---
title: "What is the impact of AI on the brain?\nWhat does scientific research say?"
description: "In 1 trial, students using an AI that answers freely scored 17% lower in the exam than those without AI. None of the 3 studies goes beyond 4 sessions."
canonical: https://projetcelsius.com/en/blog/ai-impact-on-the-brain/
author: "Guillaume Pakula"
datePublished: 2026-10-05
dateModified: 2026-10-05
language: en
translationOf: https://projetcelsius.com/blog/impact-ia-cerveau/
---

# What is the impact of AI on the brain? What does scientific research say?

**Author**: Guillaume Pakula
**Published**: 2026-10-05
**Section**: Explained
**Canonical URL**: https://projetcelsius.com/en/blog/ai-impact-on-the-brain/
**French original**: https://projetcelsius.com/blog/impact-ia-cerveau/
**Source**: Projet Celsius (https://projetcelsius.com/en/), a climate strategy consultancy of engineers and PhDs based in Paris and Marseille.

> **In a 2023 trial at 1 high school in Turkey, classes that revised mathematics with an AI that answers freely scored 17% lower in the exam than classes without AI.** With an AI that gives hints, the exam showed no measurable difference. The only one of the 3 studies selected to record brain activity covers 54 people and has not yet appeared in a peer-reviewed journal. Follow-ups lasting several months, posted online in 2026 without peer review, track grades; none measures the brain.

## Key takeaways

- In a high school trial, classes using an AI that answers freely scored **17% lower in the exam**.
- With the AI that gives hints, the same trial found **no measurable difference** in the exam.
- In a preprint of 54 people, electrode recordings found **brain signals less linked to one another** with ChatGPT.
- **Lasting effect on the brain: not measured.** The longest follow-ups, of up to 30 months, track grades.

The 3 studies we selected on the **impact of AI on the brain** measured 3 different things, during a task or just after it: **grades**, answers to a questionnaire and **electrical activity in the brain**. The study with the strongest evidence is a randomised controlled trial, an experiment in which groups are assigned at random. Run in 2023 at 1 high school in Turkey, it found that classes that revised mathematics with an AI that answers freely scored **17% lower in the exam** than classes without AI. None of the 3 studies goes beyond 4 sessions.

The only one of the 3 to record brain activity covers **54 people** and has not yet appeared in a peer-reviewed journal. Follow-ups of between 1 semester and 30 months were posted online in 2026 without peer review; they look at **grades**, and none measures the brain. This page was **updated on 5 October 2026**.

## The 3 studies: 1 published trial, 1 survey, 1 preprint

The first of the 3 studies selected, all from 2025, is a **trial** published in a **peer-reviewed** journal, meaning that other researchers check the work before publication. The second is a questionnaire survey, published in conference proceedings that are also peer-reviewed. The third is an experiment with random assignment on a small sample; it is a **preprint**, posted online without that check.

The trial appeared in June 2025 in the US scientific journal Proceedings of the National Academy of Sciences (PNAS). In autumn 2023, at 1 high school in Turkey, **nearly 1,000 students** revised mathematics over 4 sessions of 90 minutes. Classes were assigned at random to work without AI, with an **AI that answers freely** (GPT-4 in a chat interface, like ChatGPT) or with an **AI that gives hints** without handing over the answer, drawing on solutions prepared by teachers. Each session ended with an unaided exam.

Compared with the classes without AI, the AI that answers freely raised scores on the exercises by 48% and **lowered them by 17% in the exam**: 27 out of 100 instead of 32 (our calculation). Allowing for statistical uncertainty, the drop lies between about 3% and 30% (95% confidence interval, our calculation). The AI that gives hints raised scores on the exercises by 127%, with **no measurable difference in the exam** from the classes without AI. The authors write that they measured only short-term effects.

The survey was presented in April 2025 by 7 researchers from Microsoft Research and Carnegie Mellon University. It covered **319 people** who use AI at work at least once a week; between them they described **936 examples of use**. The more they said they trusted AI, the **less they reported exercising critical thinking**; the more self-confidence they had, the more critical thinking they reported. This statistical link between self-reports does not establish cause and effect.

