AI and your brain: every study on AI use and thinking

01 Key findings

1. The evidence is split. Of 56 studies from 2023 to October 2026, 19 found worse thinking, learning or memory with generative AI, 16 better, 17 mixed and 4 no clear difference.56 studies, 44 peer reviewed
2. Once the AI is gone, results depend on the design. Of 25 experiments that tested people later without AI, 8 found they did worse, such as 17% lower exam scores after plain GPT-4 ^[7], and 10 found they did better, often with tutors that guide instead of answering ^{[31] [37]}.25 randomized or controlled experiments
3. Doing the work for you is the risk. People who let AI hand them answers learned less in several trials, such as quiz scores 17% lower after learning a coding library with AI ^[44] and 57.5% vs 68.5% correct on a test 45 days later ^[5].Shen and Tamkin 2026, 52 developers; Barcaui 2025, 120 students
4. Surveys point the same way but cannot show cause. 7 of 12 surveys linked heavier AI use to weaker critical thinking, memory or grades, including the Microsoft and Carnegie Mellon survey of 319 knowledge workers ^[39].12 surveys, 319 to 666 people each
5. Brain studies are small and early. The 11 EEG and fNIRS studies had 24 to 95 participants; the best known, MIT’s 54-person preprint, found the weakest brain connectivity in ChatGPT users ^[32].11 brain-activity studies
6. One widely cited review was withdrawn. A 2025 meta-analysis of 51 studies that reported large learning gains from ChatGPT was retracted on 22 April 2026 over discrepancies in its analysis ^[62].Humanities and Social Sciences Communications 13, 528

02 Statistics to quote

03 Does using AI make you think worse?

Not in any simple way. We tracked down every study we could find, published from January 2023 to 15 October 2026, that measured what using ChatGPT or another generative AI tool does to people’s own thinking, learning, memory, creativity or brain activity, and read each one. 56 met our rules: 33 experiments, 11 brain-activity studies and 12 surveys, with 29,629 participants between them. 44 are peer reviewed; the rest are preprints or working papers.

Figure 1

What 56 studies of AI and thinking found

Number of studies by their main result for people's own thinking, learning or memory. Surveys show links, not causes. Mixed = better on one measure or for some people, worse on another.

19Worse34%

17Mixed30%

4No clear difference7%

16Better29%

34%30%29%

The kind of evidence matters more than the count. Surveys of AI users most often found worse results, but they measure people once and rely on what people say about their own thinking. Experiments that look at the work people produce with AI usually find better work, often with a cost such as overconfidence or more alike ideas. The experiments that matter most for “your brain” test people afterwards, without the AI.

Figure 2

What each type of evidence found

Share of the studies in each group, by main result. Surveys show links, not causes.

WorseMixedNo clear differenceBetter

Surveys of AI users12 surveys58%17%25%

Tested later, without AI25 experiments32%16%12%40%

Work done with AI8 experiments25%63%13%

Brain activity11 studies18%55%9%18%

04 Does AI hurt learning once you stop using it?

Sometimes. 25 experiments let some people use AI while they practised or studied, then tested everyone without it. 8 found the AI group did worse, 10 better, 3 found no clear difference and 4 found both, depending on the student or the task. A July 2026 comparison of 13 randomized trials with students reached an average gain of 0.18 standard deviations on tests taken without AI, with losses where the AI did the practice in the learner’s place ^[12].

Figure 3

The 25 experiments that tested people later, without AI

Number of experiments by their main result once the AI was gone.

8Worse32%

4Mixed16%

3No clear difference12%

10Better40%

32%16%12%40%

Did worse once the AI was gone

• Melumad and Yun (2025), 10,462 people, United States: learning from AI summaries left shallower knowledge than web links ^[42]
• Liu et al. (2026), 1,222 people, United States: solved fewer problems and gave up more once the AI was removed ^[41]
• Bastani et al. (2025), about 1,000 people, Turkey: exam scores 17% lower after plain GPT-4 was taken away ^[7]
• Kreijkes et al. (2025), 405 people, United Kingdom: note-taking beat using an LLM alone for recall three days later ^[33]
• Chan et al. (2024), 200 people, online: a misleading chatbot planted three times as many false memories ^[9]
• Barcaui (2025), 120 people, Brazil: 57.5% vs 68.5% correct on a surprise test 45 days later ^[5]
• Zhou et al. (2026), 61 people, China: the creativity boost vanished without ChatGPT while ideas stayed alike ^[55]
• Shen and Tamkin (2026), 52 people, country not stated: learning a new code library with AI left quiz scores 17% lower ^[44]

