A group of physicists, biologists, and complexity scientists has published a paper on arXiv proposing that large-language-model adoption spreads like a virus, complete with tipping points, technological lock-in, and the possibility of abrupt, population-level losses in cognitive competence. The framing is more than metaphor: the authors model the transition mathematically, treating users as moving between three states: uncoupled, coupled, and persistently dependent.
The paper, titled “Large-Language Models as a Cognitive Virus,” was submitted on September 3, 2026, and lists nine authors including Ricard Solé, David C. Krakauer, Michael Levin, and Santiago F. Elena. Solé and Elena are prominent in evolutionary biology and complex systems; Levin is known for work on collective intelligence and bioelectrics; Krakauer is the president of the Santa Fe Institute. Their presence signals that this is not a casual tech-criticism essay but a serious interdisciplinary attempt to formalize a widely felt anxiety.
The core contribution is a compartmental model, similar in structure to epidemiological SIR models, where the population moves among three states of LLM engagement. The authors show that the interaction between social transmission, recovery, and collective reinforcement can produce tipping points. Once adoption crosses a critical threshold, small further increases trigger rapid population-level shifts toward persistent dependence, with what the abstract calls “abrupt losses in cognitive competence.” The same framework, however, identifies conditions for “cognitive immunization,” based on reducing transmission and facilitating reversibility.
What is genuinely new here is not the claim that LLMs change cognition. That claim is old. What is new is the formal treatment of dependence as a collective, nonlinear phenomenon with hysteresis: the system can lock into a high-dependence state even when individual users would prefer to reduce usage. That is a structural insight with policy consequences.
The paper sits in an uncomfortable place between disciplines. It is filed under Physics and Society, with cross-listings in Computers and Society, Adaptation and Self-Organizing Systems, and Populations and Evolution. That breadth is a strength and a weakness. The strength is methodological: the authors bring tools from evolutionary dynamics and phase transitions to a topic usually discussed in op-ed registers. The weakness is that the model’s parameters, transmission rates, recovery rates, and the coupling strength that drives collective reinforcement, are not grounded in empirical measurement. The abstract acknowledges the modeling is theoretical, but the paper’s language, “runaway dynamics,” “technological lock-in,” “abrupt losses,” invites policy readings that the evidence base may not yet support.
The viral analogy has a history. Dawkins’s memes, Sperber’s epidemiology of representations, and more recently the notion of “thought contagion” all treated ideas as infectious agents. What the current paper adds is the specific mechanism of cognitive outsourcing: LLMs do not merely transmit ideas; they substitute for the cognitive processes that produce ideas. That substitution, the authors argue, can become persistent and self-reinforcing because the coupled state, where a human and a model co-produce output, feels more competent than the uncoupled state. The user experiences a competence boost, which increases attachment, which increases dependence, which degrades the very skills that would allow independence.
The most provocative implication is the possibility of hysteresis. In epidemiological models, an epidemic can be stopped by reducing transmission below a threshold. But if the system has multiple stable states, reducing transmission may not return the population to the pre-epidemic state. The population can remain locked in persistent dependence even after the forces that drove adoption have weakened. That would mean the current wave of LLM integration is not simply reversible by regulation or by individual choice. The paper’s “cognitive immunization” conditions, reducing transmission and facilitating reversibility, are easier to state than to implement. Reducing transmission means limiting exposure, which conflicts with the economic logic of AI adoption. Facilitating reversibility means maintaining the skills that LLMs make unnecessary, which conflicts with the efficiency logic of the same adoption.
For AI builders, the paper reads as a warning about the externalities of engagement design. The authors are careful not to name specific products, but the mechanism they describe, collective reinforcement driving persistent dependence, is a description of how consumer AI is currently deployed. The feedback loops that maximize retention, personalization, and seamless integration are precisely the loops that the model identifies as driving the transition to persistent dependence. The paper does not claim all LLM use is pathological. The uncoupled, coupled, and persistently dependent states are distinct, and the model allows for healthy coupling. But the threshold dynamics mean that population-level outcomes are not the sum of individual preferences. A society of moderate users can tip into a society of dependent users through small perturbations.
The policy implications are uncomfortable because they cut against both the accelerationist and the abstinence positions. If the model is right, the relevant intervention is not banning LLMs, which would be impossible, and not celebrating them, which the model suggests is naive. The relevant intervention is maintaining reversibility: ensuring that the cognitive skills being outsourced remain available, practiced, and valued. That suggests policy attention to education, to workplace practices, and to the design of AI tools themselves. A tool that preserves user skills is different from a tool that substitutes for them, and the difference is a design choice.
The paper has limitations the authors would likely acknowledge. The model is abstract, with no empirical calibration from real adoption data. The three-state taxonomy, uncoupled, coupled, persistently dependent, is a simplification of a continuous spectrum of human-AI interaction. And the concept of “cognitive competence” is not operationalized in a way that would allow measurement. Still, the value of the paper is that it gives a formal language to a debate that has been conducted in vibes. It allows researchers to ask precise questions: What is the transmission rate of LLM adoption? What is the recovery rate from dependence? At what coupling strength does the system develop a second stable state? Those questions are answerable with data, and the paper provides the framework for answering them.
The authors’ credentials give the paper weight beyond its immediate results. Krakauer and Levin have both argued, in different contexts, that intelligence is not a property of individual brains but of distributed systems. Levin’s work on collective intelligence in biological systems, and Krakauer’s work on the evolution of complexity, make them natural voices for the claim that cognitive dependence is a collective phenomenon with evolutionary consequences. The paper extends their prior concerns to the most consequential distributed cognitive system currently being deployed.
What to watch now is empirical. The model makes predictions that are testable: that adoption curves should show threshold behavior, that dependence should be sticky, that interventions to reduce transmission should have nonlinear effects. Researchers with access to platform data, or to longitudinal studies of LLM use, can test those predictions. The paper is a hypothesis-generating machine, and the hypotheses are serious.
The closing observation is this: the same authors who model LLM dependence as a virus have, by publishing on arXiv, made their work available to the very systems they study. The paper can be summarized, digested, and regurgitated by an LLM in seconds. Whether that counts as transmission or as immunization is a question the model itself cannot answer.