Does Relying on AI Weaken Your Critical Thinking? What the Research Shows

Does relying on generative AI weaken your critical thinking? The research that has accumulated since 2025 points to a consistent pattern: the more people offload their thinking to AI tools, the lower they tend to score on measures of critical thinking, and the effect is strongest in the youngest, heaviest users. That does not mean AI makes you permanently less intelligent, and the studies have real limitations. But it does mean the way you use these tools matters far more than whether you use them.

What the new research actually found

A new cross-sectional study published in Cureus, the peer-reviewed medical journal, examined dependence on generative AI among medical students and its association with critical thinking. It joins a fast-growing body of work asking the same uncomfortable question. Medical students make a revealing test population: they are bright, highly motivated, and under enormous time pressure, which is exactly the combination that makes handing your reasoning to a chatbot most tempting. Researchers in medical education have warned that generative AI can shift "from being a support tool to a threat to students' development as independent thinkers," as a 2025 paper in Advances in Medical Education and Practice put it.

The most widely cited study in this area is Michael Gerlich's 2025 paper in the journal Societies, which surveyed and interviewed 666 participants across age groups and education levels. Gerlich found a significant negative correlation between frequent AI tool use and critical thinking ability. The relationship was mediated by a specific mechanism: cognitive offloading, the practice of letting an external tool do mental work you would otherwise do yourself.

A separate survey of 712 high school and early university students, cited in a 2026 systematic review in JMIR Medical Education, reached the same conclusion: higher generative AI usage frequency correlated with lower critical thinking scores, again with cognitive offloading identified as the mediating factor.

Three different populations, three different research teams, one repeated finding. That is not proof of causation, but it is a pattern worth taking seriously.

The MIT experiment: what "cognitive debt" looks like in the brain

Survey studies can only show correlation. The closest thing we have to causal evidence comes from the MIT Media Lab's 2025 study "Your Brain on ChatGPT." Researchers divided 54 participants into three groups and had them write essays across four sessions spanning four months: one group used ChatGPT, one used a traditional search engine, and one worked with nothing but their own brain. Throughout, participants wore EEG headsets measuring neural activity.

The differences were stark. The brain-only group showed the strongest and most widely distributed neural connectivity. Search engine users showed moderate engagement. ChatGPT users showed the weakest connectivity of all, and according to the researchers they "consistently underperformed at neural, linguistic, and behavioral levels." Many could not accurately quote from essays they had submitted minutes earlier.

The researchers coined a memorable phrase for the long-term cost: cognitive debt. Like financial debt, it accumulates quietly. Each individual offload feels free, but the compounding effect is a mind that has practiced less, remembers less, and engages less.

One caveat belongs right next to that finding: the MIT paper was released as a preprint, had not completed peer review at publication, and 54 participants is a small sample. The authors themselves cautioned against overinterpreting it. It is suggestive, not settled science.

Cognitive offloading is old. This version is different.

Humans have always outsourced mental work. Writing offloaded memory. Calculators offloaded arithmetic. GPS offloaded navigation, and studies of London taxi drivers famously showed that people who navigate from memory develop measurably different hippocampal structure than people who follow turn-by-turn directions.

So why worry now? Because generative AI offloads a different category of work. A calculator executes a procedure you already understand. A search engine retrieves raw material you still have to evaluate, compare, and synthesize. A chatbot delivers the finished product: the argument already structured, the evidence already selected, the conclusion already drawn. What gets offloaded is not computation or retrieval but judgment itself, and judgment is precisely the muscle critical thinking depends on.

The search engine group in the MIT study is the telling detail. Searching is also a technology crutch, yet those participants maintained far more neural engagement than the ChatGPT group, because sifting results and constructing your own synthesis is still active mental work.

Who is most at risk

Gerlich's data showed that younger participants exhibited both higher dependence on AI tools and lower critical thinking scores than older participants. That aligns with what educators are seeing: students who reach for a chatbot before attempting a problem never get the productive struggle in which reasoning skills form.

Age is not the real variable, though. Habit is. The at-risk profile is anyone who uses AI as a replacement for thinking rather than a tool for it. A 60-year-old who lets a chatbot draft every email and make every decision is accumulating the same debt as a college student who has it write every paper. The difference between "AI as bicycle" and "AI as wheelchair" is not the tool. It is whether your own faculties stay load-bearing.