Did better once the AI was gone

• Kumar et al. (2025), 1,818 people, United States: LLM explanations taught more than correct answers alone ^[35]
• LearnLM Team (2026), 1,763 people, Sierra Leone: Gemini Guided Learning raised maths scores by 0.26 SD ^[37]
• De Simone et al. (2025), 759 people, Nigeria: six weeks of Copilot tutoring raised scores by 0.31 SD ^[14]
• Fischer et al. (2025), 334 people, Germany: an AI tutor raised exam scores by 0.23 SD ^[19]
• Pardos and Bhandari (2024), 274 people, online: ChatGPT-written hints taught as well as tutor-written hints ^[43]
• Contractor and Reyes (2026), 211 people, United States: studying with AI raised unaided test scores by 0.27 SD ^[12]
• Kestin et al. (2025), 194 people, United States: a purpose-built AI tutor doubled learning gains over active learning ^[31]
• Huang et al. (2025), 187 people, China: higher skill-test scores with less mental effort after ChatGPT training ^[25]
• Gan et al. (2024), 110 people, China: higher orthopaedics and end-of-term exam scores after studying with ChatGPT ^[20]
• Kavadella et al. (2024), 77 people, Cyprus: better knowledge-exam results after an assignment done with ChatGPT ^[29]

No clear difference, or both

• Bassner et al. (2026), 275 people, Germany: higher exercise scores with AI, no extra learning ^[6]
• Lehmann et al. (2025), 176 people, Germany; Netherlands: no overall effect on learning; it depended on how students used it ^[40]
• Kazemitabaar et al. (2023), 69 people, Canada: coding with AI helped and did not hurt skills a week later ^[30]
• Dai et al. (2025), 387 people, China: AI feedback helped some students and made others less self-directed ^[13]
• Fan et al. (2025), 117 people, China: better essays with ChatGPT, no more knowledge gained ^[17]
• Hou et al. (2026), 95 people, United States: AI with a prompting guide helped; AI without one did not ^[24]
• Ba et al. (2024), 77 people, China: same theory scores, better clinical skills after ChatGPT-assisted teaching ^[3]

05 What do brain scans show about AI use?

11 studies recorded brain activity while people used generative AI, with EEG (electrical signals from the scalp) or fNIRS (blood-oxygen changes in the outer brain). All are small, from 24 to 95 participants. Several found lower or different engagement with AI, but brain signals are hard to read as better or worse thinking, and only some studies also tested what people could do afterwards.

• Burelli et al. (2026), 60 people, Italy and Denmark: EEG and eye measures showed lower engagement when coding with AI ^[8]
• Clin Deffarges et al. (2026), 50 people, country not stated: an unrestricted chatbot gave the biggest immediate gains, mostly by giving answers ^[11]
• Gao et al. (2026), sample size not stated, China: same benefit as search, but less complex prefrontal activity ^[21]
• Jin et al. (2026), 95 people, China: an AI creativity instructor raised creativity scores, with different brain activity ^[28]
• Shrivastava et al. (2026), 30 people, India: EEG effects of ChatGPT depended on the design phase ^[45]
• Xia et al. (2026), 78 people, China: reasoning improved no differently with AI or search ^[51]
• Zhang et al. (2026), 34 people, China: same retention, weaker transfer after learning through ChatGPT ^[53]
• Baradari et al. (2025), 24 people, United States: a brain-adaptive chatbot raised engagement but not learning ^[4]
• Ge et al. (2025), 32 people, China: LLM partners lowered mental load; human partners gave more original ideas ^[22]
• Kosmyna et al. (2025), 54 people, United States: weakest brain connectivity with ChatGPT; 83% could not quote their own essay ^[32]
• Wang et al. (2025), 64 people, China: more creative designs and higher EEG concentration with AI tools ^[50]

06 What did the MIT and Microsoft studies really find?

MIT, “Your Brain on ChatGPT” (2025). 54 students from five Boston-area universities wrote essays with ChatGPT, with a search engine or with no tool while wearing EEG caps. Brain connectivity was weakest in the ChatGPT group, and 15 of 18 ChatGPT users could not correctly quote their own essay minutes later, against 2 of 18 in each other group. It is a preprint, not peer reviewed, with 18 people per group, and a published commentary questions its EEG analysis ^{[32] [57]}.