The third study, "Your Brain on ChatGPT", was posted online in June 2025 by 8 researchers from the Massachusetts Institute of Technology (MIT) and 2 other institutions. It randomly assigned **54 people**, mostly students, to **3 groups of 18**: [ChatGPT, a search engine](https://projetcelsius.com/en/ai/analyses/encyclopedia-google-chatgpt-learning/) or no tool. Each participant wrote 3 texts of 20 minutes while wearing electrodes on the scalp.

At the end of session 1, **15 of the 18 participants in the ChatGPT group** could not quote a sentence from their text, against 2 of 18 in each of the other 2 groups. On the question that followed, none of the 18 gave an exact quotation.

In sessions 2 and 3, **4 then 6 participants out of 18** in the ChatGPT group did not provide an exact quotation, against 2 of 18 in each of the other groups. The authors report no statistical test for these 2 sessions; they write that the difficulty **persisted but eased**, and that the question may have surprised participants the first time.

## What brain imaging shows: signals less linked with ChatGPT, in a preprint

Only 1 of the 3 studies selected recorded brain activity while AI was being used. Using an electroencephalogram (EEG) on 54 people in 3 groups of 18, it measured **signals less linked to one another** with ChatGPT than with no tool. As of early October 2026, there are also at least **2 MRI studies** (magnetic resonance imaging), which are preprints too: **none compares work done with and without AI**.

In "Your Brain on ChatGPT", 32 electrodes placed on the scalp record the brain's electrical activity. This technique **does not produce an image** of the anatomy, and the authors ask readers not to use the term "brain scans". From the recordings they derive a measure of connectivity, which is the degree to which the signals picked up by the different electrodes are linked to one another during the task.

This connectivity was strongest in the group with no tool, intermediate with the search engine and **weakest with ChatGPT**: according to the authors, up to 55% lower than with no tool in some of the signals. In their abstract, they infer "a likely decrease in learning skills" from their results. Their discussion refers to long-term costs, which their protocol of 3 sessions of 20 minutes **did not measure**.

The authors write that EEG does not precisely locate the deep structures of the brain, and name **functional MRI**, which tracks brain activity, as their next step. Every page of their document is marked "Preprint, under review", and on 5 October 2026 its entry on arXiv, the site where it is posted, listed **no journal**.

On 29 December 2025, 4 researchers from the universities of Vienna and Dresden posted a commentary online, also a preprint. By their calculation, a design of this kind would need **about 159 participants**, nearly 3 times the size of the study (our calculation). They note that some results are reported in contradictory ways from one page to another and that, in some of the signals, engagement was **stronger with ChatGPT** than with no tool. They judge **some interpretations overstated** and find **no strong evidence** that using AI degrades performance itself.

The first of the 2 MRI studies, posted online in November 2025, had **15 children aged 6 and 7 and 16 adults** converse with ChatGPT for 5 minutes inside the MRI scanner. The brain networks that support control and attention were less linked to one another in the children than in the adults. **The study has no group without AI**: a comparison between 2 ages does not show what AI changes.

The second of these MRI studies compares the brain activity of 12 people judged expert at writing prompts for an AI with that of 10 people at an intermediate level, recorded with no task to perform. A third preprint, using near-infrared spectroscopy (another way of measuring brain activity), recorded, in 20 people, no physiological change linked to the assistant in Word.

## Does AI have a lasting effect on the brain? None of the studies measures it

**None of the studies cited here measures a lasting effect on the brain**; those that run for several months track grades. The high school trial consists of 4 sessions of 90 minutes, with an exam on the same day. In "Your Brain on ChatGPT", each participant wrote 3 texts of 20 minutes; the 4 months cited in the abstract refer to **the length of the data collection**. Neither of these 2 studies measures whether **the ability to learn or to memorise** changed afterwards.

The MIT study includes a session 4, the only session in which the tool changes: **18 volunteers** came back, 9 former ChatGPT users wrote with no tool and 9 participants from the no-tool group used ChatGPT. On recall of the text, the study counts 7 failures out of 9 in one group and 1 out of 9 in the other, but **its pages do not agree** on which group, as the critical commentary points out.

