A new paper argues that the way we adopt large language models looks a lot like the way a virus spreads through a population, and once enough people cross a certain line, the shift toward dependence can happen fast. The study, titled “Large-Language Models as a Cognitive Virus” and led by Ricard Solé along with eight co-authors, was submitted on September 3, 2026 and is drawing attention on Hacker News, where it climbed to 183 points in the Research category. What stands out here is that the researchers aren’t talking about AI safety in the usual sense. They’re talking about your brain.
The core finding
The team treats LLM adoption as a contagion. People move between three states: uncoupled (you don’t lean on AI), coupled (you use it regularly), and persistently dependent (you can’t work without it). The interesting part is what happens when social pressure, habit, and group reinforcement stack up together.
According to Hacker News, the paper’s central warning is about “runaway dynamics.” Cross a critical threshold, and small bumps in adoption trigger a rapid population-wide slide into dependence, paired with what the authors call “abrupt losses in cognitive competence.” In plain terms: past a tipping point, a lot of people can lose skills quickly and at the same time.
How they studied it
The researchers built a mathematical model borrowed from epidemiology, the same math used to track how diseases move through communities. Instead of infection and recovery, they modeled three forces:
- Social transmission: you adopt AI because the people around you do.
- Recovery: you step back and regain independent habits.
- Collective reinforcement: the more a group depends on AI, the harder it is for any one person to opt out.
When those forces interact, the model produces tipping points and “technological lock-in.” That’s the state where reversing course gets extremely hard because the whole system has settled into dependence.
Why it matters
This is significant because it reframes the AI skills debate as a collective problem, not an individual one. You might personally decide to keep your writing or coding sharp, but if your team, your industry, and your tools all assume AI dependence, the model says you get pulled along anyway.
For practitioners, the practical takeaway is the flip side of the warning. The same framework points to what the authors call “cognitive immunization.” Two levers do the work:
- Reduce transmission: slow the automatic, everyone-does-it spread of dependence.
- Facilitate reversibility: keep it easy to work without AI, so people can move back toward independence instead of locking in.
In practice, that looks like keeping non-AI workflows alive, building in regular practice without the tool, and designing team norms that treat AI as optional rather than mandatory. The model suggests those choices matter most before you hit the threshold, not after.
A note on the limits
This is a modeling paper, and that’s worth keeping in mind. The authors describe LLM adoption as involving “nonlinear collective transitions,” but the framework is an analogy built on equations, not a measurement of real-world skill loss across actual populations. The paper identifies conditions where tipping points and competence drops could happen. It doesn’t prove they already have, or pin down exactly where the threshold sits.
Still, the value here is the shape of the problem. Framing AI dependence as a contagion with tipping points gives teams and policymakers a vocabulary for a risk that’s usually discussed in vague terms.
The bigger question the paper leaves open is measurement. If cognitive dependence really behaves like a virus with a critical threshold, the next step is figuring out how close any given field already sits to it. Expect this framing to show up in the wider debate about how much thinking we hand off to machines. Full details are available in the original paper via Hacker News.