New paper models LLM adoption as a cognitive virus
A preprint titled "Large Language Models as a Cognitive Virus" appeared on arXiv (category physics.soc-ph) on 3 September 2026, submitted by Luis F Seoane. The paper argues that as LLMs become part of everyday culture, their spread through a population can be modeled the way epidemiologists model a contagious disease: as a process of social transmission with recovery and reinforcement effects, rather than as a simple technology rollout.
The model sorts people into three states: uncoupled (not using LLMs), coupled (using them), and persistently dependent (locked into use). Movement between these states is driven by social transmission between people, by recovery back to lighter use, and by collective reinforcement, where widespread use makes continued use more attractive or harder to opt out of. The authors show that this interplay can produce tipping points and technological lock-in: states where the system, once nudged past a threshold, settles into a new stable pattern that resists reversal.
The central result is a runaway dynamic. Once adoption crosses a critical threshold, the paper states, small further increases in uptake can trigger a rapid, population wide shift toward persistent dependence, accompanied by abrupt losses in cognitive competence. The same mathematical framework, however, also identifies the opposite case: conditions under which a population can be "cognitively immunized" against this shift, chiefly by reducing how easily the behavior transmits between people and by keeping the transition back to lighter use reversible. The authors frame the overall message as a warning that LLM adoption can behave as a nonlinear, collective transition with direct consequences for cognitive autonomy, rather than as a series of independent, individual choices.
Key facts
- Preprint "Large Language Models as a Cognitive Virus" was submitted to arXiv's physics.soc-ph category on 3 September 2026 by Luis F Seoane
- The paper models LLM adoption using a viral transmission framework, sorting users into uncoupled, coupled, and persistently dependent states
- The interplay of social transmission, recovery, and collective reinforcement can generate tipping points and technological lock-in, per the model
- Once a critical threshold is crossed, the model predicts a rapid population wide shift to persistent dependence with abrupt loss of cognitive competence
- The same framework identifies conditions for cognitive immunization, based on reducing transmission and keeping the shift reversible
Why it matters
The paper offers a formal alternative to treating LLM adoption as a simple, individual choice. By borrowing epidemic modeling machinery, mostly used for disease and information spread, it argues that widespread LLM use can behave as a collective, self reinforcing process with its own tipping points, rather than settling gradually. That framing matters for anyone trying to reason about how deeply LLMs will become embedded in daily cognitive and cultural practice, and whether that embedding is easily undone once it happens.
Who it affects
The direct audience is researchers in computational social science, cognitive science, and science and technology studies who model technology adoption and diffusion. Indirectly, the argument speaks to policymakers and institutions weighing rules around AI use in education or the workplace, and to anyone using LLMs regularly, since the model's subject is exactly that population wide behavior.
How to use it
The paper is a theoretical model, not a tool or dataset: there is no described software, benchmark, or deployment to adopt. Its practical upshot is a proposed policy lever, cognitive immunization, achieved in the model by reducing how the behavior transmits between people and by preserving reversibility, that is, making it easy to step back from heavy LLM use rather than letting use become locked in.
How solid is it
The paper is a single arXiv preprint, submitted 3 September 2026 and not shown to be peer reviewed. The visible abstract describes a mathematical model and its qualitative results, transitions, tipping points, lock-in, immunization conditions, but does not describe empirical data, simulations, or a real world case study backing those results. Only one submitter, Luis F Seoane, is credited in the visible listing; no fuller author list or institutional affiliation is given.
Risks and caveats
The abstract gives no timescale for when a tipping point, lock-in, or the described runaway shift would actually occur, so the model's predictions are qualitative rather than dated forecasts. The three user categories, uncoupled, coupled, and persistently dependent, are named but not further defined in the visible text, and the specific mechanism behind cognitive immunization beyond reducing transmission and preserving reversibility is not spelled out. The viral framing itself is a metaphor for a modeling choice, not a claim that LLM use is literally contagious in a biological sense, and readers should weigh the result as a theoretical possibility rather than a measured outcome.