Microsoft and Carnegie Mellon (2025). 319 knowledge workers described 936 real tasks they did with AI. They reported using critical thinking on 555 of them (59%); the more they trusted the AI, the less critical thinking they reported, and the more they trusted themselves, the more they reported. It measured what workers said, at one point in time, so it shows a link, not a cause ^[39].

Gerlich (2025). In a survey of 666 people in the UK, heavier AI-tool use went with lower critical-thinking scores (r = -0.68), mostly through offloading thinking to devices. A correction in September 2025 replaced one duplicated table; the author says the conclusions are unaffected ^[23]. Our news explainer on AI and memory covers the older research on offloading.

07 Any study, in one sentence

Pick a study to see what it found, who took part, whether it was peer reviewed and its main limit, in a sentence you can copy. Every study has its own link and a card you can download or share.

08 Every study, in one table

Table 1

Every study in the tracker

Click a column heading to sort. People = participants analysed, or the round figure the paper gives. * = country taken from the authors' universities. Each study links to the paper; the CSV and the study picker above give the full finding, what was measured and the main limits.

Study	Area and design	People	Main finding	Result
Burelli et al. (2026) Italy and Denmark · Preprint	Brain activity Brain-activity study	60	EEG and eye measures showed lower engagement when coding with AI	Worse
Clin Deffarges et al. (2026) Not stated · Preprint	Brain activity Brain-activity study	50	An unrestricted chatbot gave the biggest immediate gains, mostly by giving answers	Mixed
Gao et al. (2026) China* · Peer reviewed	Brain activity Brain-activity study	n/a	Same benefit as search, but less complex prefrontal activity	Mixed
Jin et al. (2026) China* · Peer reviewed	Brain activity Brain-activity study	95	An AI creativity instructor raised creativity scores, with different brain activity	Better
Shrivastava et al. (2026) India* · Peer reviewed	Brain activity Brain-activity study	30	EEG effects of ChatGPT depended on the design phase	Mixed
Xia et al. (2026) China* · Peer reviewed	Brain activity Brain-activity study	78	Reasoning improved no differently with AI or search	No difference
Zhang et al. (2026) China* · Peer reviewed	Brain activity Brain-activity study	34	Same retention, weaker transfer after learning through ChatGPT	Mixed
Baradari et al. (2025) United States* · Peer reviewed	Brain activity Brain-activity study	24	A brain-adaptive chatbot raised engagement but not learning	Mixed
Ge et al. (2025) China* · Peer reviewed	Brain activity Brain-activity study	32	LLM partners lowered mental load; human partners gave more original ideas	Mixed
Kosmyna et al. (2025) United States (Boston area) · Preprint	Brain activity Brain-activity study	54	Weakest brain connectivity with ChatGPT; 83% could not quote their own essay	Worse
Wang et al. (2025) China · Peer reviewed	Brain activity Brain-activity study	64	More creative designs and higher EEG concentration with AI tools	Better
Tanveer (2026) Saudi Arabia · Peer reviewed	Memory Survey	518	More AI use went with better self-rated memory and more laziness	Mixed
Barcaui (2025) Brazil* · Peer reviewed	Memory Experiment, tested later without AI	120	57.5% vs 68.5% correct on a surprise test 45 days later	Worse
Kreijkes et al. (2025) United Kingdom (England) · Peer reviewed	Memory Experiment, tested later without AI	405	Note-taking beat using an LLM alone for recall three days later	Worse
Abbas et al. (2024) Pakistan · Peer reviewed	Memory Survey	494	More ChatGPT use went with self-reported memory problems and lower grades	Worse
Chan et al. (2024) Online panel (Prolific) · Preprint	Memory Experiment, tested later without AI	200	A misleading chatbot planted three times as many false memories	Worse
Bassner et al. (2026) Germany · Peer reviewed	Learning Experiment, tested later without AI	275	Higher exercise scores with AI, no extra learning	No difference
Contractor and Reyes (2026) United States · Working paper	Learning Experiment, tested later without AI	211	Studying with AI raised unaided test scores by 0.27 SD	Better
Hou et al. (2026) United States* · Preprint	Learning Experiment, tested later without AI	95	AI with a prompting guide helped; AI without one did not	Mixed
LearnLM Team (2026) Sierra Leone · Working paper	Learning Experiment, tested later without AI	1,763	Gemini Guided Learning raised maths scores by 0.26 SD	Better
Liu et al. (2026) United States (online panel) · Peer reviewed	Learning Experiment, tested later without AI	1,222	Solved fewer problems and gave up more once the AI was removed	Worse
Shen and Tamkin (2026) Not stated (online developers) · Preprint	Learning Experiment, tested later without AI	52	Learning a new code library with AI left quiz scores 17% lower	Worse
Bastani et al. (2025) Turkey · Peer reviewed	Learning Experiment, tested later without AI	1,000	Exam scores 17% lower after plain GPT-4 was taken away	Worse
Dai et al. (2025) China · Preprint	Learning Experiment, tested later without AI	387	AI feedback helped some students and made others less self-directed	Mixed
