In a recent post, I explored whether AI genuinely increases productivity or just makes us work more, drawing on a Harvard Business Review study by Ranganathan and Ye (2026). One of the points I raised was the cognitive cost of AI use, the idea that short-term gains in speed and output can mask deeper problems with how we think, learn, and remember.
This study from MIT’s Media Lab puts hard neuroscience behind that concern. Kosmyna et al. (2025) tracked what happens inside the brain when people use ChatGPT to write essays, and the findings are striking. Over four months, participants who relied on an LLM for writing showed weaker neural engagement, poorer memory, and less ownership of their own work compared to those who wrote without AI assistance.
The researchers call this pattern “cognitive debt,” and it accumulates quietly.
Your Brain on ChatGPT: Less Effort, Less Learning
The study compared three groups: participants who wrote essays using only their own thinking (brain-only), those who used a search engine for support, and those who used ChatGPT. Brain imaging revealed a clear gradient. Kosmyna et al. report that “brain connectivity systematically scaled down with the amount of external support: the Brain-only group exhibited the strongest, widest-ranging networks, Search Engine group showed intermediate engagement, and LLM assistance elicited the weakest overall coupling” (p. 2).
In plain terms, the more AI did the work, the less the brain engaged. The neural networks responsible for planning, memory, and executive control, the very processes that make writing a learning experience, were significantly quieter in the ChatGPT group.
This connects directly to what Shaw and Nave (2026) describe as cognitive surrender, the tendency for people to accept AI-generated outputs without critically processing them. The Kosmyna study adds a neuroscience layer to that concept. Cognitive surrender isn’t just a behavioral pattern. It has a measurable signature in the brain.
Memory Takes a Hit
One of the most unsettling findings is what happened when participants were asked to recall or quote from essays they had written just minutes earlier. The ChatGPT group struggled. As Kosmyna et al. note: “The LLM group also fell behind in their ability to quote from the essays they wrote just minutes prior” (p. 2).
This matters because memory encoding is one of the strongest indicators of genuine learning. When you wrestle with an idea, organize it in your own words, and commit it to a page, your brain stores that work. When ChatGPT handles the heavy lifting, the product may look polished, but the learning doesn’t stick. The essay exists, but the understanding behind it is shallow.
I wrote about a related finding in a previous post on how students who ask AI direct, targeted questions perform better on writing tasks (Cheng et al., 2025). The key in that study was active questioning behavior. Students who engaged with AI actively learned more. Students who let AI do the thinking learned less. The Kosmyna study confirms this pattern from a completely different angle: the brain literally does less when AI takes over.
The Ownership Problem
The study also found something worth paying attention to from an academic integrity perspective. Essays written with ChatGPT were more linguistically homogeneous. NLP analysis showed similar n-grams, named entities, and topic structures across LLM-assisted essays, often resembling default AI responses. And when asked about authorship, many ChatGPT users reported weak or partial ownership of their texts.
Kosmyna et al. found that “the reported ownership of LLM group’s essays in the interviews was low” (p. 2).
Compare that to the brain-only writers, who consistently expressed strong authorship and personal connection to their work. There’s something important happening here beyond plagiarism concerns. When students don’t feel like the work is theirs, the motivational and cognitive benefits of writing, the sense of having thought something through and produced something original, disappear.
Prior Thinking Changes Everything
One of the most hopeful findings in the study is about sequencing. When participants first wrote essays independently and then used ChatGPT in a later session, their brain activity actually increased during AI-assisted writing. Prior cognitive effort appeared to scaffold more active and strategic engagement with the tool.
This has direct implications for how we teach with AI. If students begin a task by thinking independently, forming their own ideas, and putting words on the page before involving AI, they’re more likely to use AI as a genuine thinking partner. If they start with ChatGPT from the beginning, the opposite happens. Their neural engagement drops and stays low.
The sequencing principle reinforces something I’ve been arguing across multiple posts: the first draft, the first set of ideas, the first attempt at reasoning, all of that needs to come from the student. AI can refine, challenge, and push that thinking further. But the thinking has to exist first.
Your Brain on ChatGPT: The Long-Term Picture
The four-month duration of this study is what gives it weight. Short-term, ChatGPT-assisted essays often scored well by both human teachers and AI judges. The surface-level quality was there. But over repeated sessions, the LLM group fell behind at every level. Kosmyna et al. are unambiguous: “As we demonstrated over the course of 4 months, the LLM group’s participants performed worse than their counterparts in the Brain-only group at all levels: neural, linguistic, scoring” (p. 2).
Short-term performance masked long-term decline. That’s the cognitive debt the title refers to. Each time you let AI handle the thinking, you borrow against future capacity. The interest compounds quietly, and eventually the cost becomes visible in weaker writing, weaker recall, and weaker thinking.
This doesn’t mean AI has no place in writing or learning. It means the way we use it determines whether it builds capacity or erodes it. AI directed toward mechanics, formatting, and language polishing leaves the cognitive core intact. AI that takes over ideation, argumentation, and reasoning creates debt that accumulates with every use.
The lesson for educators is clear. Teach students to think first, write first, struggle first. Then bring AI in as a support tool, not a replacement for the intellectual work that makes learning meaningful.
Reference
- Cheng, Y., Fan, Y., Li, X., Chen, G., Gašević, D., & Swiecki, Z. (2025). Asking generative artificial intelligence the right questions improves writing performance. Computers and Education: Artificial Intelligence, 8, 100374. https://doi.org/10.1016/j.caeai.2025.100374
- Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing tasks. MIT Media Lab. https://www.media.mit.edu/publications/your-brain-on-chatgpt/
- Shaw, S. D., & Nave, G. (2026). Thinking fast, slow, and artificial: How AI is reshaping human reasoning and the rise of cognitive surrender. Working paper, The Wharton School, University of Pennsylvania. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646