Start hereIn one experiment, high-school students who practised maths with a plain GPT-4 tool got lower grades on their own afterwards. Yet in a physics course, a custom AI tutor let students learn significantly more in less time. So does AI help or hurt thinking?
The verdicts
What critical thinking means
As an educational goal, the idea goes back to the American philosopher John Dewey (1910), who more commonly called it reflective thinking. He defined it as active, persistent and careful thought about a belief, in the light of the grounds for it and the conclusions it leads to.
Between 1988 and 1989 a panel of 46 scholars, educators and others agreed a longer definition, published in 1990. It was brought together at the request of a committee of the American Philosophical Association. They described it as purposeful, self-regulating judgment that interprets, analyses, evaluates and infers, and explains the reasons it rests on.
The definitions do not fully agree. Some limit critical thinking to forming a judgment. Others, as the Stanford Encyclopedia of Philosophy notes, count actions as well as beliefs as the end point.
Our reading: the studies people quote do not all measure the same thing. One of the AI studies below defines critical thinking through Bloom’s taxonomy of learning goals, and its authors say this definition is not uncontested.
Can it be taught?
On average, a little. A 2008 review by Philip Abrami and colleagues pooled 117 studies with 20,698 participants. It found an average effect of 0.341 standard deviations, with results that varied a great deal. The authors concluded that improvement cannot simply be expected to happen. Teachers must make critical-thinking goals explicit in their courses.
A 2015 review by the same group drew on 341 effect sizes from experimental and quasi-experimental studies. It found a similar average of 0.30, and effective strategies at all levels of education and in every subject area. Three stood out: dialogue, real-world problems and examples, and mentoring.
Our reading: an effect of about a third of a standard deviation is a real but modest average gain, not a transformation.
The same review found that teaching critical thinking on its own and inside subject lessons together did better than either alone. But the Stanford Encyclopedia notes that the difference was not statistically significant, so it might have arisen by chance. Most of the studies also lacked long-term follow-up.
The harder question is whether the skill travels from one subject to another. The cognitive scientist Daniel Willingham argued in 2007 that critical thinking is not a set of skills you can use anywhere. In his view, it depends heavily on knowing the subject and practising in it.
He did not say it never carries over: “such transfer does occur”, he wrote, but when and why is complex.
The Stanford Encyclopedia describes the evidence for one popular answer, teaching critical thinking across many subjects with explicit attention to the skills they share, as “scanty”.
Our reading: teaching helps on average, and knowing a subject well is what lets the habit work in that subject.
Try it: when you want to judge claims in a field, learn its basics first; the habit works best on things you know.
↑ All verdictsWhat employers say
The line “critical thinking is the No. 1 skill employers want” is usually credited to the World Economic Forum’s Future of Jobs reports. The Forum’s own summary of its 2020 report said critical thinking and problem-solving were at the top of the list. These are surveys of what employers say they expect, not measurements of what workers can do.
-
In the October 2020 report, critical thinking and analysis comes first among the skill groups employers saw rising in importance. In the same report’s list of the top 15 individual skills for 2025, it ranked fourth, behind analytical thinking and innovation, active learning, and complex problem-solving.
We searched the text of the 2023 and 2025 reports and did not find the phrase critical thinking in either. Our reading: the claim fits one 2020 chart of skill groups; in the ranked lists of individual skills, a close cousin, analytical thinking, comes first.
A second statistic often travels with these reports. It says “65% of children entering primary school today will ultimately end up working in completely new job types that don’t yet exist.” The Forum’s 2016 report gives it as “one popular estimate”, and its endnote gives a web address for Shift Happens, not a study.
We could not find a study behind it.
↑ All verdictsAI and thinking: what the studies found
A survey, a preprint and an experiment answer different questions; our guide to reading a study explains why.
A survey of knowledge workers
The paper, presented at the CHI 2025 conference, found that higher confidence in generative AI went with less critical thinking. Higher self-confidence went with more. AI also shifted critical thinking toward checking information, combining responses and overseeing the task.