De Simone et al. (2025) Nigeria · Working paper	Learning Experiment, tested later without AI	759	Six weeks of Copilot tutoring raised scores by 0.31 SD	Better
Fan et al. (2025) China · Peer reviewed	Learning Experiment, tested later without AI	117	Better essays with ChatGPT, no more knowledge gained	Mixed
Fischer et al. (2025) Germany · Working paper	Learning Experiment, tested later without AI	334	An AI tutor raised exam scores by 0.23 SD	Better
Huang et al. (2025) China · Peer reviewed	Learning Experiment, tested later without AI	187	Higher skill-test scores with less mental effort after ChatGPT training	Better
Iqbal et al. (2025) China · Peer reviewed	Learning Survey	465	More AI use went with higher self-reported achievement	Better
Jia et al. (2025) China · Peer reviewed	Learning Survey	418	More AI dependence went with lower achievement	Worse
Kestin et al. (2025) United States · Peer reviewed	Learning Experiment, tested later without AI	194	A purpose-built AI tutor doubled learning gains over active learning	Better
Kumar et al. (2025) United States (online) · Peer reviewed	Learning Experiment, tested later without AI	1,818	LLM explanations taught more than correct answers alone	Better
Lehmann et al. (2025) Germany (lab experiments); Netherlands (field study) · Preprint	Learning Experiment, tested later without AI	176	No overall effect on learning; it depended on how students used it	No difference
Melumad and Yun (2025) United States (online panels and a university lab) · Peer reviewed	Learning Experiment, tested later without AI	10,462	Learning from AI summaries left shallower knowledge than web links	Worse
Ba et al. (2024) China · Peer reviewed	Learning Experiment, tested later without AI	77	Same theory scores, better clinical skills after ChatGPT-assisted teaching	Mixed
Gan et al. (2024) China · Peer reviewed	Learning Experiment, tested later without AI	110	Higher orthopaedics and end-of-term exam scores after studying with ChatGPT	Better
Kavadella et al. (2024) Cyprus · Peer reviewed	Learning Experiment, tested later without AI	77	Better knowledge-exam results after an assignment done with ChatGPT	Better
Pardos and Bhandari (2024) Online (Mechanical Turk) · Peer reviewed	Learning Experiment, tested later without AI	274	ChatGPT-written hints taught as well as tutor-written hints	Better
Kazemitabaar et al. (2023) Canada · Peer reviewed	Learning Experiment, tested later without AI	69	Coding with AI helped and did not hurt skills a week later	No difference
Dell’Acqua et al. (2026) Not stated (Boston Consulting Group consultants) · Peer reviewed	Critical thinking Experiment, work done with AI	758	GPT-4 made consultants faster and better, but 19% less often right on a task beyond it	Mixed
Kırcaburun (2026) Turkey and online (English-speaking) · Peer reviewed	Critical thinking Survey	659	More AI use went with weaker critical-thinking habits	Worse
Yan et al. (2026) China · Peer reviewed	Critical thinking Survey	421	More AI use went with stronger self-reported critical thinking	Better
Zhi et al. (2026) United States (online panel) · Peer reviewed	Critical thinking Experiment, work done with AI	393	AI from the start helped under time pressure, hurt reasoning when time was ample	Mixed
Chang et al. (2025) Taiwan · Peer reviewed	Critical thinking Survey	413	More ChatGPT use went with higher critical thinking in nurses	Better
Gerlich (2025) United Kingdom · Peer reviewed	Critical thinking Survey	666	Heavier AI-tool use went with lower critical thinking (r = -0.68)	Worse
Kulal (2025) Not stated in the abstract · Peer reviewed	Critical thinking Survey	625	More AI use went with lower critical thinking	Worse
Lee et al. (2025) Several countries (online panel; most often UK, Canada, US) · Peer reviewed	Critical thinking Survey	319	More trust in AI went with less critical thinking at work	Worse
Tian and Zhang (2025) China · Peer reviewed	Critical thinking Survey	580	More AI dependence went with lower critical thinking	Worse
Stadler et al. (2024) Germany · Peer reviewed	Critical thinking Experiment, work done with AI	91	Less mental effort but weaker reasoning than with Google	Worse
Fernandes et al. (2026) United States (online panel) · Peer reviewed	Metacognition Experiment, work done with AI	698	Logic scores rose with ChatGPT, but people overrated how well they did	Mixed
Zhu et al. (2026) China* · Peer reviewed	Offloading Survey	589	Handing thinking to AI went with worse outcomes; using it as a scaffold, better	Mixed
Zhou et al. (2026) China · Peer reviewed	Creativity Experiment, tested later without AI	61	The creativity boost vanished without ChatGPT while ideas stayed alike	Worse
Doshi and Hauser (2024) United Kingdom (online panel) · Peer reviewed	Creativity Experiment, work done with AI	293	AI story ideas made stories more creative and more alike	Mixed
Lee and Chung (2024) Not stated in the abstract · Peer reviewed	Creativity Experiment, work done with AI	n/a	ChatGPT ideas rated more creative; a reanalysis found them less varied	Mixed
Urban et al. (2024) Czech Republic* · Peer reviewed	Creativity Experiment, work done with AI	145	Better, more original solutions with ChatGPT and less mental effort	Better
Agarwal et al. (2025) India and United States (online panel) · Peer reviewed	Writing Experiment, work done with AI	118	AI suggestions pulled Indian writers toward Western styles	Worse

