Yes, if you let it do the composing while you watch. Researchers at the MIT Media Lab put this to a direct test. Nataliya Kosmyna and seven co-authors split 54 people into three groups, one writing essays with ChatGPT, one using a search engine, one writing unaided, and tracked their brain activity with EEG across four months (arXiv:2506.08872, June 2025). Cognitive debt is the term they use for what builds up when a language model does your composing: a cost that surfaces later as a weaker hold on material you apparently just produced yourself. The ChatGPT group showed the weakest neural connectivity of the three; the unaided group’s was the strongest and most distributed. This is the study behind Agni, a learning app in development that tries to measure where people actually stand with AI rather than how worried they feel about it.
What did the MIT study actually find?
Kosmyna’s team ran four sessions over four months. In each of the first three, participants wrote an essay under one of three conditions: ChatGPT, a search engine, or no tools at all. EEG recorded brain activity throughout, and researchers scored the resulting essays. Nothing else about the task changed between groups, only who did the composing.
| Group | Tool used | EEG connectivity |
|---|---|---|
| Brain-only | None | Strongest, most distributed |
| Search engine | Web search | Moderate |
| ChatGPT-assisted | LLM | Weakest of the three |
The gap wasn’t between “used a tool” and “didn’t”. It was between outsourcing the sentence-by-sentence composing and doing it yourself, even with help finding material. Search-engine users still had to read, judge and select; ChatGPT users mostly accepted and lightly edited.
Why does writing with ChatGPT weaken your own grip on the material?
Composing a sentence forces you to choose words, resolve ambiguity and hold the argument’s shape in your head as you build it. Each of those is a small act of retrieval, and retrieval is what makes material stick. Hand the composing to ChatGPT and you skip most of that: you’re reading and approving rather than generating. Kosmyna’s group frames this as debt rather than loss, because the cost isn’t visible at the moment you write. It shows up later, when someone asks you to defend a decision, explain a paragraph in a meeting, or simply repeat what you just said you’d concluded, and you find the reasoning isn’t actually there. You have the output. You don’t have the argument that produced it.
What’s a checkable test you can run on yourself?
Close the document. Wait thirty seconds. Then try to say, out loud and unaided, what you just concluded and why, or quote one sentence of it from memory. If you can do that, the material is yours regardless of how much AI helped produce it. If you can’t, that’s not a character flaw: it’s the mechanism above at work, composing got outsourced, so little got encoded. Run this straight after finishing, while the piece is still fresh, rather than the next morning. This test tells you something different from a general skills check: it’s specific to the piece of work in front of you, not your standing on AI overall. Agni’s AI Literacy Score quiz, a fourteen-question self-assessment at couragehorizon.com/agni, is built for that broader question instead, scoring where you sit against others in your job and country rather than whether one document is actually yours.
Does this mean any use of AI erode your thinking?
No. The mechanism in Kosmyna’s data is specific: it’s what happens when a model does your composing and you approve the result. The search-engine group showed moderate connectivity, distinct from the ChatGPT group’s weakest-of-three result, even though both search and ChatGPT involved an external tool. The distinguishing act was writing the sentences yourself. Asking ChatGPT to challenge a draft you already wrote, check your maths, or find a counterargument to your own conclusion keeps the composing on your side of the line. Letting it write the conclusion and skimming the result is what put the ChatGPT group’s connectivity at the bottom of the three.
Common questions
Isn’t this the same as worrying I’m generally behind on AI? No. That’s a separate question about self-assessment anxiety, covered in two other Agni articles: How do I know if I’m behind on AI, or just anxious about it? and Does everyone your age already understand AI better than you do?. This one is about a specific mechanism instead: what gets lost when a model composes for you, checkable with a quote-or-explain test rather than a feeling of falling behind.
Does this apply to other AI tools, or only ChatGPT? Kosmyna’s team tested ChatGPT specifically, not the whole category. The mechanism they describe, composing skipped means little retrieval means weak recall, has no obvious reason to stop at one product. Any tool that writes finished sentences for you while you approve them puts you in the same position; the study just measured this one.
What do I do if I fail the quote test? Redraft the piece yourself first, even badly, then bring ChatGPT in to tighten or check it rather than generate it. The mechanism this article describes is about which half of the work you do, not which tool touches the document at all: write the sentences yourself and the retrieval that builds a real grip on the material happens regardless of what checks the draft afterwards.
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. arXiv:2506.08872