The paper’s own title calls the reductions in effort self-reported. Its authors note that participants sometimes mixed up less effort in using AI with less effort in thinking critically with it. They also say the sample leaned towards younger, more technologically skilled people. It shows an association in what people reported, not a cause.
A survey of 666 people
Critical thinking was measured partly by self-report. The author lists reliance on self-reported measures and possible sample bias among the limits, and says experiments could offer causal evidence. A correction in September 2025 replaced a duplicated table; the author states that the scientific conclusions are unaffected.
The brain-scan preprint
Many news stories about ChatGPT and the brain go back to a preprint from the MIT Media Lab, first posted in June 2025. In all, 54 people took part in the first three sessions, and 18 completed a fourth. They wrote essays with ChatGPT, with a search engine or with no tools while their brain activity was recorded by EEG.
The authors report that the ChatGPT group showed the weakest brain connectivity. As of 13 October 2026 its arXiv record shows a preprint, last revised on 31 December 2025, with no journal publication listed.
Other researchers have posted a constructive comment on its small sample, reproducibility and EEG methods. The authors’ own FAQ says it does not show that LLMs make us dumber, and asks people not to use words like “brain rot”.
The strongest experiment
Our reading: the clearest causal evidence among these studies comes from a field experiment with nearly 1,000 students at a high school in Turkey, published in PNAS in 2025. While students had GPT-4 during practice, their grades rose: by 48% with the plain tool and 127% with the tutor version.
When access was taken away, students who had used the plain tool got grades 17% lower than those who never had access. The tutor version, whose instructions asked it to guide students step by step instead of giving the answer, largely avoided the drop.
The authors’ explanation is that, without guardrails, students used GPT-4 “as a ‘crutch’ during practice problem sessions”. The study measured maths learning over a few sessions, not critical thinking in general, and the authors say they focused on short-term outcomes.
Try it: on practice work, make your own attempt first, then ask the AI for hints or explanations rather than answers.
Experiments that point the other way
When the tool is built for learning, results can be positive.
Two 2026 studies add detail, one a preprint and one a CHI 2026 conference paper.
In the conference paper, an experiment with 393 people, having an AI model from the start helped under time pressure but hurt performance when there was enough time. Starting the task without it showed the opposite pattern.
What 15-year-olds told PISA
Among frequent users, some used AI to help them learn and were taught to judge the quality of its output. They tended to do better than those who had no such guidance. These are associations, not proof that AI lowered anyone’s scores. The full results of PISA’s new Learning in the Digital World test are due in May 2027.
Our reading: surveys find that people who lean on AI more report, or score, less critical thinking, but they cannot say which causes which. The clearest experiment found that using AI to skip the work during practice left students worse off on their own. A tool built to make students do the thinking largely avoided that.
Over-trusting machines is an old problem
Researchers have a name for leaning too hard on a confident machine: automation bias, “the tendency to over-rely on automation”. A 2012 systematic review screened 13,821 papers and kept 74. It found that trust and confidence, workload, task complexity and time pressure all played a part, and that training and stressing that users are accountable helped.
In a 2023 experiment, 27 radiologists read mammograms with suggestions they were told came from an AI system. In 12 of 40 cases the suggestion was deliberately wrong. Inexperienced readers rated 79.7% of mammograms correctly when the suggestion was right and 19.8% when it was wrong. Very experienced readers fell from 82.3% to 45.5%.
The authors concluded that radiologists at every level of experience are prone to automation bias when an AI system supports them. Experience softened the pull without removing it. Why AI makes things up covers the errors chatbots make.
Try it: before you rely on an AI answer, check one of its claims against a named source.
Checking what you read: what helps, and how much
The first approach is lateral reading: leaving a page to see what other sources say about it. It is the core of the SIFT method.
