Does AI rot your brain?

A few weeks ago I needed to quickly get up to speed on research outside my own corner of social psychology: how do children decide whether someone is worth believing? What cues do they look for? In what ways is credibility different from power or status?

In the past, this would’ve taken me an afternoon at least. Sprawling database searches while I tried to find the right keywords (selective trust is the term, fwiw). A browser sprouting tabs. Rabbit holes and goose chases and occasional goldmines. Somewhere in the second hour, a rough map would start to form in my mind.

This time, I asked an AI.

Forty seconds later I had the synthesis: the major research programs, the key findings, the live disputes. The whole thing was much cleaner than my rough map would have been.

And it was good. I’m hesitant to admit that because academics like to dunk on AI, but I checked the references, looked at the evidence, and it all stood up to scrutiny. It was accurate, well-organized, and aware of the gaps in the field. If a student had turned it in as a literature review, they would’ve gotten an A.

But the experience of reading a lit review is radically different from the experience of reviewing the literature.

It was lighter. I’d arrived without traveling. The knowledge was in front of me, but it definitely wasn’t in me.

If you’ve used these tools for work, you probably know the feeling. It’s the feeling an entire genre of headlines is now trying to explain.

Oxford University Press named “brain rot” its word of the year in 2024. Researchers put EEG caps on students while they’re writing essays and find that the ones using ChatGPT showed weaker neural activity than the ones writing unassisted. Many of them, couldn’t even quote from their own essays minutes after finishing. Before that Microsoft linked up AI use with less critical thinking skills, which is more than a little ironic. And other researchers have joined the now common warning that using AI will make you dumb.

We all feel that thinking through something with LLMs is a different way of thinking. Then researchers find a way to measure that difference, and that difference often has a valence. Then headlines call it proof and sound the alarm bells.

I don’t want to be glib about this. The worries we all have about how AI might change us are real. Teachers now have an extra layer of evaluation, trying to judge whether students even wrote the thing. I’ve had beers with developers as they describe being increasingly alienated from their labor. All of us have likely felt our attention change as it is bought and sold by platforms online. The people sounding the alarm aren’t fools.

But, let’s exercise some of our critical thinking while we still have it.

In order to support the claim that AI damages cognition, researchers need a few different things. First, they need to turn “AI use” into a variable, which is hard to do because AI use is not one thing. I’ve used AI to shop, to summarize articles, to build my website, and to dialog with about topics I don’t understand. Those are all pretty different sorts of AI-use.

Second, researchers need to know what AI use is replacing. What are we comparing it to? You can control for this by having the other group in an experiment write an essay, but generalizing this comparison is very fraught.

And finally, researchers need to say what they mean by damage. Take the EEG finding, and try to avoid the seductive allure of neuroscience. A brain showing less activity while outsourcing a task to AI isn’t necessarily a brain in decline. It’s just a brain outsourcing.

That doesn’t mean these studies are useless. But it does mean we should hold their findings lightly. They suggest specific kinds of AI use may reduce certain kinds of cognitive activity. At the same time though, researchers are using AI systems to make scientific breakthroughs, which doesn’t seem like cognitive decline.

In the absence of solid evidence, the discourse is leaning heavily on a metaphor. The brain is a muscle; so thinking is exercise, offloading is disuse, and disuse is atrophy.

Each link in this chain feels obvious, but it depends on the overarching metaphor. And if we step back from it, I feel safe saying that forming a judgment is not the same as performing a pull-up. Metaphors shape what we see and this one smuggles in its conclusion. Muscles tend to get weaker if you don’t use them, so once you accept the metaphor, the decline follows automatically.

The typical rebuttal to concerns about AI-use making us dumber is that these are old concerns. And the consequences are never so bad Socrates worried that writing would rot memory, but that was fine. People fretted over calculators and GPS, and... well I guess jury’s out on that one. But, the point stands that tools have always changed us and yet here we still are.

What the rebuttal doesn’t offer is a reason. Why is it that we tend to be good at adapting to new ways of thinking? The reason, I believe, is that our minds have always been interwoven with tools. In other words, there’s never been an unscaffolded mind to protect.

We think in terms of languages and categories that we inherit. That’s cognitive scaffolding. Many couples split up the work of remembering things: you remember names and I’ll hold the directions. Or you’re like me and my wife and neither of us remember so we offload our memory to calendars and notes. Writing, paper, pencils, libraries, colleagues, rituals, institutions, a fair bit of our social worlds consist of cognitive scaffolding.

The more you widen your view of how we actually think, the less the mind looks like an individual information processor that can be damaged or enhanced. Thinking is a social and material practice that runs through our brains and our tools. Cognition has never been a private substance stored in our heads.

This pristine and solitary mind that some people want to protect from AI is a mind that never existed. We’re more robust than that.

Before LLMs were Large, a team of psychologists was studying the effects of internet-use on cognition. They found that access to the internet led people to have lower rates of recall of information. Sound familiar? But they also found that people had better recall for where to access that information. In other words, cognition wasn’t rotting, it was adapting to the affordances of a new environment.

This is where the extended-mind crowd usually declare victory and go home. The mind has always been interwoven with the world, and AI is just more world. So this is standard fare.

But that’s a little too simplistic.

A library and a casino are both parts of the world. Both environments shape the minds that spend time in them. But nobody’s gonna think they’re equivalent.

The question was never whether tools shape our thinking. Of course they do. The question is what shape will our minds be in?

The obvious answer is that they’ll be accustomed to speed. I wrote about that in The Flood: research timelines are being compressed by AI from months to hours, and when this happens, when production outruns judgment, we lean more heavily on heuristics.

But, the more nebulous change isn’t just speed. It’s the shift that unlocks speed: these tools are impacting how the assembly of thought occurs.

