AI and critical thinking

Used to skip the thinking, AI weakens critical thinking. Used to pressure-test it, AI can strengthen it. That distinction is the whole answer. The evidence does not show that AI lowers intelligence. It shows that when people lean on a model to reason for them, they do the reasoning less often, trust the output more, and lose the habit of questioning it. Critical thinking does not vanish; it migrates, from producing an answer to verifying the machine's, and it thins as confidence in the tool rises. Whether your critical thinking declines is therefore not decided by ChatGPT. It is decided by whether you think before you ask, and whether you interrogate the reasoning rather than just accept the result. Most people do neither, because the tool makes it so easy not to.

What the workplace studies found#

The clearest workplace evidence comes from a 2025 study by Microsoft Research and Carnegie Mellon, which asked 319 knowledge workers about 936 real uses of AI in their jobs. It found that the more people trusted the tool, the less critical thinking they applied, and that the character of the thinking changes: from gathering information to verifying the machine's output, from solving the problem to integrating the answer, from doing the task to supervising it. The workers who kept thinking critically were those with the confidence and skill to inspect and correct the AI. The rest tended to accept what they were given.

A separate 2025 study by Michael Gerlich, across 666 people, found a significant negative correlation between frequent AI use and critical-thinking scores, with cognitive offloading as the mechanism in between, and the effect strongest among the youngest participants, who have leaned on these tools earliest and hardest. A 2025 experiment at the MIT Media Lab used EEG to compare people writing essays with a language model, with a search engine, or unaided, and found the AI group showed the weakest brain connectivity and the lowest sense of ownership over their own work, an effect the authors called cognitive debt.

None of this is new in kind. The mechanism underneath it, cognitive offloading, was mapped by Risko and Gilbert in 2016: we hand mental work to external tools to reduce effort, and, crucially, we decide to offload based on how hard a task feels, a judgement that is often wrong. And the precedent is older still. Sparrow and colleagues, in Science in 2011, showed the Google effect: when we expect information to stay available, we remember where to find it rather than the thing itself. What was true of facts is now becoming true of reasoning. The tool that holds the answer gradually holds the thinking too.

Read carefully, none of this says AI makes people stupid. Critical thinking is a practice, and AI is very good at letting us skip the practice while still producing the output. Skip it often enough and the capability follows the effort out of the room.

Where the evidence remains uncertain#

The direction of the risk is well supported; the size is not, and it matters to say so plainly. The Microsoft and Carnegie Mellon findings are self-reported: they capture how workers describe their own thinking, not a measured before-and-after. Gerlich's result is a correlation, which cannot on its own separate whether AI use erodes critical thinking or whether people who think differently simply use AI differently. The MIT study is striking but rests on 54 participants, is a preprint, and has been questioned by its own later commentators on sample size and reproducibility. Treat it as suggestive, not settled.

There is also a real counter-current. Used deliberately, AI can raise the quality of thinking: challenging a position, surfacing a counterargument, exposing a gap in an analysis. The same tool can be a crutch or a sparring partner. Almost no study yet measures the long-run difference between the two modes of use in real workplaces, and that is the question that matters most. The responsible conclusion is that the risk is real enough, and quiet enough, to be worth designing against now, rather than waiting for a decade of proof.

Two cases that show what verification failure looks like#

In January 2025 the High Court in Pietermaritzburg, South Africa, dealt with counsel who had cited authorities that did not exist. What makes Mavundla worth remembering is the judge's own account: Bezuidenhout J tested one of the citations by asking ChatGPT, and the tool falsely confirmed the case was real. The verification method was the same class of system that had produced the error. Costs were awarded against the attorneys personally and the matter referred to the Legal Practice Council.

In March 2026 the publisher Mediahuis suspended Peter Vandermeersch, former editor-in-chief of NRC and chief executive of Mediahuis Ireland, after NRC found fabricated quotations in fifteen of fifty-three of his blog posts. Eight articles were withdrawn. He acknowledged using several AI tools. A newspaper caught its own former editor, which is both reassuring and not, and the organisation's response was to reaffirm its rules on AI use rather than restrict the tools.

These are worth holding onto because they are not stories about bad people or bad technology. In both cases capable, senior professionals produced material they had not sufficiently checked, using a tool that made the unchecked version look finished. That is the thinning described on this page, in two documented instances, with names and dates attached.

There is a third case that reframes all of it. In Ayinde v Haringey, decided by the Divisional Court in England in June 2025, the court found the threshold for contempt met over fabricated citations but declined to prosecute, partly because of "potential failings on the part of those who had responsibility for training Ms Forey, for supervising her, for 'signing off' her pupillage." A court looked at an individual AI failure and saw a supervision failure. That is the missing-rungs argument arriving from the bench.

