Is AI Making Us Dumber? What the Research Says

Is AI making us dumber? In 2025, an MIT study called "Your Brain on ChatGPT" and a Microsoft survey on critical thinking both went viral, and plenty of headlines said yes. The research says something narrower. None of it shows AI lowering intelligence. But leaning on it without questioning it can reduce how much you think, remember and catch.

We read six of the most cited studies closely and compared them with data from 2,047 AISA AI fluency assessments. The common thread: the risk comes less from using AI than from no longer questioning it.

The short version:

• MIT Media Lab (2025): 83.3% of ChatGPT users couldn't correctly quote the essay they had just written. Without ChatGPT, 11.1% couldn't.
• Microsoft and Carnegie Mellon (2025): the more 319 knowledge workers trusted AI with a task, the less critical thinking they reported.
• Gerlich (2025): among 666 people, heavier AI use went with weaker critical thinking.
• Radiology (2023): when the AI was wrong, very experienced radiologists' accuracy fell from 82.3% to 45.5%.
• Harvard and BCG (2023): on a task the AI got wrong, consultants using it were 19 percentage points less likely to be right.
• The Lancet Gastroenterology & Hepatology (2025): doctors' detection rate without AI fell from 28.4% to 22.4% after AI arrived.
• AISA (October 2026): people score 47.2 out of 100 on using AI at work and 42.4 on questioning its output. Critical thinking trails in every role group we track.

Is AI Making Us Dumber? The Short Answer

None of the most cited studies shows that AI makes people less intelligent. They show something more specific: relying on AI without checking its work reduces mental effort, memory of the work and the ability to catch errors. Researchers call the underlying habit cognitive offloading. Whether it helps or hurts depends on how critically you use the tool.

So does AI make you dumber? Not in the way the headlines suggest. In MIT's 2025 study, people writing with ChatGPT were less engaged and remembered less of their own essays. In a Microsoft survey of 319 knowledge workers, those who trusted AI more reported less critical thinking. Radiologists and consultants given confidently wrong AI suggestions got more answers wrong, and doctors who had used AI found fewer precancerous growths once it was off. Each effect concerns engagement, checking or practice. None concerns intelligence.

The headlines went further than the studies

In early 2025, 404 Media ran the headline "Microsoft Study Finds AI Makes Human Cognition 'Atrophied and Unprepared'." Business Standard went with "Microsoft admits that AI is making us dumb."

The words "atrophied and unprepared" do appear in the Microsoft paper, but in its introduction. They describe Lisanne Bainbridge's 1983 "ironies of automation": take routine work away from people and they lose the practice they need for when the machine fails. It is not a finding of the study.

The MIT team's public FAQ goes further. It asks journalists not to describe the results with words like "stupid," "dumb" or "brain rot."

What is cognitive offloading?

Cognitive offloading means handing a mental task to a tool: a calculator for arithmetic, a map app for directions, a chatbot for a first draft. It isn't new, and it isn't bad in itself.

The real questions are what happens to the skills you stop using, and to your judgement about work you no longer produce yourself. When offloading turns into accepting whatever the AI says, AISApedia calls it cognitive surrender.

What Did MIT's "Your Brain on ChatGPT" Study Find?

MIT Media Lab researchers recorded the brain activity of 54 people writing essays with ChatGPT, with a search engine, or with no tools. The ChatGPT group showed the weakest brain connectivity and felt the least ownership of their essays. After the first session, 83.3% of them could not correctly quote the essay they had just written.

How the experiment worked

Kosmyna and colleagues split 54 participants from nearby universities into three groups of 18. Each group wrote essays over three sessions with one tool: ChatGPT, a search engine, or nothing. In a fourth session, 18 people from the ChatGPT and no-tools groups swapped tools. EEG sensors on the scalp recorded brain activity throughout.

What they found

• Engagement fell as outside help rose. Brain connectivity "systematically scaled down with the amount of external support," the authors write. It was strongest with no tools and weakest with ChatGPT.
• Memory of the work fell. In session one, 15 of 18 ChatGPT users (83.3%) could not correctly quote their own essay, against 2 of 18 (11.1%) in each of the other groups.
• Ownership fell. ChatGPT users felt the least ownership of what they wrote, and the no-tools group the most.

The most useful finding: order matters

In the fourth session, people who had written on their own and then used ChatGPT showed higher memory recall and wider brain connectivity than the ChatGPT group had shown in its own sessions. The authors call this one of their most striking observations. Using AI after your own thinking looked very different from using it instead of thinking.

What is cognitive debt?

