Your Brain on ChatGPT: Is AI Making Us Think Less?
Apparently, ChatGPT is making our brains lazy. At least that is the impression you might get from the headlines that followed a recent MIT Media Lab study, usually illustrated with two brains side by side: one busily thinking, the other apparently on vacation. The images are persuasive. But is that what actually happened? It is worth taking a closer look.
The researchers asked people to write essays under three conditions: on their own, with a search engine, or with ChatGPT [1]. Fifty-four participants took part over three sessions, while the researchers recorded the brain’s electrical activity with EEG and examined the essays themselves, how well participants could recall them, and how much ownership they felt over what they had written. In a fourth session, completed by 18 of them, some participants switched conditions. The differences they observed are genuinely interesting. Those who wrote unaided showed the strongest and most widely distributed brain connectivity and ChatGPT users the weakest, and the ChatGPT group also had more trouble quoting their own essays and felt less ownership of them.
Somewhere between the laboratory and social media, however, these observations became explanations, and the explanations became warnings about what AI is supposedly doing to our brains. The study was released as a preprint, based on a small group of participants from one area and a single kind of task, and its authors themselves asked for its conclusions to be treated with caution and as preliminary.
Our neurology does not have an opinion
Our neurology provides the biological infrastructure within which cognition operates, but it does not pass judgment on the quality of our thoughts. More electrical activity does not necessarily mean better thinking, just as less activity does not automatically signal intellectual decline. To read a finding like this one, we need to keep at least three levels apart.
The biological level asks what actually happened in the nervous system: which signals were recorded, and under which conditions. The psychological level asks what the person was experiencing, understanding, remembering or believing about their own performance. The relational level asks how the interaction between the person, the task and the AI shaped what happened. These levels interact, but they are not interchangeable, and a finding at one of them does not automatically answer a question at another.
When does thinking actually begin?
There is another problem. Imagine someone who has been turning an idea over for weeks: reading, observing, struggling with contradictions, gradually building an understanding. One morning they open ChatGPT, exchange a few sentences, and write something remarkable in ten minutes. Was the thinking done in those ten minutes?
Now imagine someone else asking ChatGPT to write an essay on a subject they barely understand, then copying the result, changing a few words and submitting it. Both used AI and both produced an essay, but it is hard to believe they went through the same cognitive process, and we could not establish the difference by recording their brain activity during the writing alone. Preparation, execution and what remains afterward may involve very different processes.
The study’s own fourth session points in this direction. Participants who had first written on their own and then switched to ChatGPT showed better recall and more activity in several brain regions, while those who had used ChatGPT from the start and then had to write alone showed signs of under-engagement [1]. With 18 participants it is a small signal, but it suggests that what people bring to the interaction matters, not only whether AI is present.
This is why the researchers’ concept of cognitive debt deserves investigation rather than immediate acceptance as an established consequence. Sometimes AI substitutes for cognitive work. Sometimes it removes effort that was never useful. And sometimes it helps us reach an understanding that would otherwise take far more time or resources. These are not equivalent outcomes, and a measure that cannot tell them apart will struggle to say which one is happening.
At Reflectum we are exploring precisely these distinctions: how trust, uncertainty, cognitive preparation and the structure of human–AI interaction shape the development of understanding. Our interest is not simply whether someone performs better with AI, but whether the interaction changes what they can afterwards understand, evaluate and accomplish on their own.
The MIT study offers useful observations. It does not establish that ChatGPT damages our brains, or that reduced neural connectivity necessarily represents cognitive decline. Perhaps the most useful scientific question is not whether AI makes us think less, but whether we are learning to think differently, and whether we understand what we are gaining, what we are delegating, and what might be lost along the way. That is a conversation worth having.
One Claim. A Closer Look. is a Reflectum series that takes a single widely shared claim about AI and the mind and examines what the evidence behind it actually shows. Editorial review and source checking by Claude.
Reference
[1] Kosmyna, N., et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv preprint.
https://doi.org/10.48550/arXiv.2506.08872