This article cites recent studies published in 2025-2026, some of which are still in pre-publication (not yet peer-reviewed). The conclusions presented are cautious and call for further research. The subject is important, but scientific certainties are built over time.
The question parents are starting to ask
You watch your 10-year-old use ChatGPT to do their homework. They type in their question, the answer arrives in 3 seconds, they copy and paste, they move on. You have a slight unease without quite knowing why. "Is this damaging their brain?"
That question, which seemed paranoid 18 months ago, is now the subject of several serious scientific studies. And the first answers are starting to come in. They are not alarmist in the sense that they do not predict the end of humanity. But they raise precise questions about what happens in the brain when thinking is systematically delegated to a machine. This article takes stock of what the science really says in 2026, following on from our article on the use of AI in education.
The MIT Media Lab study: -55% brain connectivity
Let us start with the most striking study. In June 2025, researchers at the MIT Media Lab (the interdisciplinary laboratory of the Massachusetts Institute of Technology, one of the most recognised universities in the world for science and technology) published a pre-publication entitled "Your Brain on ChatGPT". Pre-publication means it is undergoing peer review and should therefore be treated with the usual methodological caution.
The protocol is sound. 54 participants aged 18 to 39, divided into three groups. The first group writes an essay with the option of using ChatGPT. The second uses only a traditional search engine. The third writes with no external help, just their own brain. The researchers measure brain activity with a 32-channel EEG (electroencephalogram, a technique that measures the brain's electrical activity via electrodes on the scalp) across four sessions spread over several months.
The results are brutal. The ChatGPT group shows the lowest neural connectivity, up to 55% lower than participants with no help. ChatGPT users gradually become more passive: by session 3, they "resort to copy-pasting" in the researchers' words. Worse still, when these ChatGPT users are asked to write on their own in session 4, their brain activity remains suppressed. As if the brain had grown used to delegation and struggled to undo it.
The researchers call this phenomenon "cognitive debt", by analogy with technical debt (a computing concept: shortcuts taken to move fast, which prove costly in the long term). In plain terms: the immediate saving in thinking that AI affords us is paid for later in the form of diminished brain capacity.
The most interesting detail from the MIT study: the participants in the "no-aid" group who started using ChatGPT at session 4 (so the opposite of the first group) showed an increase in their brain connectivity. As if timing mattered: first learn to think for yourself, then use AI as augmentation. The opposite of early delegation, which seems to weaken the brain. According to Nataliya Kosmyna, first author of the study: "timing is probably important".
The study that really shocks: children aged 6-7
The most concerning study was published in November 2025 by an Israeli team led by Prof. Tzipi Horowitz-Kraus (a cognitive neuroscience researcher at the Technion-Israel Institute of Technology, Israel's equivalent of MIT). It is the first time a team has scanned children with fMRI while they interact with ChatGPT. fMRI (functional Magnetic Resonance Imaging) measures brain activity by detecting changes in blood flow in the different regions of the brain.
The protocol: 15 children aged 6 to 7 and 16 adults had a 5-minute co-creative conversation with ChatGPT while their brains were being scanned. They responded to the same prompt and their dialogues were recorded. The researchers measured functional connectivity in seven major brain networks: cognitive control, default mode, memory, dorsal and ventral attention, frontoparietal and salience.
Results. The adults show strong connectivity in the control and attention networks, particularly in the frontoparietal network (FP: brain regions involved in planning and solving complex problems). The children, by contrast, show markedly lower engagement of these same networks. As if their brains, still under construction, "let themselves be carried" by the AI rather than mobilising their own cognitive abilities.
The researchers conclude cautiously: "children's brains are more affected by AI use than adults'". It is a very scientific way of saying what parents feared intuitively.
The counterexample that qualifies everything
To be honest to the end, one study pointing the other way must be cited. Wang et al. published in 2025 in Frontiers in Psychology (an open-access scientific journal dedicated to psychology) an EEG study of 64 design students using creative AI tools (ChatGPT, Midjourney, Stable Diffusion). Result: significantly higher concentration and superior creative performance compared with the control group using traditional software.
Why an opposite result? The key difference is the user's intention. In the MIT study, participants used ChatGPT in passive reception, asking for answers that they integrated. In the Wang study, the students actively directed the AI as a creative tool, keeping control of the process.
This nuance is crucial. It suggests that AI is not bad for the brain in itself. What is bad is passive delegation. AI used as a tool that augments human capabilities preserves cerebral engagement. AI used as a substitute for thought diminishes it. As we explored in our article on hypocrisy towards AI, the question is not "is AI good or bad", but "how is it used".