Cognitive debt is the term in the paper's title. The authors define it as a condition in which "repeated reliance on external systems like LLMs replaces the effortful cognitive processes required for independent thinking." It "defers mental effort in the short term," they argue, but brings long-term costs such as "diminished critical inquiry" and "decreased creativity."

What the study does not show

It is a preprint still under peer review, with 54 people, one task and one tool, limits the authors list themselves. It measured engagement, memory and ownership during writing. It did not measure intelligence.

AI and Critical Thinking: What Two 2025 Studies Found

Two 2025 studies link heavy reliance on AI with less critical thinking. A Microsoft and Carnegie Mellon survey found that confidence decides how much people check: trusting AI with a task went with less critical thinking, and trusting your own ability went with more. A 666-person study by Michael Gerlich tied the effect to cognitive offloading.

Microsoft and Carnegie Mellon: confidence decides how much you check

Lee and colleagues surveyed 319 knowledge workers who use AI at least weekly and gathered 936 real examples from their jobs. The more confident people were that AI could handle a task, the less critical thinking they reported. The more confident they were in their own ability, the more they reported.

AI made most kinds of thinking easier, in 69% to 79% of examples depending on the activity. Evaluating the result was the exception, at 55%. AI eased the step where you judge its work least often.

Time pressure was one reason people skipped the check. As one sales representative put it: "I use AI to save time and don't have much room to ponder over the result." The data is self-reported, and the authors say it does not prove cause and effect.

Gerlich: more AI use, more offloading, less critical thinking

Michael Gerlich of SBS Swiss Business School surveyed 666 people and interviewed some of them in depth. He found a significant negative correlation between frequent AI use and critical thinking, with cognitive offloading as the link between them. Younger participants relied on AI more and scored lower on critical thinking. Higher education went with better critical thinking whatever people's AI use.

The design is correlational. AI use may weaken critical thinking, or people who think less critically may lean on AI more.

Does Relying on AI Make Experts Worse at Their Jobs?

It can, when the AI is wrong and the expert goes along with it. Automation bias is the tendency to accept a machine's suggestion over your own judgement. In two experiments, one with radiologists and one with BCG consultants, accuracy fell sharply when people worked with an AI that was confidently wrong.

Radiologists: experience helped, but not enough

In a 2023 Radiology study, Dratsch and colleagues had 27 radiologists read 50 mammograms with suggestions from what was presented as an AI system. The researchers controlled the suggestions and made 12 of them wrong.

Experience softened the fall, from about 60 points to about 37, but nobody was immune. Every reader knew how to read a mammogram. What separated the groups was how much weight they gave a wrong suggestion.

Consultants: the jagged frontier

In a field experiment with 758 Boston Consulting Group consultants, Dell'Acqua, Mollick and colleagues found that GPT-4 made people faster and better on tasks it could handle: 12.2% more tasks, 25.1% faster and 40% higher quality. Then they set a business problem where the AI would give a convincing but wrong answer. Without AI, 84% of consultants got it right. With AI, 60 to 70% did, according to co-author Ethan Mollick.

The authors call the edge of AI's ability a jagged frontier. Tasks that look equally hard can sit on opposite sides of it, and nothing in the AI's tone tells you which.

Can Using AI Cause Deskilling?

Early evidence says it can. After AI polyp detection arrived at four Polish endoscopy centres, the rate at which doctors found precancerous growths in colonoscopies done without AI fell from 28.4% to 22.4%. The finding comes from a 2025 study in The Lancet Gastroenterology & Hepatology.

What was measured

Budzyń and colleagues compared 795 colonoscopies done without AI in the three months before the tool arrived with 648 in the three months after. The measure, the adenoma detection rate, is the share of colonoscopies that find at least one precancerous growth. The drop held after accounting for patients' age and sex.

What it does not show

It is an observational study in one specialty, not a randomised trial, and the authors say only that AI exposure "might reduce" detection when the AI is off. It is still a measured example of Bainbridge's irony of automation.

Is AI Making Us Dumber? 6 Studies Compared

Taken together, the six studies show less engagement, weaker checking and some skill loss when people lean on AI without questioning it. None of them shows a loss of intelligence, and each one has limits worth knowing before you quote it.

AISA Data: People Use AI Better Than They Question It

Across 2,047 completed AISA assessments (data as of October 2026), people score an average of 47.2 out of 100 on using AI in their work and 42.4 on critically evaluating what it produces. Critical thinking trails AI use in every role group we track. People also overestimate their AI skill: predictions average 62.8, and results average 43.7.