The honest caveats

Good critical thinking demands applying it to this research too. Three limits matter. First, correlation runs both ways: people with weaker critical thinking skills may simply lean on AI more, rather than AI causing the weakness. Cross-sectional surveys cannot untangle that. Second, the strongest experimental evidence, the MIT study, is small and was not peer reviewed when released. Third, measuring critical thinking is genuinely hard, and self-reported usage data is noisy.

What tilts the overall picture toward concern is convergence. Correlational surveys, mediation analyses, EEG data, and educator observation all point in the same direction. Waiting for perfect proof before adjusting how you use these tools is itself a gamble, and the downside of the adjustment is close to zero.

How to use AI without dulling your own thinking

Think first, prompt second

Before you ask an AI anything nontrivial, force a first attempt: sketch your own answer, estimate, or outline in two minutes. This preserves the retrieval and reasoning practice that offloading eliminates, and it makes you a sharper judge of what the AI returns, because you have a baseline to compare against.

Use AI as a critic, not an author

The healthiest pattern in the research is inverted use: you produce the draft, the argument, or the decision, and the AI attacks it. Ask it to find the holes in your reasoning, the evidence against your position, the cases where your plan fails. That configuration adds cognitive work instead of removing it.

Verify one claim per session

Generative AI states falsehoods with the same fluency as facts. Making a habit of independently checking at least one substantive claim per session keeps your verification instincts alive, and it regularly turns up errors that would otherwise slide through.

Keep some thinking deliberately manual

Pilots hand-fly takeoffs and landings even though autopilot could manage them, precisely to keep skills from decaying. Adopt the same policy: mental arithmetic on the restaurant bill, navigating without the map app on familiar routes, and regular unassisted problem-solving. Puzzles and reasoning games are a practical version of this. Timothy E. Parker, the Guinness World Records Puzzle Master whose puzzles have been solved by 180 million total solvers across 30 years and 80+ countries, has built a career on a simple premise: minds stay sharp by doing the work, not by watching it get done. If you want a baseline read on your own reasoning, you can take the Real World IQ test and see how you perform when no tool is helping.

Notice when you have stopped deciding

The subtlest failure mode is not wrong answers. It is the quiet migration of decisions you used to make yourself. Once a month, audit where AI sits in your workflow: which judgments have you delegated entirely? Pull one or two back.

The bottom line

The evidence so far says generative AI is neither poison nor vitamin. It is a lever, and levers amplify whatever you do with them. Used as a sparring partner, a critic, and a checker, AI can genuinely sharpen your thinking. Used as a substitute for it, the research suggests you will slowly get worse at exactly the skills that matter most in a world full of confident, fluent, machine-generated claims. The people who thrive alongside these tools will be the ones who can still think without them, and the only way to stay in that group is to keep testing your reasoning against real problems, regularly and without assistance.

FAQ

Does using ChatGPT actually lower your IQ?

No study has shown that AI use lowers IQ scores. What the research shows is a negative association between heavy AI reliance and critical thinking performance, mediated by cognitive offloading, plus reduced neural engagement during AI-assisted work in the MIT EEG study. Skills that go unpracticed decay; that is different from a drop in underlying capacity, and it appears to be reversible through deliberate practice.

Is cognitive offloading always bad?

No. Offloading routine work, like storing phone numbers or automating a calendar, frees attention for harder problems and is generally beneficial. The concern in the recent research is offloading judgment itself: letting a tool evaluate evidence, structure arguments, and draw conclusions for you. That is the specific kind of offloading that correlated with lower critical thinking scores.

Are young people really affected more?

In Gerlich's 666-participant study, younger participants showed both higher AI dependence and lower critical thinking scores than older ones. Researchers believe developing thinkers are more vulnerable because they are still building the skills being offloaded, which is why the effect worries educators most at the high school, university, and professional-school level.

How much AI use is too much?

The research measured frequency of reliance, not hours, so there is no established safe dose. The pattern across studies suggests the risk lives in how you use it: reaching for AI before attempting your own answer, and accepting its output without verification. If you think first, use AI to challenge your work, and routinely check its claims, heavy use looks much less risky than light-but-passive use.