TL;DR: In 2025 and 2026 a wave of studies asked whether leaning on AI weakens the mind, and the honest answer is more useful than either headline. The MIT Media Lab essay study found that people who wrote with ChatGPT showed the weakest brain connectivity of any group and that most could not quote a sentence they had just produced. A Microsoft and Carnegie Mellon survey of 319 knowledge workers found that the more people trusted the AI, the less critical thinking they applied, and that self-reported effort dropped across nearly every reasoning task. A separate study of 666 people found a strong negative link between frequent AI use and critical-thinking scores, mediated by offloading and sharpest in younger users. But none of these prove AI makes you permanently dumber, and all measure short-term or self-reported effects. The real finding is conditional: offloading a skill you already own is leverage, and offloading a skill you were meant to build is decay. This piece gives you the research table, an editorial Cognitive Offloading Ledger, and a two-question test to run before each delegation. Decide what to hand over like a CEO, and keep the reps that matter like a student.
Every few months a study lands with a headline saying AI is making us stupider, and every few months a rebuttal says the study was tiny, unreviewed, or misread. Both reactions miss the point. The question that matters to anyone building a career in the assistant era is not “is AI bad for the brain” in the abstract. It is narrower and far more practical: which specific mental steps can you safely hand to a machine, and which ones, if you hand them over, quietly stop working. That is a decision you make dozens of times a day, usually without noticing. This guide is built to help you make it on purpose.
Cognitive offloading is not new and it is not inherently harmful. Writing a phone number down instead of memorizing it is offloading. So is using a calculator, a checklist, or a shared calendar. Cognitive scientists have studied the trade-off for years: offloading reduces the internal effort a task demands, which is exactly why it is useful and exactly why it carries a cost when the internal effort was the point. What changed in 2025 is that the tool got good enough to absorb the entire task, not just the storage step. The research wave is really an attempt to measure what happens when the thinking itself, and not just the remembering, gets delegated.
What the three studies actually found
Three studies anchor the current conversation. It is worth separating what each one measured from what the headlines claimed, because the gap is where most of the confusion lives.
The MIT Media Lab study, led by Nataliya Kosmyna and colleagues and released as a preprint in 2025, put 54 participants into three groups to write essays: one using ChatGPT, one using a search engine, and one using nothing but their own memory. Using EEG to track brain activity, the researchers reported that the brain-only group showed the strongest and most distributed neural connectivity, the search group sat in the middle, and the ChatGPT group showed the weakest. When asked to quote a line from an essay they had written minutes earlier, roughly 83 percent of the ChatGPT users could not do it. The team described this as accumulating “cognitive debt.” The critical caveat, which the authors themselves stress, is that this is a small, unreviewed preprint measuring a single narrow task, and lower brain activity is not the same as lasting damage.
The Microsoft and Carnegie Mellon study, presented at the 2025 CHI conference, took a different route: it surveyed 319 knowledge workers about 936 real tasks they had done with generative AI. Its central finding was a relationship, not a verdict. The more confidence a person had in the AI’s ability, the less critical thinking they reported applying; the more confidence they had in their own ability, the more they applied. Workers also reported that AI reduced the effort they spent on the classic reasoning activities, and the survey noted that the work did not disappear so much as shift, moving from producing answers toward verifying, integrating, and stewarding them.
The third study, published by Michael Gerlich in the journal Societies in early 2025, surveyed 666 people across age and education groups and combined the numbers with 50 interviews. It found a strong negative correlation between frequent AI-tool use and critical-thinking scores, and it identified cognitive offloading as the mechanism linking the two. The effect was most pronounced among younger participants, who used the tools more and scored lower. As a correlational study, it cannot prove the direction of cause, and the author is explicit about that limit.
| Study | Sample and method | Core finding | What it does not show |
|---|---|---|---|
| MIT Media Lab (Kosmyna et al., 2025 preprint) | 54 people, EEG, essay writing across three tool conditions | Weakest brain connectivity in AI writers; about 83% could not quote their own essay | Small, unreviewed, single task; not proof of lasting harm |
| Microsoft and Carnegie Mellon (CHI 2025) | 319 knowledge workers, 936 self-reported tasks | Higher trust in AI linked to less critical thinking; effort shifts to verification | Self-reported, correlational; no direct skill test |
| Gerlich (Societies, 2025) | 666 people, survey plus 50 interviews | Strong negative link between AI-use frequency and critical thinking, via offloading | Cannot prove direction of cause; younger users overrepresented in effect |
Table: CEOtudent synthesis of three 2025 studies from their published methods and results. Sample sizes and figures reflect what the authors reported.
Read together, the studies do not say AI rots your brain. They say something more precise and more actionable: when you let the tool do the reasoning, the measurable signs of you doing the reasoning go down, and the more you trust the tool the more of the reasoning you hand over. Whether that is a problem depends entirely on whether the reasoning was one you needed to keep.
The distinction that resolves the panic
The single most useful idea for a working professional is that offloading is not one thing. There is a decisive difference between offloading a skill you already own and offloading a skill you are supposed to be building.
A senior engineer who lets an AI draft boilerplate they could write in their sleep is buying time to spend on architecture. The reps they are skipping are reps they no longer need. A first-year student who lets the same tool write every essay is skipping the exact reps that would have built the ability to structure an argument. Same tool, same action, opposite consequence. The Microsoft survey’s finding that expert effort shifts toward verification is the healthy version: the human keeps the judgment and delegates the production. The MIT result, where writers could not recall their own words, is the unhealthy version: the human delegated the part that was supposed to become memory and skill.