Why children's brains are particularly vulnerable
To understand why researchers are worried about children specifically, we need to understand how the human brain develops. A few useful neuroscientific reminders.
The plasticity window. The human brain is said to be "plastic" (capable of changing) at any age, but this plasticity is exponentially stronger in children. Between the ages of 0 and 12, the brain builds its main circuits. The habits formed during this window leave lasting traces. It is during this period that we learn to speak, read, do arithmetic, but also to concentrate, to tolerate effort, to persevere. If AI becomes the habit for providing instant solutions, these effort circuits may not build correctly.
The development of executive functions. "Executive functions" are the set of cognitive abilities that allow us to plan, to inhibit distractions, to sustain attention, to memorise actively. They develop mainly between the ages of 4 and 25 in the prefrontal cortex (the front part of the brain, behind the forehead). If these functions are not regularly called upon by real cognitive challenges, their development may be slowed.
Effort as cerebral nourishment. Contemporary neuroscience converges on one observation. The brain needs cognitive effort to develop, just as a muscle needs resistance to grow. Struggling to memorise a poem, searching for a word for a long time, labouring over a maths exercise: these uncomfortable moments are precisely the ones that build abilities. If they are systematically avoided with AI, the brain does not receive what it needs to build itself.
Think of a child who is always given pre-chewed food. They would not fall ill, but they would not learn to chew, their jaw muscles would not develop properly, their digestion would be affected. In the long term, they would suffer for it, without any single moment being "the" problem. AI for children could work like this: no dramatic immediate effect, but a silent accumulation of undeveloped skills.
The figures that worry researchers
To understand the urgency, we need to look at the scale of the phenomenon. In 2025-2026, children's use of AI is exploding.
- 26% of American teenagers aged 13 to 17 used ChatGPT for their homework in 2024, compared with 13% in 2023 (Pew Research Center, one of the largest American statistical research institutes)
- 72% of American teenagers have tried a personified AI companion such as Character.AI, as we documented in our article on AI companions
- 86% of students use AI in their studies according to HEPI (the British Higher Education Policy Institute)
- 52% of teenagers chat with an AI every day
- Snapchat My AI has 150 million users, mainly teenagers
For younger children, the figures are less well studied but probably comparable. Children aged 6-10 use AI via their parents' tablets, via family voice assistants (Alexa, Google Home), via educational apps. Most children born after 2018 will be the first generation to grow up with generative AI as a default tool. This transition is historically unprecedented.
The specific areas that could be affected
Several areas of cognitive development are potentially affected. Let us list them with caution.
Memory. If information is always accessible within 3 seconds via an AI, the brain learns that it does not need to memorise. Researchers speak of the "Google effect", already documented before AI: we remember better where to find a piece of information than the information itself. AI could amplify this effect to problematic levels. Yet living memory is an essential substrate of complex thought.
Sustained attention. Holding your attention for 30 minutes on a difficult subject is a skill that has to be built. If AI provides answers broken down into a few seconds, children never learn to sustain long attention. Several studies converge: adolescents' sustained attention capacities have fallen by 30 to 50% over 15 years according to standardised norms.
Problem solving. Solving a difficult problem involves trying several approaches, failing, starting again. This effort trains specific brain circuits. If AI gives the solution directly, those circuits do not develop. Children may become competent at asking questions of an AI, but incompetent at solving on their own.
Writing. Writing means structuring your thought. When you look for the right word, rephrase a sentence, hesitate between two structures, you learn to think. If AI writes in our place, we lose that fundamental exercise. Several French secondary-school teachers report it: the essays are technically perfect (probably AI-generated) but the actual spoken skills show an impoverishment.
The ability to get things wrong and learn. AI almost always gives the right answer (bar a few hallucinations). Yet a crucial part of learning comes from error: getting it wrong, understanding why, correcting it. If children no longer get the chance to get simple problems wrong (because AI spares them that), they may never develop this metacognitive ability.
What cannot (yet) be measured
The studies cited all concern brain activity during or just after AI use. Nobody knows what happens at 10, 20 or 30 years. The first children to have used ChatGPT at age 6 in 2023 will be 9 this year. The first serious longitudinal data will probably arrive around 2030-2035.
Until then, we are navigating blind. Parents are making educational choices with no visibility on their long-term consequences. Schools are deploying or banning AI without solid data. Governments are regulating or leaving things be without a stable scientific basis. This ignorance is uncomfortable but real.
A few indirect signals exist. Average IQ measured in developed countries has stagnated or fallen since the 2000s, after a century of continuous increase (the reversed Flynn effect, documented by James Flynn, a New Zealand psychologist). Several researchers attribute this decline to smartphones and screens, with AI now adding to these factors. But the causal links remain to be demonstrated.
