MIT Found That 83% of ChatGPT Users Couldn't Remember What They'd Just Written. Then the Researcher Had to Clarify What the Study Did Not Show. Both Facts Matter.

In June 2025, MIT Media Lab researcher Nataliya Kosmyna and her team published a pre-print titled 'Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.' They recruited 54 participants — mostly college students in Boston — and divided them into three groups: one wrote essays using ChatGPT, one used Google Search, one wrote with no tools at all. Each wore an EEG headset measuring brain activity across 32 neural regions. Each essay was scored by human teachers and an AI judge. The results: ChatGPT users showed the weakest neural connectivity, the lowest cognitive engagement, and the worst memory recall. When asked to reproduce what they had written, 83% of ChatGPT users could not. The brain-only group had the strongest neural activity and the best recall. The search engine group sat in the middle. The headlines wrote themselves: 'AI Is Making Us Dumber.' 'ChatGPT Is Rotting Your Brain.' 'MIT Proves AI Destroys Critical Thinking.' Then Kosmyna herself had to step in and clarify that the study showed nothing of the sort.

What the Study Actually Found — and What It Did Not

Kosmyna publicly stated that her study did not find that AI made participants 'stupid,' 'dumb,' or caused permanent damage to critical thinking or intelligence. What it found was more specific: when people outsource a cognitive task entirely to an AI, they engage less cognitively with it, and they retain less of what the AI produced. This is not surprising. If you hand a task to someone else entirely, you put less mental effort into it. What the study does not — and cannot — tell you is whether AI use in general reduces cognitive ability over time. That is a different question, and the study did not test it. AI researcher Ethan Mollick, one of the most respected voices on AI in education, described the mainstream coverage as a 'massive misinterpretation,' noting that 'It says something important about cheating with AI, but it doesn't tell us anything about LLM use making us dumber overall.'

The Studies That Surround It: A Fairer Picture

The Three Gaps That Explain Everything

Marketing AI Institute's Mike Kaput articulated the framework that makes sense of all of this research. There are three gaps created by AI use that compound if you are not deliberate about addressing them. The Verification Gap: AI generates output, but you must fact-check it. Most people skip verification unless they already suspect an error. The errors you do not suspect are the costly ones. The Thinking Gap: evaluating AI output critically requires cognitive capacity that is not unlimited. Under time pressure, the temptation is to accept what the AI produced rather than engage with it. The Confidence Gap: when you present AI-generated material without deeply engaging with it, you develop a false confidence that collapses the moment someone asks you to explain it without the AI.

These gaps do not mean AI is making you dumber. They mean passive AI use — without deliberate engagement — creates conditions for skill atrophy. The distinction is yours to make every time you open a prompt box. The tool does not make the choice. You do.

What Dario Amodei Does — And Why It Matters

Anthropic CEO Dario Amodei, who leads the company that makes Claude, described a personal rule that captures the entire debate in one observation. He still types out his own notes in every meeting rather than using an AI notetaker. 'If I just have the notetaker, there's less cognitive load,' he said. 'But that cognitive load is actually what embeds it in my memory networks.' The person who runs one of the world's most powerful AI companies understands that the cognitive friction of note-taking is not a bug — it is the mechanism of memory formation. He chooses to preserve that friction deliberately. That is the right frame for every knowledge worker making decisions about AI use.

The Practical Framework: How to Use AI Without Paying the Cognitive Price

• Engage first, AI second — every time: The MIT study's most important buried finding is that people who wrote independently first and then used ChatGPT produced the best creative work — stronger arguments, higher originality, better recall. Spend 5-10 minutes engaging with any task before asking AI to assist. Write your initial thoughts, identify the questions you're trying to answer, form your first hypothesis. The AI then extends your thinking rather than replacing it.
• Treat verification as non-negotiable, not optional: The Microsoft/Carnegie Mellon finding that high AI trust correlates with reduced critical evaluation is the most actionable piece of research in this entire space. Make a rule: for any AI-generated content that will be shared, published, or acted on, you verify the key claims, numbers, and logic before using it. Not because you distrust the AI — because verification is the cognitive engagement that keeps you from being fooled by a plausible-sounding error.
• Know which tasks are worth outsourcing and which are not: There is a meaningful difference between outsourcing the formatting of a bibliography (low cognitive value, appropriate to outsource) and outsourcing the argument structure of an analysis (high cognitive value, the part you need to engage with). Map your work to these categories. Use AI heavily on the low-value-to-your-development tasks. Protect your cognitive engagement on the high-value ones.
• The student test: The AI math tutoring study's 17% exam performance drop is the clearest warning for learners. If you use AI to complete assignments, you will score well on those assignments. You will perform worse when AI is unavailable — exams, job interviews, live client situations. For anything you are trying to learn, not just complete, the correct rule is: understand it yourself first, use AI to check and extend, never to replace the initial understanding.
• Periodic AI-free zones: The researchers found that LLM-to-brain participants showed weakened performance even after switching back to independent work — suggesting dependency can develop. Build regular practices of doing cognitively demanding work without AI assistance. Writing, analysis, problem-solving, planning — not as productivity, but as maintenance of the cognitive capacity that AI use gradually erodes if never exercised independently.

The Longer View: What History Actually Suggests

Calculator use became widespread in schools in the 1970s and 1980s. Arithmetic fluency declined. Computational speed — the ability to do mental multiplication quickly — atrophied in most people who went through calculator-enabled education. This is measurable and real. And it turned out that fast mental multiplication was less important to the overall trajectory of human intellectual capability than the educators who worried about it believed. The more interesting question about AI is the same one: what does the tool make possible that was not possible before, and is the gain worth the cognitive dependency it creates? The research so far is early, limited in scope, and frequently misrepresented. The MIT study is one specific finding about one specific task with 54 participants. What we can say accurately in April 2026 is this: the concern is legitimate, the 'AI is rotting your brain' headline is wrong, and deliberately cultivating active rather than passive AI use is both the prudent response to the concern and the correct way to get the most from the technology.