AI Brain Rot Is Real. What an MIT Study Found, and What to Do.
A couple of weeks ago I sat down to write landing page copy. It was something that would have taken me five minutes in the past. My first instinct was to open Claude before I’d even thought about what I wanted the page to say. I caught the compulsive move, and pushed myself to write without Claude’s help. A few sentences in, I noticed something: the writing felt curiously hard, and the words came more slowly than they used to.
It turns out I’m not alone. People are calling it different things depending on their sphere of influence or profession: AI brain rot, AI brain fry, AI brain drain. The underlying idea of each is the same: you’ve quietly outsourced a bunch of your thinking to AI, and now when you attempt to do the thinking yourself, you don’t have the continuous flow and output that you previously had. Your thinking skill and your creative craft are out of shape.
What the MIT study measured
In June 2025, Nataliya Kosmyna and her colleagues at the MIT Media Lab released a preprint called Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. Fifty-four participants wrote SAT-style essays in three groups: one using ChatGPT, one using Google search, one using nothing but their own brain. Everyone wore an EEG cap.
The ChatGPT group showed the weakest neural connectivity across the regions involved in planning, integration, and recall. The brain-only group lit up the most. The search group sat in the middle.
After the essay was done, the researchers asked everyone to quote a sentence from what they’d just written. Most of the ChatGPT users couldn’t. They didn’t remember their own essays because, in some real sense, they hadn’t written them.
Across four sessions the pattern got worse. Participants who used ChatGPT engaged their brains less in the second session than the first, and less in the third than the second.
When the researchers eventually took ChatGPT away and asked them to write unassisted, the brain activity didn’t return to baseline. The authors call that residue cognitive debt, and the unsettling part is that it doesn’t immediately lift just because the AI was removed.
The paper is what’s called a preprint, meaning the researchers shared it before formal peer review, and the sample is small. Treat it as a signal rather than a verdict.
This finding fits a larger pattern that scientists have been documenting for decades: when you stop using a mental skill, your brain stops maintaining it.
The clearest example came from a 2000 study by neuroscientist Eleanor Maguire on London taxi drivers. Back then, getting licensed required memorizing every street in the city, and Maguire’s team found that the drivers’ brains had physically changed from the effort. The posterior hippocampus, the part of the brain that handles spatial memory, had grown noticeably larger than in non-drivers. When GPS arrived a few years later, the next generation of drivers didn’t need to memorize the streets, and that brain advantage disappeared. The same shift has happened in smaller ways with calculators and mental math, and with spell check and spelling: once a tool reliably handles the job, the underlying skill stops being maintained.
So the mechanism is old and well-documented; the reason it feels more significant in this example is that AI is involved in far more of what a knowledge worker does in an average day than any of the older tools ever did, which leaves a lot more skills at risk of atrophy.
What AI brain rot feels like
A few patterns come up repeatedly when people talk about this.
You catch yourself opening ChatGPT or Claude before a thought has had time to form, a similar reflex to the one that pulls you toward your phone when a feeling of discomfort or boredom shows up.
You feel a disconnection from the work itself. Once the novelty of “hey, the machine did this pretty well!” wears off, you’re left with an empty feeling, a lack of accomplishment. You haven’t exercised that most human of activities: creating.
It feels like a time-saver until you count the hours you spend trying to put yourself back into what the model produced. Whatever you used to make from a blank page was made of small particular choices that came from you and no one else: the phrase no one else would have written, the chord change that surprised you, the crop you spent ten minutes deciding on. The model’s first draft has none of those. You re-prompt and re-edit and ship something that mostly works, but it doesn’t quite read or sound or look like you wrote it.
If any of this sounds familiar, you’re not imagining it, and a lot of other people you respect are quietly experiencing some version of the same thing.
Why this is different than social media brain rot
It’s tempting to lump this in with social media brain rot, since both often circle around the idea of not being able to focus. The problem is significantly different, however. Doomscrolling costs you time and attention while you’re scrolling, but the skill you would have used in that time is still there when you finally close the app. Letting AI do a task however, means the skill the task would have required isn’t getting exercised at all, and over a few months that compounds differently than a daily hour of Instagram does.