When Socrates was fretting about writing, he was worried about its impact on memory. The google study above was about recall. Same with the EEG study-- they were looking at whether students could remember what they wrote. Focusing on memory is strategic because it’s easy to assess: you either remember or you don’t. But it’s also a pretty narrow slice of thought.

That matters because with LLMs the cognitive process being outsourced isn’t primarily memory. It’s synthesis.

Outsourcing synthesis isn’t new, of course. Teachers synthesize. Review articles synthesize. Books synthesize. Synthesis has always been both internal and external.

The difference is that we typically only encounter those syntheses after we’re already a ways down the road of inquiry. You have to put in some effort to find them. And once you do find them, they often rub up against the idea you were bringing to them.

These sources of synthesis have authors, edges, publication dates, they have visible texture and grit. They resist your ideas in ways that matter.

The chatbot is different because it offers nearly immediate synthesis. I asked for a summary of selective trust and got one in less than a minute. This moves assembly up into the opening moments of inquiry, and the synthesis arrives fully-formed. This is precisely where confusion and friction would normally be doing its work.

I know friction is having a bit of a moment as a good thing. I’m about to jump on that bandwagon, but it’s worth remembering that a lot of friction is just pain and annoyance. It’s producing a report you don’t care about or figuring out how to replace the battery in your car. A lot of people have a lot of friction in their lives already. Cognitive offloading is not bad in itself.

But when it comes to learning, some friction is irreplaceable.

That AI summary on selective trust was good, but I hadn’t learned the topic by simply reading it. It was only after I’d read the articles it referenced and then followed their references. I had to get confused and try to explain that confusion in my own words.

That sort of friction is what learning is. And it’s not something I’d want to offload, because it’s something I enjoy doing.

One of my favorite philosophers, John Dewey, spent fifty years insisting that thinking is the process of inquiry itself, not the beliefs that process settles on. It’s the journey, not the destination, dude.

But the cliche hides the deeper point that we all too often focus on the ends and forget that they’re means towards the next end. AI delivers ends. Ends that are custom fit to whatever your particular problem or question might be.

The usual gloss on all this is reassuring: you’re the editor now, not the writer; the curator, not the maker. But if all you receive are finished products, what’s the cognitive process we should be watching?

Not construction or recall. It’s judgment.

This is why there are all these headlines now about taste. Taste is judgment when you’ve got lots of options. It’s the ability to sort through what matters and what doesn’t, what’s a helpful frame and what isn’t.

At least some of what’s being tracked as cognitive deterioration is likely just a cognitive shift from construction to judgment. We’ve got a new tool and the work is relocating.

That’s a fine shift for those of us who’ve already developed judgment. But what about everyone else?

My kids are in elementary school, which means their cognitive capacity for judgment and evaluation... let’s say they aren’t quite there yet. But they are under construction. And they’re building judgment the only way evaluators get built: by the loop of confusion and correction, doubt and belief.

As that loop spins they’re absorbing information, but they’re also learning the standards of communities that care about getting things right.

When my kiddo writes his report on stonefish, the point is not the report (the world’s supply of stonefish knowledge will survive without it). The point is the assembly. The searching, choosing, misunderstanding, revising, summarizing, explaining.

If he delegates synthesis now, before his evaluator exists, the tool occupies the place where his thinking was scheduled to be built.

This isn’t just an issue with kids. There’s a whole slew of tedious entry level positions where people are able to see evaluation close up and practice and learn what good work feels like from the inside. Of course, these are also the jobs most likely to be offloaded.

It’s worth noting that this is what alignment looks like at scale. Institutions train newcomers what to value, what sort of work to lift up. This is a messy process for sure, and often leads to misalignment, but it’s what we’ve got. And we’re now in the process of rerouting a fair amount of that process.

Teachers are the ones watching this formation process up close, and as far as I can tell they’re concerned. Meanwhile CEOs continue to push for more use.

This is all why I’m less interested in one-off studies about AI and brain rot and more interested in the way this technology may accelerate various forms of social erosion that had started well before its arrival. Public education, democratic structures, the soaring gap between the rich and the rest of us. All of these were in bad shape well before LLMs hit the scene.

And if we take extended cognition seriously, which I do, then it’s worth noting that the tools the mind is surrounded by are way less important to how we think than the environment in which those tools and mind are embedded. You know what damages cognition? Stress, precarity, not knowing who to trust.

Is AI making me dumber?

The question isn’t wrong, it’s just a bit small. It’s sized for a individual isolated mind that’s just one small part of something much larger.

Zoom out a bit and memory looks like something shared across notes and tools and friends.

Zoom out a bit more and thinking itself looks less like rummaging through the basement and more like working out in the garage with the doors open to the community.

And zoom out once more and you’ll see that those communities sit inside schools and institutions and economies that were wobbling before any of this showed up.

I opened by saying the summary was in front of me but not in me. That’s still true, but it’s probably the wrong distinction. The sort of thinking I care about was never really “in me.” It runs through the tools I use, my teachers and friends, and the broader social ecosystems we all work within.

So if we want to know how AI-use is changing our cognition, we’ve gotta ask which layer of thought do you mean? These tools may make us smarter and faster at an individual level but create instability at a social level, which then makes it harder to think clearly. Studying any piece of that is going to miss the whole.

The better question won’t have a straightforward answer: what kind of thinking does AI make possible, what does it foreclose, and are we equipped to track the difference?

The ability to distinguish reliable syntheses from those that confirm what we think, is a judgment that gets built through friction, in communities, over years. This means that every time we use these tools, we’re not just affecting our own cognition; we’re also shaping how the next person’s judgment will or won’t get built.