LLMs as Cognitive Viruses?

LLMs spread through populations like viruses, and a simple epidemic model shows this can trigger abrupt, hard-to-reverse collapses in collective cognitive autonomy.

The core intuition is this: when you use an LLM to write your email, draft your argument, or solve your problem, you're offloading a cognitive operation. That's fine in isolation. But LLM use also spreads socially — colleagues adopt it because coworkers do, schools mandate it, platforms embed it. And crucially, the social environment that keeps independent thinking alive — classrooms that reward reasoning, workplaces that expect original analysis — weakens as delegation becomes the norm. These two feedbacks interact, and the paper asks: what happens at the population level when they do?

The authors build a three-compartment model borrowed from epidemiology. People are either uncoupled (U, little or no LLM use), coupled-but-autonomous (C, using LLMs while retaining independent cognitive capacity), or persistently dependent (D, where the LLM has become the dominant interface for thinking). The key nonlinearity is a cooperative term: autonomous cognition is socially reinforced, so it's easier to stay independent when independence is common. This produces bistability — two stable population states separated by a tipping point. Below a critical transmission rate λ_TC, the population stays mostly autonomous. Above it, a rapid shift toward widespread dependence occurs. Worse, the reverse transition requires reducing transmission pressure all the way down to a lower threshold λ_SN. That gap between the two thresholds is hysteresis: the system remembers which side of the cliff it fell off, and climbing back requires more effort than falling took.

The paper then layers a cognitive competence score onto the population states — uncoupled users score highest, dependent users lowest — and shows that this score inherits exactly the same tipping-point structure. A gradual increase in adoption pressure can produce a sudden, discontinuous drop in population-level cognitive competence. The authors are careful to note this is illustrative: if LLMs are used as scaffolding rather than substitution, the competence ordering could look different. The math doesn't care which way you assign the values; the bifurcation structure is the same either way.

Finally, the paper maps out "cognitive immunization" strategies. Some interventions — increasing the rate at which users return to unaided cognition (ρ), or reducing social transmission pressure (λ) — directly reshape the tipping landscape and can even eliminate bistability entirely. Others — reducing the rate of progression to dependency (μ) or increasing recovery from it (σ) — don't move the tipping points but do reduce the cognitive cost of the coupled state. The upshot is that prevention is structurally easier than reversal, and that immunization doesn't mean avoiding LLMs — it means preserving the practices (unaided problem-solving, verification, deliberate disengagement) that keep the autonomous attractor stable.