As our SIFT page explains, a lasting benefit measured against a comparison group has not been established.
The second is prebunking, or inoculation: short games or videos that show people a manipulation technique before they meet it. In a 2022 study in Science Advances, five short videos improved recognition of techniques such as false dichotomies across six randomised experiments with 6,464 people.
In a field test on YouTube the effect was smaller. Recognition rose by about 5% on average, measured by a single test question within 24 hours of the advert.
Whether these lessons help people tell reliable from unreliable news, and not just name a trick, is disputed. A 2023 reanalysis of five studies found that two inoculation games “did not improve discrimination” but made people answer false more often to everything.
Two of that reanalysis’s five authors also co-wrote the 2022 YouTube study. The 2023 critique was written by two researchers who were not among the authors of either. Our logical fallacies page weighs the two and leans, cautiously, toward the larger reanalysis, because it covers far more experiments.
Try it: when a page makes a claim, open a new tab and see what others say about who is behind it. In the Bad News study, the effect lasted when people were tested at regular intervals, so it may help to repeat the habit.
↑ All verdictsQuick answers
What is critical thinking?
The philosopher Robert Ennis puts it as reasonable and reflective thinking focused on deciding what to believe or do. The idea goes back to John Dewey, who more commonly called it reflective thinking.
Can critical thinking be taught?
Yes, modestly. Two reviews of classroom studies (2008 and 2015) found an average gain of about a third of a standard deviation, and how far the skill carries from one subject to another is still argued over.
Is critical thinking the top skill employers want?
Partly. In the World Economic Forum's 2020 report, critical thinking and analysis led the skill groups but ranked fourth among individual skills. In the 2023 and 2025 reports the top core skill is analytical thinking.
Does ChatGPT make you dumber?
The research does not show that. Two surveys show associations, not causes, and the brain-scan study is a small preprint that has not been peer reviewed. In one maths experiment, students who used a plain GPT-4 tool as a crutch did worse on their own later.
Has the MIT "Your Brain on ChatGPT" study been peer reviewed?
Not as far as we can find. As of 13 October 2026 its arXiv record shows a preprint, last revised on 31 December 2025, with no journal publication listed.
Use this
Use AI after you have done some thinking, not instead of it.
- On practice work, make your own attempt first, then ask the AI for hints rather than answers.
- Before you rely on an AI answer, check one of its claims against a named source.
- When a page makes a claim, open a new tab and see what others say about who is behind it.
Did I do the thinking here, or did I hand it to the tool?
Quick quiz
Nothing is scored or saved. Recalling it once now is what makes it last.
1. In the World Economic Forum's 2020 report, where did critical thinking and analysis rank among individual skills?
Answer: Fourth.Read it again ↑
2. What is the MIT "Your Brain on ChatGPT" study?
Answer: A preprint with 54 participants.Read it again ↑
3. In the maths experiment, which version of GPT-4 largely avoided the drop in grades?
Answer: A tutor that guided rather than gave answers.Read it again ↑
How we checked this · Last checked October 2026
- What we read in full: the Stanford Encyclopedia entry, the Delphi Report, Ennis's outline, Willingham's article, the four World Economic Forum reports (searched as text), the Lee, Gerlich, Bastani, Kestin and Roozenbeek papers, and the Gerlich correction.
- We also read the OECD's PISA 2025 press release, the abstracts of the other papers and preprints, and the Kosmyna authors' FAQ.
- We could not read the full texts of the Abrami and Huber & Kuncel reviews (their effect sizes come from the abstracts), the full PISA 2025 report, the full texts of the two 2026 studies, or the BBC programme on the 65% figure.
- On 13 October 2026 we re-checked the MIT preprint: its arXiv record still shows version 2 of 31 December 2025 and no journal reference, and Crossref and OpenAlex searches found no journal version.
- The same day we found that the 2026 learning preprint had been revised (version 2, 4 October 2026); the page now follows its current abstract.