A trial from Brazil, published in late 2025, measured memory after several weeks. It gave a surprise test to **85 students** who had worked on the same subject **45 days** earlier, with or without ChatGPT: **57.5% correct answers with ChatGPT**, against 68.5% without. Of the 120 students assigned at random, 35 did not take this test.

The authors of "Your Brain on ChatGPT" call for studies that **follow the same people over time**. As of early October 2026, there were at least 3 follow-ups of 1 semester or more on generative AI, posted online in 2026. They are working papers, **not peer-reviewed**, and they track **grades**.

In 1 county in China, a 30-month follow-up of **26,811 middle and high school students**, with no random assignment, finds that adopting an AI is associated with an 18% rise in homework grades and a **20% fall in monthly exam scores** within 6 months. The loss is small among AI users whose homework time stayed close to that of non-users. In an op-ed, the authors caution that these orders of magnitude do not carry over directly to other settings, though they see them as a warning.

In **18 middle schools in Tennessee**, a 2-year trial had students work on the Khan Academy exercise platform with its AI tutor, set up to guide. It found **a small gain in mathematics**, which its authors judge close to the gain from the platform used without AI; the students made little use of the tutor.

At a large public university in the United States, a trial with 2,379 students gave some of them access, for 1 semester, to an AI tutor whose default setting **gave the explanation directly**. Between groups within the same course, the final grade was **0.37 standard deviations lower** with this access (a standard deviation is a measure of how spread out grades are); across all students, the gap is smaller and statistically less certain. This final grade is the one the instructor awards for the course, with no independent test, and the authors judge it hard to infer an effect on learning from it.

## Are schoolchildren and university students more exposed? The studies do not say

**None of the studies cited here measures whether AI affects schoolchildren more than adults.** In the high school trial, the students who had the AI that answers freely most often asked it for **the answer**; the authors see this as **a crutch** and give it as their explanation for the drop in the exam. When the authors put each of the 57 exercises to it 10 times, this AI gave the right answer in only **51% of cases**.

The students **were not aware of the drop**: those in the group with the AI that answers freely did not think they had done worse in the exam, and those in the group with the AI that gives hints thought they had done better, although their scores did not bear this out.

**The design of the tool** changed the result. With the AI that gives hints, the exam score stayed on a par with that of the classes without AI. At Harvard, a trial published in June 2025 covered **194 students** in an introductory physics course: those who followed a lesson at home with an AI tutor whose instructions were written by teachers **did better in the test** they took straight afterwards than those who followed the same lesson in class, in groups. Its authors do not assume that this result holds in every context.

In January 2026, Inserm, the French public medical research institute, published **the opinions of 3 researchers**; the page cites no study on the brain. The neurobiologist Jean-Philippe Lachaux describes cognitive offloading, the act of entrusting a mental task to a tool. In his words (our translation, as for the other opinions below), this delegation encourages the loss of skills that go unused. He suggests that students **predict the answer** before asking the AI, so as to learn from the gap. Cédric Naudet, a specialist in education sciences, fears that **the educational divide** will widen. The researcher Margarida Romero considers that everything depends on the tool and how it is used.

**Only 1 of the studies cited here involves primary school pupils**: one of the 2 MRI preprints, which compares 15 children aged 6 and 7 with adults. The participants in the other studies are middle school students, high school students, university students or working adults. In France, the framework for the use of AI in education published by the Ministry of National Education in June 2025 provides for awareness-raising **without hands-on use** of generative AI in primary school. It authorises its use in class **from the 4e class** (the third year of collège, the French lower secondary school), under the teacher's supervision. The framework treats as **fraud** the use of an AI to do homework without permission and without personal work.

## Learning with AI: only answers versus hints were tested

The studies cited here tested **only 1 difference in use**: receiving the answer or receiving hints. Trying on your own before asking rests on an opinion and a hypothesis; checking the answers rests on what users reported.

The authors of the high school trial write that **hints designed by teachers** may mitigate the negative effect; they **did not test** the case of a student who asks an AI that answers freely for a hint. The authors of "Your Brain on ChatGPT" have a hypothesis about starting without a tool; it rests on the **18 participants in their session 4** and on a result that their document reports in 2 contradictory ways.