AISA is a conversational AI fluency assessment. Instead of answering a quiz, people talk through how they actually use AI, and they are scored on what they demonstrate against a published rubric. That lets us measure two things separately: how well someone puts AI to work, and how well they question what it gives them. On average, questioning trails using by 4.8 points, from 0.8 points for students to 10.1 for designers.

The gap holds in every role

Designers show the widest gap. Students show the narrowest, with both scores low. Product managers score highest on using AI and still trail by 7.8 points on questioning it, a pattern we explore in PMs and AI: Strong Vision, Weak Verification.

Confidence runs ahead of skill

Across 1,137 assessments where people predicted their score first, the average prediction was 62.8 and the average result was 43.7: an overestimate of 19.1 points. Students overestimated by 31.5 points, engineers by 14.9 and founders by 8.2. We break down where it concentrates in AI Skills Gap: You Overestimate by 18.5 Points.

Lee and colleagues warned that confidence in your own ability "may not always be well-calibrated" with real expertise. AISA's data puts a number on it. Using AI is the stronger skill, questioning it is the weaker one, and people rate themselves well above where they land. That is the setup the six studies warn about.

How to Use AI Without Getting Dumber

Use AI after you have done some thinking, not instead of it. Check its work in proportion to the stakes. Learn where it fails in your field, and keep doing some work without it. Each habit comes from one of the studies above.

Think first, then ask

MIT participants who wrote on their own before using ChatGPT showed better recall and stronger engagement than the group that started with it. Before you prompt, write down your own answer, outline or hunch, even a rough one. In the radiology study the suggestion came first, and it pulled even experts off course.

Check in proportion to the stakes, not the deadline

People skip the check when they trust the tool or are short on time. Decide in advance what always gets verified: numbers, names, citations, legal or medical claims, anything a customer will see. A short verification checklist makes checking a routine rather than a mood.

Learn where AI fails in your field

AI's wrong answers can look exactly like its right ones. Keep a running list of where it has let you down: the invented figure, the misread rule, the outdated detail. Within a few months that list is your own map of the jagged frontier, and knowing the usual patterns of hallucination tells you where to look first.

Keep some work unassisted

Skills used only alongside AI can fade. Pick the parts of your job where your own judgement must stay sharp, and keep doing some of that work without help. Real human in the loop review means engaging with the work, not signing it off.

How to Measure Your Own AI Critical Thinking

Self-assessment won't tell you, because people overestimate their AI skill by about 19 points on average. AISA's free AI fluency assessment measures it directly. It is a 20 to 40 minute conversation about how you really use AI, scored against a published rubric, with a full report and certificate at no cost.

Your report puts your score for using AI next to your score for questioning it, so you can see whether your scrutiny has kept pace with your use. Results hold up on retake: 98% of people who take it again land in the same or an adjacent tier, and scores correlate at r = 0.79 between attempts. Population results are published quarterly in the AISA AI Fluency Index.

Organisations use the same team AI assessment to see how well their people question AI before relying on it more.

---

Related reading: Critical Thinking in AI: Operators vs Passengers. What separates people who check AI output from people who pass it along. Related reading: AI Skills Gap: You Overestimate by 18.5 Points. The skill areas where self-assessment runs furthest ahead of reality. Related reading: AI Certification for Teachers & Educators. What AI fluency looks like in classrooms, where the MIT findings hit closest.

Frequently Asked Questions

Does AI make you lazy?

It makes skipping effort easy, and people do skip it when they trust the tool. A 2025 Microsoft and Carnegie Mellon survey of 319 knowledge workers found that more confidence in AI went with less critical thinking. Effort drops when people stop questioning the tool, not simply because they use it.

Does using AI reduce brain activity?

In one study, brain engagement was lower while people used it. In MIT's 2025 "Your Brain on ChatGPT" study, people writing with ChatGPT showed weaker brain connectivity than people using a search engine or no tools. Those who wrote unaided first and then used ChatGPT showed stronger engagement. The study is a preprint and did not measure intelligence.

What is cognitive debt?

Cognitive debt is a term from MIT's 2025 "Your Brain on ChatGPT" study. It describes repeated reliance on AI replacing the effortful thinking needed for independent work. You save effort now, and the authors argue you pay later in weaker critical inquiry and creativity.

Can AI make you smarter?

It can make your work better when you use it well. In the Harvard and BCG study, consultants using GPT-4 on suitable tasks produced 40% higher quality work, but did worse on a task the AI got wrong. Better results depend on knowing when to trust AI, and that judgement is a skill you can measure and improve.