In June 2025, a team at the MIT Media Lab posted a preprint called “Your Brain on ChatGPT.” They recorded brain activity from 54 adults writing short essays, some with ChatGPT, some with Google, and some with no tools at all [1]. A week later, TIME’s headline read “ChatGPT May Be Eroding Critical Thinking Skills, According to a New MIT Study.”
I am an AI researcher. Part of my work is building large language model tools that pull structured information out of medical records, and deep-learning tools that analyze medical images, so I use and build these systems in my own work.
My answer to the title question is: not on this evidence. When you let AI write the text for you, you remember less of it and feel it is less yours. That is a real cost if the point of writing was to learn something. But it looks like cognitive offloading on one task, the same trade we make whenever a tool does part of the work for us. It is not evidence of brain damage.
What the MIT team did
Participants, all aged 18 to 39 and all from universities around Boston, were randomly assigned to one of three groups: ChatGPT only, Google search with no AI, or no tools [1]. Each wrote an essay on an SAT-style prompt in 20 minutes, in each of three sessions, while wearing an electrode cap that recorded brain activity (electroencephalography, or EEG) [1]. Right after each essay, a researcher asked them to quote a sentence from it [1].
Two things keep me cautious. This is a small study: 18 people per group, and only 18 participants in total came back for an optional fourth session [1]. And it had not been peer reviewed. The authors explain in their project FAQ that they released the preprint early to collect feedback faster, with peer review only just started.
The EEG part got the headlines. Brain connectivity, a measure of how strongly activity at different spots on the scalp is coupled, was strongest with no tools, weaker with search, and weakest with ChatGPT [1]. That describes how brains were engaged during a 20-minute task with a tool in hand. It is not a scan of brain structure, and it says nothing about intelligence.
The recall result is the convincing part
The finding I take seriously is much simpler. In the first session, 15 of the 18 ChatGPT users could not quote anything from the essay they had finished minutes earlier [1]. When the researchers checked accuracy, none of the 18 quoted their own essay correctly [1]. In the other two groups, most people could.
Imagine two students writing the same essay. One drafts it herself; the other pastes in a chatbot’s version and tidies it up. An hour later, ask each of them to quote a line. The first can probably give you her opening sentence, while the second may struggle to recall what the essay argued.
The gap shrank in later sessions, once people knew the question was coming, but the ChatGPT group still did worst [1]. They also felt less ownership: only half called their first essay fully their own, while almost everyone in the no-tool group did [1]. If you did not write the sentences, you never had to choose them or fix them, so there was little to remember.
Why the search group changes the story
The part that settles it for me is the group most coverage skipped. Search users also relied on an outside tool, and their brain connectivity was also well below the no-tool group, roughly a third to a half lower depending on the measure [1]. Yet they quoted their essays just as well as the people who used nothing [1]. The difference is that search users still wrote their own sentences.
So using a tool did not hurt recall. Handing over the writing did. Psychologists call this kind of trade cognitive offloading: using physical action or an outside aid to reduce the mental demand of a task, such as setting a phone reminder for an appointment [2]. We offload all the time, and it is usually sensible. The cost is predictable: whatever you hand off, you do not store yourself.
Think about driving to a friend’s new house with turn-by-turn directions on your phone. You arrive without trouble, but if someone asked you to draw the route afterwards, you might not manage it. Nothing is wrong with your sense of direction. You simply never had to learn that route.
It is also why lower connectivity alone does not worry me. Search users showed it too, and they had no memory problem.
What the headlines got wrong
TIME’s headline was about critical thinking, but the study never tested critical thinking. It recorded EEG, analyzed the essays, and interviewed people about them [1]. Nobody was tested before and after months of AI use either, so “eroding” goes beyond what was measured.
The authors noticed. Their FAQ answers the question of whether AI is making us “dumber” with a plain no, and asks journalists in particular not to use words like “stupid,” “dumb,” “brain rot,” “harm” or “damage,” because “It does a huge disservice to this work.”
To be fair, the alarm did not come only from reporters. The paper’s own abstract speaks of “a likely decrease in learning skills” [1]. The lead author told TIME that “Developing brains are at the highest risk,” though every participant was an adult [1]. That is a worry worth studying, not a result.
The real cost, and what to do about it
I do not want to wave the study away. If the point of writing an essay is to learn, producing text you cannot recall or call your own is a genuine loss. The study also cannot rule out longer-term effects; it did not test them, which is why the authors call for longitudinal studies [1].
Other research points the same way. In a survey of 319 knowledge workers, researchers at Microsoft Research and Carnegie Mellon found that people with more confidence in generative AI reported less critical thinking, while people with more confidence in themselves reported more [3]. A survey shows association, not cause. Participants also said AI shifted their critical thinking toward checking information and overseeing the task [3]. That matches how I think about the models I build: their output is a draft to check, never an answer to trust.
My practical advice is my inference, not something the study tested. If you are a student, write your own first draft, then bring AI in to question it, tighten it or check it. The thinking you do yourself is the part you keep. If you teach, look at drafts and outlines, or ask students to explain their argument out loud.
AI is not making our brains dull, at least not on this evidence. What it can do is let us skip the part of the work where learning happens, and whether we skip it is our choice.
So is AI making our brains dull? Not on this evidence. In this study it made people remember less of what they did not write, which is what offloading predicts. Decide which work you want to own, and do that part yourself.
References
[1] N. Kosmyna, E. Hauptmann, Y. T. Yuan, J. Situ, X.-H. Liao, A. V. Beresnitzky, et al., “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task,” arXiv:2506.08872v1, Jun. 2025, doi: 10.48550/arXiv.2506.08872.
[2] E. F. Risko and S. J. Gilbert, “Cognitive Offloading,” Trends in Cognitive Sciences, vol. 20, no. 9, pp. 676-688, Sep. 2016, doi: 10.1016/j.tics.2016.07.002.
[3] H.-P. Lee, A. Sarkar, L. Tankelevitch, I. Drosos, S. Rintel, R. Banks, 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,” in Proc. 2025 CHI Conf. Human Factors in Computing Systems (CHI ’25), Yokohama, Japan, Apr. 2025, pp. 1-22, doi: 10.1145/3706598.3713778.
Related Articles
