure more visible and prepare more precise feedback.

But it does not automatically know the classroom scene, the learner’s history, the didactic validity of its interpretation or the causes of the difficulty. Any causal explanation must remain hypothetical.

03 · Evidence and inference

Observed data, hypothesis and inference must remain distinct.

The report requires an explicit separation between what was actually observed, what is being supposed and what is inferred. A confident AI formulation does not turn a hypothesis into a fact.

A plausible explanation is not an observation.

04 · No diagnosis

AI must not diagnose the learner.

The corpus forbids automatically assigning an inner cause to the learner and recalls that Noosophy replaces neither clinical diagnosis nor specialist expertise.

AI may support a situated pedagogical analysis; it must not produce a school, psychological or cognitive identity from a small number of traces.

05 · Withdrawal

Assistance must be able to decrease.

The central criterion remains the same as with any scaffolding: useful support must be progressively removable. If the learner can no longer act as soon as AI disappears, the capacity remains dependent on the support.

The relevant test is not “does the learner succeed with AI?” but “what can they still recognise, explain, verify or transfer when assistance decreases?”

06 · Transfer

Transfer must be tested outside the assisted dialogue.

A fluent conversation with AI can create the impression of a more general mastery than is really present. The report asks that the capacity be tested in several contexts and through new tasks.

Transfer requires the learned structure to remain available when the cues supplied by the tool change or disappear.

07 · Feedback

Automatic feedback is not an autonomous school judgement.

AI may prepare questions, compare some traces or help formulate feedback. It must not decide alone on orientation, sanction or a durable conclusion about the learner.

Evaluation involves criteria, context, responsibility and sometimes major consequences. It must remain under human control and contestable.

08 · Data

Data minimisation is part of the pedagogical architecture.

The report asks that data collection be reduced and AI use made understandable. Technical convenience does not justify accumulating traces about the learner.

The question is not only what AI could infer, but which data are actually necessary for the stated pedagogical function.

09 · Teacher responsibility

The teacher remains responsible for the real learning scene.

AI does not automatically see relationships, school codes, classroom climate, fatigue, the clarity of an instruction or institutional effects.

The teacher remains responsible for the capacity being targeted, the task selected, the relevance of the scaffolding and the interpretation of real effects.

10 · Proof-act

Redo, explain or transfer without delegation.

To test an educational use of AI, choose a precise capacity and verify what the learner can accomplish after the assistance: reformulate in their own words, solve a variation, justify a strategy, correct a response or transfer the method to another context.

The proof-act does not consist in obtaining a better assisted production. It consists in observing whether a capacity becomes genuinely more available in the learner.

Condensation question: After AI assistance, what can the learner now do, verify or transfer without the tool doing it in their place?