AI affects critical thinking through cognitive offloading: when a tool does the reasoning, you stop practicing it. Three 2025 studies point the same way. A survey of 666 people found that heavier AI use went with lower critical thinking scores, explained by offloading. A survey of 319 knowledge workers found that the more people trusted the AI, the less critical thinking they did, while self-confidence went with more. A randomized trial with nearly 1,000 high school students isolated the cause: an AI that gave answers raised practice scores but cut exam scores by 17% once it was gone, while an AI told to give hints instead showed no drop. The link between AI and critical thinking therefore depends on which way the questions flow. If you ask and it answers, you offload the thinking. If it asks and you answer, or you write your own answer before asking, the reasoning stays yours.
On this page+
The link between AI and critical thinking comes down to who does the reasoning. AI weakens critical thinking when it reasons for you and can strengthen it when it makes you reason yourself. The research so far points at one variable: which way the questions flow. If you ask and it answers, you offload the thinking. If it asks and you answer, you keep it.
You are probably here because you use AI most days and something feels slightly off. Emails come out faster, research takes minutes instead of hours, and yet you are less sure you could explain your own conclusions without the tab open. That suspicion is reasonable, and it is also only half the story. This piece covers what the studies actually found, why an answer on tap costs more than it seems to, and a simple test for whether your own use is building the skill or replacing it.
How does AI affect critical thinking?
AI affects critical thinking mainly through cognitive offloading: handing a mental task to a tool so you do not have to do it yourself. No single AI critical thinking study settles the question, but three from 2025 give the clearest picture so far.
In the first, Michael Gerlich surveyed and interviewed 666 people in the UK1. People who used AI tools more often scored lower on critical thinking, and the link ran through offloading: the more they handed over, the lower they scored. Younger participants leaned on AI most and scored lowest, while more education went with stronger critical thinking whatever people's AI habits were. In the second, a team at Microsoft Research and Carnegie Mellon asked 319 knowledge workers about 936 real tasks. The more people trusted the AI, the less critical thinking they reported doing. The more they trusted themselves, the more they did.
Both of those are surveys, so neither proves cause. The third is a randomized trial with nearly 1,000 high school students, and it does3. Students given an AI that handed out answers did better during practice and 17% worse on the exam once the tool was gone. Students given an AI told to offer hints instead of answers showed no such drop.
That last result matters most for AI and critical thinking, because it isolates the variable. Same technology, same students, same material. The only difference was whether the AI did the reasoning or asked the student to do it. For the school-side evidence in more detail, see what the research says about AI in classrooms.
Why does a ready answer weaken the thinking it replaces?
A ready answer weakens your thinking because it skips the part where the skill gets used. Critical thinking is not a fact you hold; it is a set of moves you make: noticing an assumption, weighing one reason against another, asking what would change your mind. An answer arrives with those moves already made, out of sight, and all that is left for you is to accept it or not.
The Microsoft Research study found that AI does not remove critical thinking from work so much as change its job2. The effort moves toward checking information, fitting the output into your own work, and overseeing the result. That is still thinking. The catch is the confidence finding: the people most sure the AI was right were the least likely to do the checking. So overreliance on AI grows quietly. The better the tool gets, the more reasonable it feels to skip the one job it left you.
There is a second, subtler cost. Checking facts is not the same as checking reasoning. You can confirm every statistic in an AI's answer and still have adopted its framing of the problem, its choice of which options exist, and its sense of what matters. The negative impact of AI on critical thinking is rarely a wrong fact. It is a borrowed structure you never noticed you were standing on.
We see this in sessions often. Someone arrives with a pros and cons list a chatbot wrote for them, sometimes a very good one. Asked which item on it they actually believe, they go quiet. The list is sound; it is just not theirs, and a decision built on it tends to wobble the first time someone pushes back.
Can AI improve critical thinking?
Yes, when it asks rather than answers. The same trial that found the 17% exam drop also found the fix3. The tutor version of the AI was told to give hints, not solutions, and to respond to common mistakes. Students using it improved far more during practice than students using the answering version, and they did not lose ground on the exam. The researchers found that students used the answering version "as a crutch", asking for solutions and copying them, while they used the tutor to ask for help or try an answer themselves.
The same pattern shows up for college students, where the use that helps during term is the one that fails on exam day. And it matches what Gerlich told Big Think when his study came out: in his view, the right use of AI can increase critical thinking skills, if it serves critical discussion rather than replacing your own work.
The honest limit: the causal evidence so far is one trial, in school maths. The AI impact on critical thinking in adult decisions has only been measured through surveys. The direction is consistent across all three studies, though, and it lines up with a much older finding about learning: you keep what you work out, not what you are told.
Side by side, the two ways of combining AI and critical thinking look like this:
AI answers, you accept
AI asks, you answer
Who does the reasoning
The tool
You
What you practice
Prompting and skimming
Weighing, testing, deciding
What you can defend later
The answer, if you remember it
The reasons, because you built them
When the tool is gone
Performance drops
Performance holds
Where the effort goes
Checking output (often skipped)
Thinking the problem through
Which way do your questions flow?