Sources
- Stanford Encyclopedia of Philosophy, “Critical Thinking” (David Hitchcock), §§1, 11 and 12.Scholarly
- P. A. Facione, Critical Thinking: A Statement of Expert Consensus for Purposes of Educational Assessment and Instruction (“The Delphi Report”), 1990, ERIC ED315423.Scholarly
- R. H. Ennis, “The Nature of Critical Thinking”, University of Illinois, revised May 2011.Scholarly
- P. C. Abrami et al., “Instructional Interventions Affecting Critical Thinking Skills and Dispositions”, Review of Educational Research 78(4), 2008.Scholarly
- P. C. Abrami et al., “Strategies for Teaching Students to Think Critically”, Review of Educational Research 85(2), 2015.Scholarly
- C. R. Huber and N. R. Kuncel, “Does College Teach Critical Thinking? A Meta-Analysis”, Review of Educational Research 86(2), 2016.Scholarly
- D. T. Willingham, “Critical Thinking: Why Is It So Hard to Teach?”, American Educator, Summer 2007.Scholarly
- World Economic Forum, The Future of Jobs (2016), The Future of Jobs Report (2020, 2023, 2025).Primary
- C. N. Davidson, “How can you prepare for a career in an era of automation? BBC interview”, 1 June 2017.Primary
- H.-P. Lee et al., “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers”, CHI ‘25.Scholarly
- M. Gerlich, “AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking”, Societies 15(1), 2025, and correction, Societies 15(9), 2025.Scholarly
- N. Kosmyna et al., “Your Brain on ChatGPT”, arXiv:2506.08872 (preprint), and the authors’ FAQ.Scholarly
- M. Stanković et al., “Comment on: Your Brain on ChatGPT”, arXiv:2601.00856 (preprint).Scholarly
- H. Bastani et al., “Generative AI without guardrails can harm learning: Evidence from high school mathematics”, PNAS 122(26), 2025.Scholarly
- G. Kestin et al., “AI tutoring outperforms in-class active learning”, Scientific Reports 15, 2025.Scholarly
- Z. Contractor and G. Reyes, “Experimental Evidence on the Learning Impact of Generative AI”, arXiv:2607.08849 (preprint, 2026; revised October 2026).Scholarly
- J. Zhi, H. Kumar and M. Lee, “Investigating the Effects of LLM Use on Critical Thinking Under Time Constraints: Access Timing and Time Availability”, Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (April 2026); also arXiv:2603.08849.Scholarly
- OECD, “PISA 2025: Students’ reading and mathematics performance declined sharply across the OECD”, press release, 8 September 2026.Primary
- K. Goddard, A. Roudsari and J. C. Wyatt, “Automation bias: a systematic review of frequency, effect mediators, and mitigators”, JAMIA 19(1), 2012.Scholarly
- T. Dratsch et al., “Automation Bias in Mammography”, Radiology 307(4), 2023.Scholarly
- S. Wineburg et al., “Lateral reading on the open Internet: A district-wide field study in high school government classes”, Journal of Educational Psychology 114(5), 2022.Scholarly
- J. Roozenbeek et al., “Psychological inoculation improves resilience against misinformation on social media”, Science Advances 8(34), 2022.Scholarly
- R. Maertens et al., “Long-term effectiveness of inoculation against misinformation”, Journal of Experimental Psychology: Applied 27(1), 2021.Scholarly
- A. Modirrousta-Galian and P. A. Higham, “Gamified inoculation interventions do not improve discrimination between true and fake news”, Journal of Experimental Psychology: General, 2023.Scholarly
- A. Simchon et al., “A Signal Detection Theory Meta-Analysis of Psychological Inoculation Against Misinformation”, Current Opinion in Psychology 67, 2026.Scholarly
- S. Y. N. Wang et al., “Limited effectiveness of psychological inoculation against misinformation in a social media feed”, PNAS Nexus 4(6), 2025.Scholarly