Writers feel this first because writing is thinking communicated through words, but it shows up anywhere reasoning used to happen on its own. Engineers who lean on Copilot to draft their functions lose the architectural intuition that used to tell them where bugs would live. Lawyers using AI to summarize their case files lose the slow attention that used to catch the buried detail a summary would skip. With designers, the complaint is that they’ve stopped having opinions of their own, because by the time they’d normally be forming one, the model has already shown them four options that feel like the right answer. Students notice it when they generate an essay and don’t recognize their own voice on the page, or at an even deeper level they feel a cloud of malaise because they don’t even care about writing anymore.
Who’s most at risk
Roughly speaking, anyone whose work is mostly thinking and creating. That includes knowledge workers, writers, researchers, students, designers, and founders. If the day’s main deliverable is a document, a decision, a diagnosis, or a piece of code, a design, a piece of music…you’re in the affected population.
People with ADHD have a particular challenge, because AI is exceptionally good at removing the three things that usually make deep work hard: getting started, staying with one thread, and seeing the underlying structure. For the immediate task, that can be helpful. The longer-term cost is that the skills that recover slowly during a good ADHD week, the ones that come back when you’ve slept enough or finally found the right medication, only come back when you’re using them. Outsourcing the hard parts to a model means none of those abilities ever get resurfaced from where ADHD hides them.
It’s also worse for people who are early in their career, because there’s no pre-AI baseline to fall back on. Someone whose first thousand pieces of professional writing were AI-assisted doesn’t have a version of their unaided self to recover, and won’t quite know what’s missing. The Kosmyna paper hints at this without saying it outright: the brain-only group, when later allowed to use ChatGPT, handled it better than the AI-first group did. The order people learn things in seems to matter, which is mildly reassuring if you spent any of your formative years thinking without a model doing a lion’s share of the work.
The “just use it less” problem
The obvious response to all of this, and particularly the problem of AI brain rot, is to use AI less. The problem is the same as with every other distraction, which is that the decision to use Claude gets made many times an hour and each individual decision feels small. I’ll just have it draft this one. I’ll just have it summarize this one. I’ll just have it generate the title and I’ll write the body. Every one of those is reasonable taken on its own, and the compound effect is what the EEG caps were measuring.
That’s why willpower-based advice fails for AI use the same way it fails for Instagram. The tool has been engineered to be the path of least resistance, and trying to out-decide it in the middle of the afternoon when you’re already tired is not a realistic strategy. The fix that works is to take the option off the table during the parts of the day where the cost is highest.
Using AI better
None of this is an argument against using AI. That would be a stupid argument to make in 2026. The model can produce a workable outline of almost anything in seconds, compress a six-hour research pass into ten minutes, and surface options across a problem space you wouldn’t have even considered on your own. Used well, it makes work better and faster.
Using AI well means knowing when it’s the right tool for the job and where it belongs in your workflow. That kind of judgment is what your own thinking has to do, but it can only do it if it’s in working order.
It matters which tasks you’re asking AI to do. Running an analysis on a massive spreadsheet or pulling together research summaries across myriad sources is something the model does faster and often better than you would, and nothing of you gets lost in the handoff. The atrophy shows up in other work, the kind meant to reach another human, where the small particular choices about how it reads, sounds, or looks are what make it feel like a person made it and speak the way you want it to speak, sound, or
What helps
The cognitive-debt framing from the MIT paper is the encouraging part, oddly, because debt implies a balance you can pay down. The brain-only group’s later performance with ChatGPT is the evidence that the balance is recoverable. Starting from your own thinking and bringing AI in as a tool produces different results than starting from AI and trying to bolt your thinking on afterward.
Four things help.
Draft first, prompt second. Whatever you were about to ask AI to write, write the bad version yourself first. Two paragraphs of your own real thinking, then bring AI in to sharpen it, push back on it, or check the facts. It sounds slower in the abstract, but after about a week it stops being slower in practice, because you’re working with something you believe in and the AI has something concrete to react to instead of having to invent from scratch. The other useful side effect is that the next morning, the work still reads as yours.