In the same series: [the impact of AI on jobs](https://projetcelsius.com/en/blog/ai-impact-on-jobs/), [its environmental impact](https://projetcelsius.com/en/blog/environmental-impact-of-ai/), [whether to boycott AI](https://projetcelsius.com/en/blog/should-you-boycott-ai/), [reducing its impact](https://projetcelsius.com/en/blog/reduce-ai-environmental-impact/) and [choosing a more eco-friendly AI](https://projetcelsius.com/en/blog/most-eco-friendly-ai/). The MIT study devotes a section to the energy cost of a query, which we detail in [how much energy a ChatGPT query uses](https://projetcelsius.com/en/blog/chatgpt-query-energy-consumption/).

## Key takeaways

- **High school trial**: 17% lower in the exam with the AI that answers freely, compared with classes without AI; no measurable difference with the AI that gives hints.
- **Study with electrodes**: brain signals less linked to one another with ChatGPT, in a preprint with 3 groups of 18 people.
- **Quoting your own text**: 18 of 18 participants in the ChatGPT group gave no exact quotation in session 1, then 4 and 6 in the next sessions, against 2 or 3 out of 18 in the other groups.
- **Lasting effect on the brain: not measured.** The follow-ups of between 1 semester and 30 months posted online in 2026 concern grades, and they are working papers that have not been peer-reviewed.

## Frequently asked questions

### Does AI make you dumber?

None of the studies cited here shows this, and none measures intelligence. Asked whether these tools can be said to make us "dumber", the authors of "Your Brain on ChatGPT" answer "No!" and ask that such words be avoided; they do, however, infer from their results a likely decrease in learning skills, in a study of 54 people that has not yet appeared in a peer-reviewed journal. In a trial conducted at 1 high school in Turkey, classes that revised with an AI that answers freely scored 17% lower in the exam than classes without AI, while an AI that gives hints produced no measurable difference.

### What did the MIT study find about ChatGPT and the brain?

In this study, 54 people, mostly students, were assigned at random to 3 groups of 18 and wrote 3 texts of 20 minutes while an electroencephalogram (EEG) recorded their brain activity. The signals recorded were most closely linked to one another with no tool and least with ChatGPT. In session 1, 15 of the 18 participants in the ChatGPT group could not quote a sentence from their text, against 2 of 18 in each of the other 2 groups; in the next 2 sessions, 4 then 6 of 18 did not provide an exact quotation, against 2. This study, posted online in June 2025, has not yet appeared in a peer-reviewed journal, and a commentary by 4 researchers judges its sample too small.

### Does AI damage your memory?

Only 1 of the studies cited here, a preprint, measured recall of a text written with ChatGPT, straight after the writing: in session 1, 15 of 18 participants could not quote a sentence from it, against 2 of 18 in the other groups, and the gap was smaller in the next 2 sessions. None of the studies cited here on generative AI measures memory in general. A trial from Brazil, published in late 2025, gave a surprise test to 85 students who had worked on a subject 45 days earlier, with or without ChatGPT: 57.5% correct answers with ChatGPT, 68.5% without.

### Are the effects of AI on the brain reversible?

None of the studies cited here measures this. In "Your Brain on ChatGPT", the number of participants in the ChatGPT group with no exact quotation fell from 18 to 4 as early as session 2, even though they were still using ChatGPT. Session 4, in which 9 former users wrote with no tool, gives a result on this point that the document reports in 2 contradictory ways.

### Is AI dangerous for children's and teenagers' brains?

Only 1 of the studies cited here involves primary school pupils: an MRI preprint that observed 15 children aged 6 and 7 for 5 minutes of conversation with ChatGPT and compared them with adults. It has no group without AI. The high school trial involves teenagers: it measured grades in mathematics, without observing the brain. The other studies involve middle school students, high school students, university students or working adults. In France, the Ministry of National Education provides for awareness-raising without hands-on use in primary school and authorises the use of generative AI in class from the 4e class (the third year of collège, the French lower secondary school), under the teacher's supervision. This page reports studies; for a health concern, the right person to consult is a health professional.