This maps directly onto the CEO-and-student frame this publication keeps returning to. A CEO delegates work they have already mastered so they can operate at a higher altitude; they do not delegate the judgment that makes them a CEO. A student is in the phase where the doing is the point, so delegating the doing defeats the purpose. Most of us are both at once, depending on the task. The skill is knowing which hat you are wearing at the moment you reach for the tool. This is the same muscle behind deliberate practice in the age of AI, where the goal is to keep struggling with the parts that build capability while offloading the parts that do not.
The Cognitive Offloading Ledger
To make the distinction usable, here is an editorial framework that sorts a typical professional’s AI use into three columns. It is a judgment tool, not a measurement: the ratings are our synthesis of the research logic above, meant to prompt a decision rather than to report a study.
| Mental step you might offload | Keep it yourself when… | Safe to delegate when… | Danger sign |
|---|---|---|---|
| Recalling facts and figures | You are learning a field and need a mental map | You have the map and just need a lookup | You can no longer reason about a topic without the tool open |
| Structuring an argument | You are still learning to think in that domain | The structure is routine and you will edit heavily | You accept the AI’s outline without being able to defend it |
| First-draft writing | Writing is how you discover what you think | The format is templated and low-stakes | You cannot summarize your own document from memory |
| Analysis and interpretation | The judgment is the core of your value | The analysis is mechanical and you verify it | You ship conclusions you did not check |
| Decisions and trade-offs | The choice carries real consequences | The option space is trivial | You have stopped forming your own view before asking |
Table: CEOtudent editorial framework. Columns describe decision heuristics, not measured outcomes.
The pattern across every row is the same. Offloading is safe when you could still do the task without the tool and are choosing not to for speed. It becomes decay when the tool has become load-bearing for something you never actually learned. The danger-sign column is the early-warning system: each one is a moment where verification quietly stopped, which is the exact behavior the Microsoft study linked to over-trust.
The two-question test
You do not need to consult a table every time you open a chat window. You need two questions, asked in the half-second before you paste a prompt.
First: am I offloading the storage or the thinking? Handing over where a fact lives is cheap and almost always fine. Handing over the act of reasoning is the move that carries a cost, and it is worth a beat of attention. Second: is this a skill I want to own, or a task I want gone? If you would be happy never getting better at this, delegate freely. If getting better at it is part of who you are trying to become, do the rep yourself and use the AI to critique your attempt rather than replace it.
That second mode, using the tool as a sparring partner instead of a ghostwriter, is where the research and the practical advice converge. Ask the AI to find the weak point in your argument after you have made it. Ask it to grade your analysis, not to produce it. This keeps the human doing the load-bearing cognition while still capturing the speed, and it is the same principle behind treating AI evaluation as a skill in its own right, covered in the evaluation skill of judging AI output. The related idea of what happens to the brain when the tool does the thinking is explored further in our piece on cognitive offloading and what research says about your brain, and the broader question of which mental capacities compound over a career runs through the meta-skills that make every other skill easier to learn.
What this means for the next few years
The studies will keep coming, and they will keep getting better designed. The current batch is early, mostly correlational, and in one case built on 54 people and an EEG cap. It would be a mistake to treat any of them as settled science, and an equal mistake to wave them away because they are imperfect. The direction of the signal is consistent across three independent methods, and it matches decades of prior work on how skills fade when the reps stop. That is enough to act on, even before the definitive study exists.
The professionals who will do well are not the ones who refuse the tools out of fear, nor the ones who hand over everything for speed. They are the ones who make offloading a decision instead of a default: automate the mastered work on purpose, protect the reps that build judgment, and keep verifying rather than trusting. AI is not making careful people worse at thinking. It is making the careless gap widen faster. Which side of that gap you land on is, as it has always been, a choice you make one task at a time.
Frequently asked questions
Does using AI permanently damage your brain?
No study has shown permanent damage. The MIT work found weaker brain activity during and shortly after AI-assisted writing, but that is a measure of engagement in the moment, not proof of lasting harm. The stronger and better-supported claim is that skills you stop practicing tend to fade, which is true of any skill and any tool.
Is all cognitive offloading bad?
No. Offloading is how humans have always extended their minds, from writing to calculators to search engines. It becomes a problem only when you offload a skill you actually needed to build or keep. Offloading a mastered task is leverage; offloading a formative one is decay.
How do I use AI without losing my thinking edge?
Do the reps that matter yourself and use AI to critique rather than replace them. Ask it to find flaws in your reasoning after you reason, grade your draft after you write it, and challenge your decision after you decide. Keep verifying its output instead of trusting it by default, which is the single behavior the research links most closely to weaker critical thinking.
Which age group is most affected?
The Gerlich study found the strongest negative effect among younger participants, who used the tools most heavily and scored lowest on critical thinking. This is consistent with the idea that offloading is riskiest during the years when core reasoning skills are still forming.
Are these studies reliable?
They are early and imperfect. The MIT study is a small, unreviewed preprint; the Microsoft and Gerlich studies rely partly on self-report and cannot prove cause and effect. Their value is that three different methods point the same way, which raises confidence in the direction even while the exact size of the effect remains open.
Sources
- Kosmyna, N. et al. Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab, 2025 preprint.
- Lee, H. P. et al. The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. Microsoft Research and Carnegie Mellon University, Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems.
- Gerlich, M. AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 2025, Volume 15, Issue 1.
- Risko, E. F. and Gilbert, S. J. Cognitive Offloading. Trends in Cognitive Sciences, 2016.
- Sparrow, B., Liu, J. and Wegner, D. M. Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips. Science, 2011.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