What neuroscience concretely recommends to parents
Several research teams are beginning to publish practical recommendations, based on the current state of knowledge. The main points of convergence.
Before age 8: limit as much as possible. The brain is building its fundamental circuits of effort and attention. This period is precious and fragile. Most researchers recommend keeping AI almost entirely at a distance until at least age 8, ideally 10. Exposure to screens in general should also remain limited (the WHO recommends no more than 1 hour a day before age 5, ideally none before age 2).
Between 8 and 14: supervised and limited use. The child can start using AI for specific purposes (exploring a subject they are passionate about, checking a piece of information), but under parental supervision and for limited periods of time. The reflex of always asking AI must be countered, in favour of independent research, reading and human dialogue.
Between 14 and 18: supervised autonomy. The teenager can use AI for real schoolwork, but must learn to direct the AI rather than passively receive from it. Ask them why they accept a given answer, how they verified it, what they learned beyond the answer itself.
At any age: protect effortful rituals. Deep reading, handwriting, memorisation, mental arithmetic, sport, music played on a real instrument, manual crafts. These activities engage the brain circuits that AI risks atrophying. They should remain central to any education, even in the AI era. As we explored in our article on what we pass on to our children, these "non-transferable" skills paradoxically become more valuable in the AI era.
The generational contrast taking shape
If the neuroscientists' hypotheses are confirmed, we could see an unprecedented generational contrast over the next 20 years.
Generation before 2010 (who learned without AI, and uses AI as an adult): "classic" cognitive abilities + AI augmentation = clear advantage.
Generation 2010-2020 (who discovered AI in adolescence): partially developed abilities, still conscious use of AI = mixed situation, to be observed.
Generation born after 2018 (who grow up with AI from childhood): real risk of underdeveloped cognitive abilities if use is heavy and passive. Potential advantage if use is well supervised.
This polarisation could create a cognitive inequality between children raised with firm parental limits (a brain more classically developed) and children left with free access to AI (a brain potentially impoverished in certain respects). Educated families aware of the risks will probably set strict limits. Less informed families will let their children use it heavily. The result at 15-20 years could be a major cognitive gap between social classes, even wider than the one cultural access inequality already creates.
The children of Silicon Valley tech executives are overwhelmingly schooled in screen-free schools (Waldorf, Steiner, Montessori). The founders and leaders of tech companies (former executives of Apple, Google, Meta) ban or strictly limit screens for their own children. This dissonance between what they sell to the world and what they consume at home should give us pause. They know something that the marketing communication for their products hides.
The wider debate opening up
Beyond the figures and the studies, this subject raises major educational and political questions for the years ahead.
Should AI be banned in schools before a certain age, as mobile phones were banned in schools in France in 2024? Several European countries are considering it. Opponents say it is backward-looking and that we should instead teach critical use. Supporters say it is better to preserve cognitive functions that are still developing.
Should AI publishers be held responsible for use by children, as tobacco or alcohol manufacturers are held responsible? OpenAI announced in 2025 a "ChatGPT for children" version with reinforced protections. But these protections remain largely declarative and easily circumvented.
Should neuroscience research on these topics be massively funded, when budgets remain derisory compared with the profits of AI companies? In the United States, the NIH (National Institutes of Health, the main federal medical research agency) currently devotes around 50 million dollars a year to studying the cognitive effects of screens and AI. By way of comparison, OpenAI invests that sum in a few days on infrastructure.
These political choices are being made, or not being made, without major public debate. As we explored in our article on the limits of AI regulation, regulators are structurally several years behind the technology. On this specific issue, that lag could prove costly.
What we should honestly take away
Three cautious observations by way of conclusion.
First observation: the early scientific data are concerning but not catastrophic. The MIT and fMRI studies on children suggest real negative cognitive effects in cases of passive use. These effects are probably not irreversible in the short term, but they accumulate. "Cognitive debt" is a useful concept for thinking about the phenomenon.
Second observation: how AI is used matters more than AI itself. Passive delegation = negative effects. Active direction = potentially positive effects. This distinction should shape all future education policy.
Third observation: we don't have the long-term data and we won't have it for another 10 years. This ignorance demands caution. When in doubt about what properly shapes a child's brain, it is better to refrain from massive exposure than to regret it in 20 years.
If you are a parent, the practical message is simple. Protect your child's time. Cultivate the activities that genuinely engage their brain: reading, construction games, sport, music, conversation, imagination. AI can remain an occasional tool, in limited doses, supervised. It must not become the default mode of access to knowledge and thought. To go further, you can consult our article on AI in education, which describes how usage is actually unfolding in schools, and our article on what we pass on to our children, which raises the complementary question of what remains worth teaching.