Children and AI: productive resistance and the question of cognitive debt

An answer arrives. What happened to the thinking?

A child who copies an AI explanation may hand in the right words without having chosen the reasoning behind them. Another child might use the same system to question a first attempt, check an unfamiliar term and then work through a new example independently. Both have used AI. Their opportunities to learn are very different.

Three terms help describe that difference: cognitive offloading, productive resistance and cognitive debt. They are connected, but they do not refer to the same thing. Offloading describes what work is delegated. Productive resistance describes a way of protecting useful thought during an activity. Cognitive debt names a concern about what may remain undeveloped when the effort needed for learning is repeatedly bypassed.

Offloading is ordinary. The learning goal changes its meaning.

External aids are part of learning. A diagram can help organise ideas; a reminder can free attention for a demanding task. The relevant question is whether the aid removes a barrier or takes over the target skill. If the lesson is about comparing evidence, asking a system to choose the strongest evidence can remove the central learning opportunity. If the lesson is about interpreting a result, help with an unrelated formatting task may leave that opportunity intact.

This distinction is especially relevant for children who are still acquiring foundational skills. Adults may already have knowledge they can use to question a generated answer. A learner meeting a concept for the first time may have fewer ways to recognise a plausible mistake or notice that a step has been skipped. That is a reason to make adult guidance and the learning purpose explicit, rather than treat convenience as sufficient evidence of educational value.

Productive resistance: preserve the thinking that matters

Here, productive resistance means deliberately keeping a useful challenge in the learning process: make a prediction before seeing the explanation, identify a reason, test a counterexample or decide which revision to accept. It is a teaching approach, not a clinical term or a universally standardised intervention. Resistance is productive only when the learner has a realistic route forward and appropriate support.

Advait Sarkar’s proposal for AI as a provocateur offers a related design perspective. Instead of simply completing work, a system could question assumptions, introduce alternatives and prompt critical evaluation. The proposal concerns tools for thought and knowledge work; it does not prove that a particular chatbot improves children’s learning. It helps explain why an educational tool may need to ask a useful question rather than supply the quickest complete answer.

For an illustrative science task, a child first predicts which object will float and gives a reason. A permitted AI prompt asks for one case that might challenge that reason. The child then compares the suggestion with an observed result and revises the explanation. The learning record contains the child’s prediction, evidence and decision. The system contributes a challenge; it does not decide what the child should believe.

Cognitive debt: a concern to investigate, not a diagnosis

The phrase gained attention through Kosmyna and colleagues’ Your Brain on ChatGPT preprint. The study compared essay writing with an LLM, search and no tools. It involved 54 participants in the first three sessions, with 18 completing a fourth session. The authors reported differences in EEG connectivity, recall and reported ownership of the writing. The study concerned adult essay writers, not children. The project page describes preliminary findings and substantial limits; the paper should be read as a preprint.

Cognitive debt is a useful question about learning that may have been deferred: what will the learner be unable to explain or do later because a tool performed the relevant work now? It is not an individual medical diagnosis, an established measure of permanent damage or a conclusion that can be read directly from a child’s screen time. The adult study does not establish a long-term developmental effect in children.

A practical sequence: attempt, question, explain, transfer

1. Attempt: ask for a brief first response, sketch or prediction that reveals the learner’s starting point.
2. Question: offer a hint, clarification or counterexample that helps the learner inspect the difficulty.
3. Explain: ask what changed and why. Keep the learner responsible for choosing the final reasoning.
4. Transfer: return later with a fresh example or discussion that uses the same idea without the generated answer.

The sequence is not a requirement to withhold help until a child becomes frustrated. An unclear instruction, missing prerequisite or inaccessible format may need immediate correction. A child who needs assistive technology should continue to receive it. Remove unnecessary barriers while keeping the intended intellectual work visible.

What parents and schools can agree on

Instead of a vague instruction to “use AI responsibly”, identify what assistance is allowed for a specific task. Can a child ask for a definition? A hint? Feedback on their own draft? A complete worked example? State how assistance should be acknowledged and what the child should still be able to explain. Different tasks may reasonably have different boundaries.

UNICEF’s 2025 guidance places children’s development, safety, privacy and inclusion among the requirements for child-centred AI. That wider lens matters: an activity should not require a child to disclose personal information or use an unsuitable service just to demonstrate a learning skill. Schools should check age eligibility and institutional arrangements before selecting a tool.

The goal is a more capable learner

The relationship between these concepts is practical. Decide what can be offloaded, preserve a useful amount of challenge, and check for understanding that has been deferred. The measure of success is a child who can question, explain and apply more of an idea afterwards. A polished answer becomes educationally valuable when the learner can make sense of it and carry the understanding into the next task.