Adolescents, ChatGPT and the brain

EEG records the brain’s electrical activity through the scalp, and the study, led by Nataliya Kosmyna, used it to compare how strongly brain regions worked together while people wrote. The group writing unaided showed the strongest connectivity and the group using the language model the weakest. The language model group also struggled to quote from essays they had only just written. The authors gave the name ‘cognitive debt’ to the idea that mental work handed to AI saves effort now and is paid for later.

Two facts about that study matter when reading it. Every participant was between 18 and 39 years old, so no adolescent took part. The study is also a preprint, meaning that it has not yet passed peer review, and a critical comment on its methods has already been published. Our paper adds a caution of its own, that lower brain activity need not mean poorer thinking, because in studies of expertise it often marks more efficient processing.

Executive functions are the processes that let a person plan, hold information in mind, resist impulses and adapt when circumstances change. Brenden Tervo-Clemmens and colleagues combined four datasets covering 10,766 people aged 8 to 35 and found that these functions develop quickly between roughly 10 and 15, then level off, reaching adult-like performance at around 18 to 20. The steep part of the climb belongs to adolescence, and the adults in the Kosmyna study had largely finished it.

The developmental research we review holds that executive functions are shaped by use. On that account, the demands that schoolwork makes on an adolescent are part of how the capacity is built. The tools, meanwhile, have already reached the classroom.

Cognitive offloading means using an external tool to reduce the work the mind has to do, and by that definition a calendar counts, as does a calculator. People worried about calculators too, and those worries proved overblown. What has changed with generative AI is the kind of work that can be handed over.

The paper’s central idea is what we call the cognitive offloading paradox. Generative AI gives the most immediate help with tasks that draw on executive function, such as outlining an essay or judging a source, and those are the tasks whose effortful practice is thought to build executive function during adolescence. When a 16-year-old types ‘write an outline for my essay on climate policy’, the model holds the possible arguments in mind and decides which to keep and in what order. What is left for the student is to judge the result. Younger adolescents monitor their own thinking poorly, and they are inclined to trust text that reads fluently. We suggest that a cost which may be temporary in an adult could, in an adolescent, be developmental. The book makes the longer argument.

The paper separates two ways of using the same tool. Used as a scaffold, the AI supports a student who is still doing the thinking, whereas a substitute produces the finished outline or essay and leaves the student to approve it.

We argue that current tools default to substitution, partly because users tend to prefer tools that minimise effort. The same tool, used the same way, may also scaffold one student and substitute for another. We expect the students least able to manage their own learning, who are often the youngest, to be the most likely to let it do the thinking.

The paper is a research argument, and it gives no guidance for individual families or schools. It does propose a question to ask of any exchange between a student and an AI tool, which is where the cognitive work happened: whether the student generated the options and chose between them, or approved a finished answer. The answer can be read from the record of the conversation, the prompts and the replies, and it is what studies of adolescents will have to measure.

‘None of this has been demonstrated. What follows is a hypothesis.’From the paper’s introduction

The paper does not show that ChatGPT damages adolescent brains. No study has yet followed adolescents over time to see whether a particular way of using AI changes how their executive functions develop. The evidence that involves adolescents directly is correlational, so it cannot tell us whether AI use weakens executive function or whether students with weaker executive function are drawn to AI.

The one brain-imaging study of children that we discuss, led by Tzipi Horowitz-Kraus at the Technion, is a small preprint. It scanned 15 children aged 6 and 7 and 16 adults during a short creative exchange with ChatGPT.

In the scanner

The children’s control and attention networks were less connected than the adults’.

Outside it

Their creativity scores were no different.

A clear difference in the brain came with no difference in what the children made. I read that as a reason to hold back the strongest version of our own argument.

It would be one-sided to set out only the risks, and much of the evidence is encouraging. A meta-analysis of 69 experimental studies by Deng and colleagues found that help from ChatGPT tends to improve academic performance and higher-order thinking, and to lower the mental effort students report. A second meta-analysis, of 29 experiments, found a moderate positive effect on higher-order thinking that was stronger for students who already regulate their own learning well. In the studies we review, the benefits are clearest where the tool prompts and questions whilst the student keeps doing the work. One experiment gave students a chatbot and found that their essays improved and their knowledge did not, a result its authors called ‘metacognitive laziness’. We also note that the published record probably leans towards negative findings, because those draw more attention.

