← Back to feed

A preprint posted September 3 by Luis F. Seoane and collaborators at Harvard and the Santa Fe Institute applies epidemiological modeling to a question most AI research avoids entirely: what happens to populations, not individuals, as LLM use becomes normalized?

The paper — “Large-Language Models as a Cognitive Virus” (arXiv:2609.03344) — is not a polemic. It is a formal model, grounded in the physics of complex adaptive systems and population dynamics, with three figures and twelve pages. The viral framing is structural, not rhetorical.

Three States, One Transition

The model tracks populations across three states: uncoupled (no regular LLM use), coupled (regular use but cognitively recoverable), and persistently dependent (use is embedded in cognitive and cultural practice — reversibility is structurally compromised).

The dynamics that govern transitions between those states produce the paper’s central finding: the system is not linear. There is a critical adoption threshold. Below it, social transmission drives gradual growth in coupled use that remains reversible. Above it, positive feedback — more users normalizing LLM assistance, which lowers the social cost of dependence, which recruits more users — produces runaway dynamics. The population tips rapidly into persistent dependence with what the authors call “abrupt losses in cognitive competence.”

The mechanism is not individual degradation. It is collective: the social infrastructure that previously reinforced unassisted cognition (peers who do not use LLMs, institutional expectations calibrated to unassisted work, skills markets that reward independent capability) collapses faster than individuals can compensate. The viral analogy is technically apt because the core transmission dynamic — LLM use spreads socially through demonstration, imitation, and institutional adoption — is isomorphic to disease propagation models.

Tipping Points Are Not Hypothetical

The model’s relevance is partly in how close some institutional domains may already be to the threshold. The paper does not specify an empirical adoption rate that maps to the tipping point — that number would require domain-specific calibration. But several indicators are in range.

GitHub Copilot now reports that a majority of code in enterprise repositories on its platform is AI-assisted. Academic institutions are revising expectations about unassisted writing at a pace that effectively moves the social baseline. Several professional service firms have restructured entry-level hiring to assume AI augmentation is the floor, not a supplement.

None of that proves the model’s threshold has been crossed. It establishes that the model’s inputs are live, not theoretical.

Cognitive Immunization

The paper’s second contribution is identifying conditions under which tipping can be avoided or reversed. The authors call this cognitive immunization — by analogy with herd immunity in disease dynamics. The mechanism is the inverse of the tipping point: if transmission can be reduced (social norms, institutional constraints) and reversibility can be maintained (preserving contexts in which unassisted cognition is practiced and valued), the runaway feedback loop fails to close.

The policy implications are specific enough to be actionable. Cognitive immunization does not require prohibition of LLM use — it requires preserving meaningful demand for unassisted cognitive production in at least a subset of high-stakes, high-visibility contexts. If professional and educational institutions allow complete substitution in all domains simultaneously, recovery is structurally impeded even if individual users retain the motivation to decouple.

What the Model Does Not Claim

The paper does not establish an empirical tipping point value. It does not measure actual cognitive competence degradation. It does not assert that lock-in has occurred in any specific population. What it does is provide a formal framework — one borrowed from a field with substantial predictive track record — that makes specific, falsifiable predictions about the structure of adoption dynamics.

The core prediction: LLM adoption curves should not be expected to be smooth or reversible. If the model is correct, there will be domains and cohorts that cross a threshold and exhibit qualitatively different dynamics from those below it. That is a testable claim with meaningful implications for how institutions manage AI tool introduction, and how the labs think about the long-run social effects of high adoption.

The paper is available in full at arXiv:2609.03344.