Cognitive Debt: The Hidden Cost of Letting AI Think for You
By Minerva Next Team | | 10 min read
MIT's viral EEG study found that outsourcing your thinking to ChatGPT builds cognitive debt. Here is the science and how to use AI without dulling your mind.
Ask a chatbot to write your essay and you get a finished draft in seconds. Ask yourself, an hour later, to quote a single sentence from that essay, and you may find you cannot. That gap between having a document and having learned anything from it now has a name, and a growing pile of evidence behind it: cognitive debt.
The phrase comes from a 2025 MIT Media Lab study that wired up students' brains while they wrote and found something uncomfortable. The people who leaned on AI produced perfectly acceptable essays and remembered almost nothing about them. Their brains, quite literally, had checked out. For students racing through assignments with a chatbot open in the next tab, this is the most important learning-science finding of the year, and it is not the simple "AI rots your brain" headline you may have seen. The real story is more useful, and it hinges entirely on how you use the tool.
What "cognitive debt" actually means
In the study Your Brain on ChatGPT, researchers Nataliya Kosmyna, Pattie Maes, and colleagues asked 54 people to write essays across several sessions while wearing high-density EEG caps that measured brain activity. Participants were split into three groups: one wrote with no tools at all ("brain-only"), one used a search engine, and one used ChatGPT.
The pattern in the brain data was stark. The brain-only writers showed the strongest and most widely distributed neural connectivity, the kind associated with executive function, attention, and memory formation. Search-engine users sat in the middle. The ChatGPT group showed the weakest connectivity of all, with reduced activity in the alpha and beta bands that support internal focus and working memory.
The behavioral results were even more telling. When asked shortly afterward to quote from the essay they had just submitted, more than 80% of the ChatGPT group could not reproduce a single line, compared to almost none of the brain-only group. Many reported little sense of ownership over the work. The essay existed; the learning did not. The researchers coined "cognitive debt" for this shortfall: each time you let the tool do the mental heavy lifting, you skip the effortful processing that builds understanding and memory, and the bill comes due later when you actually need to know the material.
It is not just essays
You might dismiss one essay-writing study as a quirk. The trouble is that it lines up with a larger body of recent research pointing the same direction.
A 2025 study by Michael Gerlich in the journal Societies, titled AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking, surveyed and interviewed 666 people across a wide range of ages and backgrounds. It found a significant negative correlation between frequent AI-tool use and critical-thinking scores, and that the link was driven by "cognitive offloading," the habit of handing mental tasks to an external tool. The effect was strongest in the youngest participants, aged 17 to 25, who used AI the most and scored lowest on independent reasoning. Correlation is not causation, and heavy AI users may simply have different habits to begin with, but the direction of the finding is hard to ignore.
The same theme shows up among working adults. In a large study presented at CHI 2025, Microsoft and Carnegie Mellon researchers Hao-Ping Lee, Advait Sarkar, and colleagues surveyed 319 knowledge workers about the impact of generative AI on critical thinking. Across roughly 936 real examples, they found that the more people trusted the AI, the less critical thinking they reported doing. Higher confidence in your own ability had the opposite effect: it was linked to more scrutiny of what the AI produced. In other words, the danger is not the tool itself but the moment you stop checking it.
Why your brain needs the struggle
None of this should be surprising once you remember how learning works. Memory is not a recording device that captures whatever passes in front of it. You remember things in proportion to how much you actively process them, a principle cognitive scientists call effortful processing, or "desirable difficulty."
When you wrestle a concept into your own words, retrieve a fact from memory instead of looking it up, or generate an answer before checking it, you force your brain to build and strengthen the neural connections that make the knowledge durable. This is why active recall and self-testing beat rereading, and why the productive frustration of figuring something out yourself pays off weeks later on the exam. The effort is not a side cost of learning. The effort is the learning.
Generative AI is, in a sense, a machine for removing exactly that effort. It hands you the polished output and quietly skips the messy middle where understanding is actually forged. The EEG data from the MIT study is a picture of a brain that has been relieved of the struggle, and has therefore been relieved of the learning too. Cognitive debt is what you get when the difficulty disappears but the need to know the material does not.