The quickest way to check your own habit is to look at who asked the last real question. Scroll back through your recent AI conversations and mark each exchange. If you asked and it answered, the questions flowed toward you as finished answers. If it asked you something and you had to think before replying, they flowed the other way. Most people find the first kind outnumbers the second by a wide margin. That is not a moral failing; the tools are built to answer, and answering is what we ask for.
For AI and critical thinking, it does mean the ratio is the thing to change, not the hours. Heavy use where you are mostly doing the reasoning is closer to practice than light use where you mostly accept. Critical thinking and artificial intelligence are not opposites; an answer on tap simply leaves nothing for the thinking to do.
An answer you did not build is one you cannot check. The thinking has to come from your side of the screen.
That guess-first habit is built into Inwitt AI, an AI thinking partner app that asks rather than answers. When you want facts during a session, the findings stay locked until you have made your own guess, and then you choose which results are relevant, with the results kept neutral rather than framed as advice. The point is the same as the exercise: your reasoning goes first, so the information has something to test.
What does critical thinking look like when you use AI?
It looks like doing the reasoning yourself and using the tool to test it, not to produce it. The Microsoft Research team's three shifted jobs (checking, integrating, overseeing) are real, but they only protect the skill if they involve your own judgement. Four habits turn them into practice:
Answer first. Write your view before you ask, as in the exercise above. It turns the AI's reply into a second opinion rather than the first one.
Ask for the objection, not the answer. "What is the strongest case against this?" keeps the conclusion yours and puts the tool to work on stress-testing it.
Name what would change your mind. Before you read the reply, decide what kind of evidence would move you. Otherwise every answer either confirms you or overrules you.
Keep the decision in your words. If the final reasoning is a paste, you have a document, not a conclusion. Rewrite it in a sentence you would say out loud. Seeing it drawn out, as on a thinking board built from your own words, makes gaps obvious.
These track the parts of the critical thinking process that most people skip: stating the question precisely, and deciding in advance what would count as an answer. An AI can do the middle steps quickly. The first and last ones only work if they come from you.
Common mistakes with AI and critical thinking
Treating fact-checking as thinking. Verifying the numbers in an answer is useful and still leaves the framing untouched. Ask what the answer assumed about the problem, not only whether its facts are right.
Asking it to "be critical" and adopting the critique wholesale. A generated critique is just another answer. If you swap your view for its objection without weighing either, you have outsourced the judgement twice. Our reasoning is better at defending conclusions than testing them, which is why reasoning critically takes deliberate effort; the effort has to be yours.
Quitting AI entirely. Stopping is not the same as thinking. The trial evidence says the design and direction of use matter, not abstinence, and a tool that asks can be real practice.
Better prompts instead of more of your own sentences. Prompt tricks improve the output. They do nothing for your reasoning unless you write something of your own before and after.
Measuring dependence in hours. The question is not how much you use it but who reasons when you do. An hour of being questioned is different from five minutes of accepting.
Questions people ask about AI and critical thinking
Will AI replace critical thinking?
No, but it can replace the practice that keeps it sharp. Someone still has to judge whether an answer fits the situation, and that judgement is critical thinking. If you stop exercising it, what AI replaces is your share of the work, not the need for it.
Does using AI make you lazy?
It makes the easy path easier, which is not quite the same thing. The surveys show people who trust AI more put in less critical effort; the effort is still available when you choose to spend it. Answering first before you ask is the cheapest way to keep spending it.
Is AI bad for your brain?
There is no good evidence that AI damages the brain. What the studies measure is habit: people who hand over more reasoning score lower on reasoning tests. That is closer to an unused skill than an injury, and practice is the remedy.
Why does critical thinking matter more with AI?
Because the answers are fluent whether they are right or wrong, and fluency reads as confidence. The more polished the output, the more your own judgement is the only thing standing between a plausible answer and a good one. The methods for improving critical thinking that hold up in research apply here too: real decisions, on paper, with questions.
The pattern across the research on AI and critical thinking is plain enough: when the questions flow toward you, the thinking stays yours. The exercise above does that for one decision, with a pen. A session does it at the pace of a conversation: one question at a time, your answers on the thinking board, and a receipt at the end that holds your reasoning in your own words. If you would rather be asked than answered, join the Inwitt waitlist.
Footnotes
Gerlich, "AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking", Societies 15(1), 6 (2025). Mixed-method survey and interviews, 666 participants; correlational. A 2025 correction replaced a duplicated table; the author states the conclusions are unaffected. Journal record↩
Lee, Sarkar, Tankelevitch, Drosos, Rintel, Banks and Wilson, "The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers", CHI 2025 (Microsoft Research and Carnegie Mellon University). Paper page↩↩2
Bastani, Bastani, Sungu, Ge, Kabakcı and Mariman, "Generative AI without guardrails can harm learning: Evidence from high school mathematics", Proceedings of the National Academy of Sciences 122(26) (2025). Randomized trial, nearly 1,000 students in grades 9 to 11. Journal record↩↩2