09 What do meta-analyses of AI in classrooms say?

Meta-analyses pool many classroom studies, often small and not randomized, and report larger average gains than the randomized trials that tested students without AI. The spread is wide: the critical-thinking review’s prediction interval runs from harm to large benefit ^[59]. One widely cited review, of 51 studies, was retracted in April 2026 after outside readers raised discrepancies in its analysis ^[62].

Table 2

Meta-analyses: studies pooled together

g = Hedges' g, a standardised difference; 0.2 is small, 0.5 medium and 0.8 large by a common rule of thumb.

Review	What it pooled	Outcome	Result	Status
Sun, Liu, Li and Cai (2026) [59]	26 studies, 47 effects	Critical thinking	g = 0.54 (95% CI 0.31 to 0.78); the 95% prediction interval runs from -0.59 to 1.68	Current
Zhao, Yue, Sun, Jiang and Li (2025) [60]	29 experiments and quasi-experiments	Higher-order thinking	g = 0.61; strongest for problem solving, weakest for creativity	Current
Contractor and Reyes (2026), pooled comparison [12]	13 randomized trials plus their own	Tests taken without AI	average 0.18 SD; losses where the AI did the practice for the learner	Working paper
Wang and Fan (2025) [61]	51 studies	Learning performance	reported g = 0.87; retracted on 22 April 2026 over discrepancies in the meta-analysis	Retracted

10 How we found and checked the studies

1. 01

   Search

   1,409 records from 32 keyword searches of OpenAlex (titles and abstracts, 2023 onwards), 1,287 works cited by 23 recent reviews and key studies, 1,283 records from searches for EEG, fNIRS and fMRI studies of AI use, the 13 trials compared by Contractor and Reyes ^[12], and a news search for 2026 studies.

