An Unlikely Team Tackles AI Brain Rot

There’s a new paper making waves called “Large Language Models as a Cognitive Virus,” from applied mathematicians in biophysics and complexity theory—the same general intellectual neighborhood that gave us Ian Malcolm. Their question is not whether an LLM can write a decent memo. It’s what happens to a culture, and to our capacity to think, if we lean on these systems too hard. Once you get past the math, it offers a useful answer about interventions available right now.

The model puts everyone into three buckets: U, people who don’t use LLMs; C, people who use them while retaining independent thought; and D, people dependent enough that they’ve largely handed over the work of thinking. The equations update those groups until the system settles down. This is standard population-dynamics machinery. The point is not memorizing derivatives; it’s adjusting pressures in the world and seeing how population behavior changes.

One parameter measures pressure on non-users to adopt LLMs. Turn it up, and the model hits a tipping point: a small increase can move society from mostly non-use to mostly use. It also shows hysteresis. Once widespread use takes hold, lowering pressure back to where it was may not reverse it; you have to lower it much further. Prevention matters more than hoping for a clean rollback. The bad outcome is not AI use. It’s lots of people in D, no longer doing much thinking of their own.

The first intervention is keeping adoption pressure low enough that people remain non-users. The authors basically say good luck doing that everywhere. Fair enough, but don’t throw it out. In a university, the point is cultivating a mind, not finishing problem sets efficiently, so strong limits on AI for cognitive work make sense. Journalism is another case: if you’re selling professional human thinking and writing, having reporters compose with an LLM is a strange arrangement. Restrictions, norms, and culture can make non-use sensible inside particular institutions.

A second option is creating pressure for ordinary users to step away. That sounds fanciful until you look at youth smoking. Truth-style campaigns changed the social meaning: cigarettes stopped looking rebellious and started looking like a product powerful old guys were using to exploit you. Young people already have some of that instinct with AI. They grew up in social media; they’re not automatically inclined to trust Silicon Valley saying, hey, you have to use this thing. Cultural retreat is not impossible.

The third intervention, and the paper’s preferred one, is maintaining a firewall between competent use and dependency. Preserve activities where humans do hard cognitive work even when a machine could do it faster. That’s what I’ve been calling cognitive fitness. We deliberately maintain physical fitness through diet and exercise. We should treat complex thought the same way: care about what information we take in, and exercise the brain. I like this because you can start today; you don’t need to wait for your employer, district, or a regulatory cavalry.

So, three rules. Use LLMs to remove tedium, not the friction of thinking. Let them handle mechanical tasks; don’t use them because forming the idea or drafting feels difficult. Second, don’t pass off machine-written prose as your own. If AI creates a meeting summary or memo, label it. When people read text they believe came from you, they’re engaging with a model of your mind; pretending otherwise is psychological fraud. Third, read human writing rather than AI compressions. That strain may be exactly the workout you need.

The paper has drawn criticism because its equations and parameters are not direct measurements of the real world. Fine. That is a limitation, but these models can reveal general dynamics and test changing conditions. I found the framework clarifying because the AI debate is usually a muddle of boosterism, anger, and disconnected fears. Different settings call for different responses. But when a mathematical model and actual life both land on preserving human cognition, that’s a signal worth taking seriously. All right. That’s enough complexity theory for one day. Stay deep.