Relocated, not destroyed#

Critical thinking is being relocated by AI, not destroyed, and most people have not noticed the move. The work of thinking is shifting from generating an answer to checking one, and checking is the harder discipline. It requires you to hold an independent view against which to test the machine, and that view is what you lose if you ask the machine first and think second. The danger is that AI removes the moment where you would have thought for yourself, and does it so smoothly that nothing feels lost. Thinking for you would at least be noticeable.

This is why I put a single practical principle at the centre of it: on any judgement-heavy task, think first, then consult. Form an initial position before you open the model, so you keep an independent reference point to evaluate its answer against. It is a small habit with a large effect, because it preserves the one thing the evidence says is at risk: your capacity to notice when a confident output is wrong. The failure mode is outsourcing the first move, the framing of the problem, which is where judgement actually lives. Using AI is not the variable.

The distinction between a crutch and an amplifier is what SuperSkills calls the Augmented Mindset: working with AI so it extends your capability rather than replacing it. A person with a strong augmented mindset uses AI to attack their own reasoning, not to avoid it, and can always say where the tool helped, where it misled, and where their own judgement overrode it. That is critical thinking with AI in the room, rather than critical thinking handed to it. And it connects directly to the wider account of judgement and to the missed reps: every time the tool does the reasoning, the person misses the repetition that would have built the capability to do it themselves.

What to do about it#

For yourself: think before you ask. On anything that requires judgement, write your own position first, even a rough one, then use AI to test and extend it rather than to produce it. Interrogate the reasoning, not just the result, because a fluent answer can be confidently wrong and the fluency is what disarms you. Work unaided sometimes, deliberately, to keep the capability exercised, the way a musician still practises scales. And treat cognitive convenience as a cost as well as a benefit: the easier it is to skip the thinking, the more it is worth asking whether you should.

For managers and teams: make reasoning visible. Ask people to show their thinking, not only their output, because output quality has stopped being a reliable signal of whether the person can think. Keep some work AI-free on purpose, especially for people still building their judgement. Reward the person who can explain and defend a recommendation over the one who simply produced a polished one. And treat verification as real, skilled work rather than a rubber stamp, because in an AI-assisted team the checking is the thinking.

The market already values this. The World Economic Forum's 2025 Future of Jobs report names analytical thinking as the single most sought-after core skill among employers. As AI makes fluent output cheap, the ability to question it becomes the scarce and valuable thing. Critical thinking becomes more valuable and more fragile at the same time, which is close to the opposite of obsolete.

Key research and primary sources

Go to the study rather than the article reporting it.

• Mavundla v MEC: Department of Co-operative Government and Traditional Affairs KwaZulu-Natal [2025] ZAKZPHC 2. High Court of South Africa, KwaZulu-Natal Division, Pietermaritzburg.
• Ayinde v London Borough of Haringey; Al-Haroun v Qatar National Bank [2025]. Divisional Court, England and Wales.
• The Irish Times (2026). Mediahuis suspends top journalist after admission of using false AI material.
• Lee, H.-P. et al. (2025). The Impact of Generative AI on Critical Thinking. Microsoft Research and Carnegie Mellon, CHI 2025.
• Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), 6. See also the correction published in September 2025.
• Kosmyna, N. et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt. MIT Media Lab preprint.
• Risko, E. F. and Gilbert, S. J. (2016). Cognitive Offloading. Trends in Cognitive Sciences, 20(9).
• Sparrow, B., Liu, J. and Wegner, D. M. (2011). Google Effects on Memory. Science, 333(6043).
• World Economic Forum (2025). The Future of Jobs Report 2025.

Related SuperSkills research#

This page sits within a wider body of work: AI and human judgement, the Augmented Mindset, the missed reps, decision quality in the AI era and drift versus design. In his own words, see the Box of Amazing essay Are You Flying or Are You Being Flown?, where Rahim Hirji set out algorithmic drift and the atrophy of skill under autopilot. See also using AI without dependency and how humans learn with AI. The mechanism is defined at cognitive offloading. The graded evidence, including what each study does not support, is in the evidence base. On the collective effect, does AI make everyone think alike? The claims themselves, banded by evidence strength and including what remains unknown, are in what we actually know about AI and human capability.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. This work draws on research across more than 200 organisations in 30 countries over seven years. Findings on this page are attributed to the studies that produced them, and kept separate from the interpretation, which is the author's. It is a living reference, reviewed and updated as significant new evidence appears.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.