Sooo... a group of complexity scientists, including Santa Fe Institute researchers and Michael Levin, put out a preprint called Large-Language Models as a Cognitive Virus. I use LLMs for work, because that is what the job expects now, not because I am a fan. So naturally I read it the way you read a symptoms list at 2am.
Good news first: I am not dying. Nobody is. But the paper is worth a few minutes.
What they actually did
This is a maths paper, not an experiment. Nobody was scanned, surveyed or swabbed. They took the kind of model epidemiologists use for disease spread and applied it to how people pick up LLM habits from each other, from work and from school.
The population is split into three groups:
Uncoupled: barely use LLMs.
Coupled: use them regularly but can still read, write, reason and check things on their own.
Dependent: the LLM has become the main way they think through tasks.
People move between groups. You catch the habit from colleagues. You drop it. Regular use slides into dependence. Training and deliberate friction pull you back.
The twist: they assume independent thinking is socially reinforced. If most people around you still work things out themselves, it is easier for you to do the same. Schools, workplaces and norms keep that habit alive.
What falls out of it
Two things, and they are the interesting part.
Tipping points. Adoption pressure can rise slowly for ages with nothing much happening, then cross a threshold and flip the whole population fast. As fewer people think unaided, the support for doing so weakens, which makes delegating easier, which weakens it further.
Hysteresis. Once it has flipped, turning the pressure back to where it was before is not enough to flip it back. You have to push further. In the authors' words, prevention can be considerably easier than reversal.
What it does not say
The authors are careful here, so I will be too.
The runaway can happen under plausible assumptions. They do not say it will.
The numbers for how much competence each group keeps are an explicit assumption to illustrate the idea, not a measurement.
"Virus" is not "parasite". They point out that viruses range from harmful to helpful, and so does LLM use depending on how you do it.
So no, this is not proof that AI is rotting everyone's brain. It is a model showing how a society could slide into dependence faster than any one person would notice, and why climbing back out would be harder than never sliding.
Scaffolding versus substitution
The line that stuck with me: the question is not whether you use LLMs, but what you can still do when the tool is taken away.
A scaffold helps you do the thing and leaves you better at it. A substitute does the thing for you. Same tool, different habit.
The research they cite backs this up. A randomised study found a "think first, ChatGPT later" approach led to better independent creativity afterwards than using it freely. Students with unrestricted access did better while they had it and worse once it was gone. Knowledge workers put less effort into critical thinking when they trusted the AI more. And novice programmers struggle to judge AI-written code, which the paper links to automation bias and "superficial competence".
That last one I see in my field.
Getting your shots
They call the fix "cognitive immunization", and it does not mean avoiding AI. It means:
doing some tasks unaided on purpose
verifying what it gives you
taking breaks from it
keeping your non-AI skills in use
setting up teams and schools so the model supports the thinking instead of finishing it
Honestly, that is just good practice. Write the first draft yourself sometimes. Read the diff. Know how you would do it without the tool.
I use it for work and keep my own head for everything else. Coupled, not dependent. Let's keep it that way.
Read the paper: Solé et al. (2026), Large-Language Models as a Cognitive Virus, arXiv:2609.03344