Carve out AI-free blocks. Pick the part of the day where the thinking matters most. For most people that’s the first ninety minutes after they sit down, before the inbox eats the morning. Make those minutes AI-free, not because there’s anything noble about it but because the underlying skill only stays in shape when you’re using it. Ninety minutes is long enough to get into real work and short enough that you don’t dread it; if ninety feels like too much during the first week, drop to sixty and build up.
Practice retrieval. Whenever you’re about to look something up or ask an AI to summarize something, try first to write down what you can already remember on your own. Cognitive psychologists Roediger and Karpicke published the foundational paper on this in 2006, and the finding has been replicated a lot since: trying to retrieve a piece of information strengthens your memory of it more than re-reading it does. It takes about thirty seconds and it works. Using your memory and creative skills can help deter AI brain rot.
Add friction where you don’t trust yourself. Most of these protocols fail not because people stop believing in them but because AI is one tab or tap away all day, the decision to use it gets made many times an hour, and willpower runs out at some point. An external block is the part that picks up where discipline can’t.
That’s where Freedom comes in. You set a session, pick the AI tools you don’t want to be able to access during deep work, and they’re blocked across your devices until the session ends. Freedom allows you to sit with the work, without an option to escape it.
Once you’ve thought about what you want and applied yourself, then by all means use AI and get the job done.
What to try this week to help with AI brain rot
Pick a ninety-minute block sometime this week, use Freedom to block ChatGPT or Claude and whichever other models you usually reach for during that window, and write the bad first draft of whatever you’d normally have outsourced. Then pay attention to how it goes.
What people tend to notice on the first attempt is that the work goes faster than they expected. The resistance to starting without help eats up more time than the writing itself does, and once you’re past that resistance and get some momentum going, the writing itself moves at roughly the pace it always has. The other surprise tends to come the next morning, when you can usually quote a few sentences from your own draft back to yourself, which for a lot of writers is the first time that’s been true in a while.
Being able to do that is what the EEG caps in Cambridge were measuring, and it’s the closest thing to a self-administered test of whether the cognitive debt is starting to come down.
Final thoughts
Working on this article reminded me of something. If I sit on the couch all day and eat junk, I lose fitness and my body shows it. The same is true for my thinking and creative skills: if I let AI do the work, they get out of shape. They need exercise (regular, effortful use) and good food (books, real conversation, the kind of input that wasn’t picked out for me by an algorithm).
I recently bought a hoodie from one of my favorite music plugin companies. It says “Keep Music Human,” with “Don’t let machines have all the fun” in small text underneath. They build a ton of AI music tools, but they know the tools are only useful when a human is making the choices.
AI gives us real new capabilities. But when we’re creating for other humans, the best first step is probably to just be human.
Frequently Asked Questions About AI Brain Rot
AI brain rot is the informal term for the cognitive dulling that people are reporting after months or years of heavy ChatGPT use. The MIT Media Lab’s June 2025 EEG study suggests there’s a real cause underneath it: when you offload thinking to an AI, the brain regions that would have done that work engage less, and that pattern can persist even after you stop using the AI for a given task. The authors call the residue cognitive debt.
No. Brain fog is usually a symptom of something physiological, like poor sleep, illness, perimenopause, long COVID, a concussion, or medication side effects. AI brain rot is closer to skill atrophy. It can feel similar in the moment, with slow recall, shallow thinking, and trouble getting started, but the cause is behavioral rather than physiological, which means the fix is behavioral too.
The evidence is early but encouraging. Decades of cognitive offloading research suggests that skills do come back when you start using them again, though the speed depends on how long they were dormant. The most useful implication of the Kosmyna paper is that the order seems to matter: people who built the skill first and then added AI as a tool appear to handle the tool better than people who started with it.
Probably, though it hasn’t been studied with EEG the way essay writing has. Engineers writing publicly about their own Copilot use describe the same general pattern: faster shipping, weaker architectural intuition, less ability to predict where bugs are going to live. The same protocol applies. Write the hard parts yourself, and let AI handle the boilerplate and the review.
