A research group led by Nataliya Kosmyna at the MIT Media Lab conducted four sessions to measure what happens in the brain when people write essays using ChatGPT versus without it. Fifty-four participants, divided into three groups—one using the language model, one using traditional search, and one relying solely on their own thinking—along with an EEG that records neural connectivity across all frequency bands. The first result is as expected: The AI model group consistently showed the weakest connectivity, while the “brain-only” group showed the strongest. Those who rely on external tools activate fewer neural connections. That’s understandable—every tool comes at a cost, and reducing effort is often precisely the point.
The session that’s not so easy to wrap up is the fourth one. That’s when the AI-model group switched to writing without assistance—and their connectivity didn’t return. It remained below the level of those who had worked without an AI model from the start. No reset, no baseline to return to as soon as you put the tool down. The authors call this “cognitive debt.”
The term has no moral connotation. It’s used in an accounting sense, by analogy with technical debt in software development: What you don’t think through yourself today doesn’t disappear, but is carried forward as a deficit—and can’t be cleared in the short term simply by putting in more effort next time. That’s the inconvenient part. The convenient part was that you handed off the work and finished faster.
Before this sounds like more than it is: The authors themselves acknowledge the limitations of their study, and these are significant. It is a preprint, not yet peer-reviewed, based on a small, purely academic sample from a single setting. The testing was text-based and conducted exclusively with ChatGPT—a single model, a single modality. The EEG only captures the surface anyway; what happens in the hippocampus, where memory is formed, would have to be shown by an fMRI, which is still missing. They measured essay writing in an educational context: The model provides the text that the user adopts—not AI in general, nor the AI model acting as a second reader of one’s own draft.
It is precisely this narrow scope that makes the finding useful. It specifies what it applies to—the outsourcing of writing itself—and what it does not. Anyone who reads “AI makes people stupid” into it is stretching it too far; anyone who dismisses it for that reason overlooks the fact that it aligns with more robust findings: the generational effect in memory research, and Sternberg’s “use it or lose it.” We retain and understand what we generate ourselves better than what we read. The fourth session merely shows the other side of the same coin.
It also reveals something that got lost in the headlines. In the other direction—the “brain-only” group switched to the AI model—memory performance was higher than among those who had written using the AI model from the start. Their own preliminary work acted as a buffer. Those who thought for themselves first can use the AI model afterward without building up the same sense of guilt. The order matters.
That’s the practical point, and it has little to do with resisting technology. Cal Newport advocated for focused work long before language models existed; AI merely shifts his question—it doesn’t eliminate it. The question now is: What needs to happen first for the model to be useful afterward? Putting your own thought process first—spending the first few minutes on a problem without tools, until you’ve formed your own understanding—is not an exercise in virtue. It’s capital preservation. The AI model expands on a judgment that’s already there; it cannot replace a judgment that was never formed, only make it appear to be replaced.
This diagnosis, too, was generated using an AI system. It analyzed sources, challenged my assumptions, and developed drafts. But not at the beginning. It all started with the question of what actually matters in this finding—and I answered that for myself before the system came into play. It’s the difference between the first and final instance, and according to Kosmyna, it’s not merely stylistic. It’s visible in the EEG.
The bill that’s being put off here won’t come in the form of a crisis. It will come as a slowly declining ability that no one notices, because the results were always presentable. Whoever decides what to hand over to the machine also helps determine what they still believe they are capable of doing on their own once it is no longer there. This is not a question for later. It is the order of the day.
Sources
- Nataliya Kosmyna et al. (MIT Media Lab), “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Tasks,” arXiv:2506.08872, 2025 (preprint, under review): https://arxiv.org/abs/2506.08872 · Study website: https://www.brainonllm.com/
- Cal Newport, *Deep Work: Rules for Focused Success in a Distracted World*, Grand Central Publishing, 2016 (German edition: *Konzentriert arbeiten*, Redline 2017)
- Norman J. Slamecka & Peter Graf, “The Generation Effect: Delineation of a Phenomenon,” Journal of Experimental Psychology: Human Learning and Memory, 4(6), 1978, pp. 592–604: https://doi.org/10.1037/0278-7393.4.6.592
- Robert J. Sternberg, “Does AI increase cognitive abilities, decrease them, or a little bit of each?”, Frontiers in Education, 11:1759062, 2026 (“use it or lose it” principle): https://doi.org/10.3389/feduc.2026.1759062
- Wiki: manage-ai-decision-research (Threads 5–6) · system-competence-and-judgment