### How can you use ChatGPT to learn?

The studies cited here tested only one difference in use: receiving the answer or receiving hints. In the high school trial, the AI that gives hints, whose instructions included solutions prepared by teachers, produced no measurable difference in the exam, whereas the AI that answers freely lowered the score by 17%. The other 2 approaches were not tested: trying on your own before asking (a researcher's opinion and a hypothesis) and checking the answers (a practice reported in a survey).

## Sources

- **The 3 studies**: [Bastani et al., Generative AI without guardrails can harm learning: Evidence from high school mathematics](https://doi.org/10.1073/pnas.2422633122), PNAS, vol. 122, No. 26, 25 June 2025, and [the manuscript posted online by the author](https://hamsabastani.github.io/education_llm.pdf). [Lee et al., The Impact of Generative AI on Critical Thinking](https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/), proceedings of the CHI conference, April 2025. [Kosmyna et al., Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task](https://arxiv.org/abs/2506.08872), arXiv preprint, version 1 of 10 June 2025, version 2 of 31 December 2025, with [the project website](https://www.brainonllm.com/) and [its FAQ](https://www.brainonllm.com/faq).
- **The critical commentary**: [Stanković, Hirche, Kollatzsch and Doetsch, Comment on: Your Brain on ChatGPT](https://arxiv.org/abs/2601.00856), arXiv, 29 December 2025. It is also a preprint.
- **The other studies**: [Kestin et al., AI tutoring outperforms in-class active learning](https://www.nature.com/articles/s41598-025-97652-6), Scientific Reports, 3 June 2025: 194 students included out of 233 enrolled, 2 lessons, test taken straight after each lesson. [Barcaui, ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention](https://doi.org/10.1016/j.ssaho.2025.102287), Social Sciences & Humanities Open, vol. 12, 2025: 120 students assigned at random, 85 tested at 45 days. Working papers, not peer-reviewed: [Strömberg, Lei and Wu, The Generative AI Learning Penalty: Evidence from Chinese Secondary Education](https://cepr.org/publications/DP21577), CEPR, 2 June 2026, and [their op-ed on VoxEU](https://cepr.org/voxeu/columns/generative-ai-learning-penalty-secondary-school), 4 September 2026; [Oreopoulos and Low, One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment](https://www.nber.org/papers/w35620), NBER, August 2026; [Liu et al., The Effects of Course-Integrated AI Tutoring on Student Performance and Engagement](https://edworkingpapers.com/ai26-1598), EdWorkingPaper 26-1598, posted online on 30 September 2026. Imaging preprints: [Horowitz-Kraus et al., Lower engagement of cognitive control, attention, modulation networks and lower creativity in children while using ChatGPT: an fMRI study](https://www.biorxiv.org/content/10.1101/2025.11.07.687207v2), bioRxiv, November 2025; [Al-Khalifa et al., The Prompting Brain](https://arxiv.org/abs/2508.14869), arXiv, 2025; [Russell et al., Neural and Cognitive Impacts of AI](https://arxiv.org/abs/2506.04167), arXiv, June 2025.
- **In France**: [Inserm, Quel est l'impact de l'IA sur l'éducation ?](https://www.inserm.fr/actualite/quel-est-limpact-de-lia-sur-leducation/) (What is the impact of AI on education?, in our translation), 5 January 2026: opinions of Jean-Philippe Lachaux, Margarida Romero and Cédric Naudet. [Ministry of National Education, Cadre d'usage de l'IA en éducation](https://www.education.gouv.fr/cadre-d-usage-de-l-ia-en-education-450647) (Framework for the use of AI in education, in our translation), 13 June 2025. Both pages are in French.
- **Images**: Opening photo: John Campbell, [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Brain_model_12.jpg), CC BY 2.0, cropped. Electroencephalography cap: Thaismay, [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Treinamento_de_EEG-57.jpeg), CC BY-SA 4.0. Exam desks: Jérémy Barande / École polytechnique, [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Exam_tables_-_France.jpg), CC BY-SA 2.0.