The paper does not call for AI to be banned from classrooms, which it describes as impractical and counterproductive. It does not recommend waiting for conclusive evidence either. Teenagers are using these tools now, and, as we put it in the paper, ‘the developmental window does not pause for the publication cycle’.

We propose five lines of research.

1. Repeat the Kosmyna experiment with adolescents aged 14 to 17.If the reductions in connectivity seen in adults are present or larger at this age, the hypothesis gains support. If they are absent, that is equally important to know.
2. Follow executive function across a school year.Standardised tests at several points in the year, alongside diaries of AI use, in students who use the tools to different degrees and in different ways.
3. Test scaffold against substitute in an experiment.Adolescents assigned at random to an AI that asks questions and prompts revision, an AI that produces the work, or independent work, with their cognitive engagement measured in each condition.
4. Compare early, middle and late adolescence.Groups aged 10 to 13, 14 to 16 and 17 to 20, because the effect is unlikely to be the same at 13 as at 17.
5. Ask adolescents themselves.In a RAND survey we cite, 67 per cent of students agreed that using more AI for schoolwork harms critical thinking.

Does the paper show that ChatGPT harms teenagers’ brains?

No. The paper states that none of this has been demonstrated and that it is putting forward a hypothesis. No study has yet followed adolescents over time to see whether a way of using generative AI changes how their executive functions develop, and the evidence that involves adolescents directly is correlational.

What is the cognitive offloading paradox?

It is the paper’s name for a specific risk. Generative AI gives the most immediate help with tasks that draw on executive function, such as planning an essay or evaluating a source, and those are the tasks whose effortful practice is thought to build executive function during adolescence.

Who took part in the original ‘Your Brain on ChatGPT’ study?

Fifty-four adults aged 18 to 39. The study, led by Nataliya Kosmyna, recorded EEG while participants wrote essays with a large language model, with a search engine or unaided. It is a preprint and has not yet been peer reviewed. No adolescents took part.

Does the paper say schools should ban AI?

No. The paper says a ban would be impractical and counterproductive. It also says that waiting for conclusive evidence is not an option, because teenagers are using the tools now. It asks for research on adolescents, organised around the difference between AI used as a scaffold and AI used as a substitute.

Pereira Campos, J., & Koff, T. (2026). Your brain on ChatGPT, but whose brain? The missing adolescent in AI-cognition research. Frontiers in Developmental Psychology, 4, 1885225. https://doi.org/10.3389/fdpys.2026.1885225

The paper was peer reviewed and published on 8 July 2026. It is open access under a Creative Commons Attribution licence, so the full text is free to read and to reuse with credit.

Read the paperAll publicationsA related essaySpeakingPress kit

Studies named on this page

• Kosmyna, N., and colleagues (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv preprint. arxiv.org/abs/2506.08872
• Stankovic, M., and colleagues (2026). Comment on: Your Brain on ChatGPT. Preprint. arxiv.org/abs/2601.00856
• Tervo-Clemmens, B., and colleagues (2023). A canonical trajectory of executive function maturation from adolescence to adulthood. Nature Communications, 14, 6922. doi.org/10.1038/s41467-023-42540-8
• Klarin, M., and colleagues (2024). Adolescents’ use and perceived usefulness of generative AI for schoolwork. Frontiers in Artificial Intelligence, 7, 1415782. doi.org/10.3389/frai.2024.1415782
• Horowitz-Kraus, T., and colleagues (2025). Lower engagement of cognitive control, attention, modulation networks and lower creativity in children while using ChatGPT: an fMRI study. bioRxiv preprint. doi.org/10.1101/2025.11.07.687207
• Deng, R., and colleagues (2025). Does ChatGPT enhance student learning? A systematic review and meta-analysis of experimental studies. Computers & Education, 227, 105224. doi.org/10.1016/j.compedu.2024.105224
• Zhao, Y., and colleagues (2025). Does generative artificial intelligence improve students’ higher-order thinking? A meta-analysis based on 29 experiments and quasi-experiments. Journal of Intelligence, 13, 160. doi.org/10.3390/jintelligence13120160
• Fan, Y., and colleagues (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56, 489–530. doi.org/10.1111/bjet.13544
• Schwartz, H. L., and Diliberti, M. K. (2026). More Students Use AI for Homework, and More Believe It Harms Critical Thinking: Selected Findings from the American Youth Panel. RAND Corporation. rand.org/pubs/research_reports/RRA4742-1.html