The plot twist: AI can also make you smarter
Here is where the "AI is bad for your brain" narrative falls apart. In the very same month the MIT study went viral, a different Harvard experiment reached what sounds like the opposite conclusion.
In a randomized controlled trial published in Scientific Reports, Gregory Kestin and colleagues compared students learning physics through an AI tutor versus in-class active learning. The students using the carefully designed AI tutor learned roughly twice as much, in less time, and reported feeling more engaged and motivated than peers in a strong active-learning classroom. The key detail is in the design: this was not a chatbot spitting out answers. It was built with expert-authored scaffolds, step-by-step reasoning, guardrails against giving away solutions too early, and prompts that pushed students to think rather than copy.
So which is it, brain poison or super-tutor? Both findings can be true because they describe two completely different behaviors. One study measured people using AI to avoid thinking. The other measured people using AI to be guided through thinking. As learning scientists Vitomir Kovanovic and Rebecca Marrone argued in a widely shared analysis of the MIT study, the alarmist reading misses the nuance. Even in the MIT data, the group that did their own thinking first and only then brought in AI showed higher brain engagement than the group that reached for the chatbot from the start. Sequence, it turns out, is everything.
How to borrow from AI without going into debt
The lesson is not to swear off AI. It is to make sure the tool amplifies your thinking rather than replacing it. A few habits keep you on the right side of that line.
Think first, then prompt. Before you open a chatbot, spend a few minutes generating your own answer, outline, or rough draft. Even a bad first attempt primes your brain to engage with whatever the AI gives back, and it preserves the effortful processing that builds memory. The order matters more than almost anything else.
Use AI to challenge you, not to finish for you. Instead of "write my essay on the causes of the French Revolution," try "here is my argument, now poke holes in it" or "quiz me on this chapter and do not show me the answers until I try." Turning the tool into a sparring partner or a Socratic tutor keeps you in the driver's seat, which is exactly the design that made the Harvard tutor work.
Retrieve before you look up. When you hit a gap, resist the reflex to paste the question straight into a chatbot. Try to pull the answer from memory first. Struggling for even ten seconds before checking dramatically improves how well the answer sticks, because you have activated the retrieval pathways the AI would otherwise let you skip.
Rewrite everything in your own words. If you do use AI to explain a concept, close the tab and reconstruct the explanation yourself, out loud or on paper. If you cannot, you have found a genuine gap, which is far better to discover now than in the exam hall. This one step converts passive reading of an AI answer into active generation.
Verify, do not trust. Remember the Microsoft finding: confidence in the tool is where critical thinking goes to die. Treat every AI output as a draft from a fast but unreliable collaborator. Check its claims, question its logic, and stay the expert in the room.
The bottom line
Cognitive debt is real, and the students most exposed to it are the ones using AI exactly as it is easiest to use: to make the work disappear. The MIT brain scans are a warning about a specific behavior, not a verdict on a technology. The same category of tool, pointed in the opposite direction, doubled learning in a Harvard classroom.
The difference comes down to a single question you can ask yourself every time you reach for a chatbot: is this tool doing my thinking, or is it making me think harder? Get that answer right, and AI becomes one of the most powerful study aids ever built. Get it wrong, and you graduate with a transcript full of essays you cannot remember writing.
At Minerva Next, we are building a learning platform designed around that distinction. Instead of handing you finished answers, IntelliMind turns your notes and course material into active recall prompts, quizzes, and Socratic questioning, so the AI does the coaching while your brain does the learning. The goal is not to think for you. It is to make you think better.
References
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 Task. MIT Media Lab (arXiv:2506.08872). https://www.media.mit.edu/publications/your-brain-on-chatgpt/
Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), 6. https://www.mdpi.com/2075-4698/15/1/6
Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/
Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI Tutoring Outperforms In-Class Active Learning: An RCT Introducing a Novel Research-Based Design in an Authentic Educational Setting. Scientific Reports, 15, 17458. https://www.nature.com/articles/s41598-025-97652-6
Kovanovic, V., & Marrone, R. (2025). MIT Researchers Say Using ChatGPT Can 'Rot' Your Brain. The Truth Is a Little More Complicated. The Conversation. https://theconversation.com/mit-researchers-say-using-chatgpt-can-rot-your-brain-the-truth-is-a-little-more-complicated-259450