2. 02

   Rules

   A study is in if people used a generative AI tool; it measured their thinking, learning, memory, critical thinking, creativity, offloading or brain activity; and it was an experiment with at least 50 people, a brain-activity study of any size, or a survey of at least 300. Journals, conference papers, preprints and working papers count.

3. 03

   Out

   Studies of how well AI itself performs, non-generative AI, opinion pieces, interview-only studies, study plans without results, and classroom comparisons that were not randomized. The table below lists 25 studies we read and left out, with the reason.

4. 04

   Reading

   We read each paper’s full text where it was open (42 studies) and otherwise the abstract on the publisher’s page or a research database (14). Every number on this page is quoted from the paper, not from news coverage.

5. 05

   Status

   Crossref for retractions, corrections and errata; arXiv for later versions and journal publication; Crossref for authors, titles and volumes. Two corrections and one corrigendum are noted in the data; one retraction is shown.

6. 06

   Result

   Each study gets one result for its main measure of people’s own thinking: worse, better, no clear difference, or mixed when it was better on one measure or for some people and worse on another.

Studies we read but left out

Table 3

Studies we read but left out, and why

Study	Why it is not in the tracker
Georgiou (2025), arXiv 2507.00181	Experiment with 40 people, below our 50-person rule
Wecks et al. (2024), SSRN 4812513	Observational study of 193 students, below our 300-person rule
Kıyak et al. (2025), Medical Education	Tests a warning label on AI advice, with no comparison without AI
Jain et al. (2026), Cognitive Processing	The abstract does not say people were randomly assigned; full text closed
Essien et al. (2024), Studies in Higher Education	Mixed-methods study of 107 people
Fitriani et al. (2026), AI in Lifelong and Continuing Education	Survey of 93 students
Farhat (2026), Applied Cognitive Psychology	Interviews with 45 users only
Hou et al. (2025), British Journal of Educational Technology	Studies how critical thinking shapes reliance on AI, not the reverse
Jiang and Chen (2026), Computers in Human Behavior: Artificial Humans	Measures what students hand to AI, not the effect on their thinking
Lau et al. (2026), Computers in Human Behavior Reports	Builds a questionnaire
Zhang et al. (2024), IJETHE 21, 34	Lists effects students perceive; no measure of thinking
Davor et al. (2026), JMSTE	The outcome is confidence in problem solving, not thinking
Misiejuk et al. (2026), Computers in Human Behavior Reports	Analyses 281 prompts from 122 submissions
Liu et al. (2026), Interactive Learning Environments	Sample size not given; the outcome is learning accountability
Zhang et al. (2025), Digital Studies in Language and Literature	A study plan; no results yet
Thinking with AI (2026), OSF registration	A registration; no results yet
Juárez et al. (2026), Sustainability	EEG with ChatGPT but no comparison condition
Xia et al. (2026), Interactive Learning Environments	Measures how human the AI seems, not thinking
Gopalla et al. (2025), AHFE International	Compares three chatbots for ease of use, 12 people
Li, Liu and Ye (2025), SSRN 5533921	Randomized, but the sample size is not in the abstract we could read
Wang et al. (2026), International Journal of Industrial Ergonomics	Could not read the abstract or the paper
Lokare et al. (2026), Research Square	Could not read the abstract or the paper
Yurt and Kuşci (2025), Current Psychology	Could not read the abstract or the paper
Yang et al. (2026), International Journal of Information Management	Could not read the abstract or the paper
Deng et al. (2025), Computers & Education (meta-analysis)	Could not read the abstract or the paper

Randomized experiment

People are assigned by chance to use AI or not, so differences afterwards can be put down to the AI.

Preprint

A paper posted publicly before peer review, in which independent experts check the methods.

Cognitive offloading

Handing mental work to something outside your head, such as a notebook, a search engine or a chatbot.

Standard deviation (SD)

A measure of spread; a gain of 0.2 SD is small and 0.5 SD is medium by a common rule of thumb.

EEG and fNIRS

EEG records electrical activity through scalp sensors; fNIRS uses light to track blood oxygen in the outer brain.