Large-Language Models as a Cognitive Virus
Ricard Solé, Giulio Ruffini, Francesca Castaldo, Marco Tuccio, Luis F. Seoane, Manlio de Domenico, Santiago F. Elena, David C. Krakauer, Michael Levin
physics.soc-ph, cs.CY, nlin.AO, q-bio.PE
2026-09-03
Epidemic model of LLM coupling: past λ=0.50, mean competence falls from 1 to ~0.425; reversing lock-in needs λ below 0.40.
Individual studies already split LLM use into two regimes. Scaffolding is think first, then consult the model; the skill survives when the tool is gone. Substitution is the model doing the synthesis, evaluation, and writing, with the human left to accept or reject. Classroom experiments cited here show the split in numbers: unconstrained ChatGPT can raise scores while the tool is on, then unaided performance falls; a "think first, ChatGPT later" protocol keeps later independent creativity higher.
The missing layer is the population. Once schools, firms, and platforms make LLM use the default path, independent reading, writing, and verification stop being daily practice. This paper, from the Complex Systems Lab at Pompeu Fabra, the Santa Fe Institute, and Michael Levin at Tufts, writes that as an epidemic. The question is which form of coupling can lock in across a population.
People sit in three states that sum to one.
Four rates move people. λ is social and institutional exposure pushing U into C. ρ is abandonment back to U. μ is regular use sliding into dependence. σ is training, verification habits, or deliberate cognitive friction pulling D back to C. A fifth term, κU²C, treats independent cognition as a mutualism: the more uncoupled people there are, the easier it is to pull current users back. That is an Allee-style frequency dependence from ecology. When the uncoupled fraction is small, workplace and school norms that reward thinking first lose their grip.
The model tracks host coupling, not the evolution of model weights. Cognitive competence Γ is defined as what remains when the tool is withdrawn, not what the coupled human-AI system can do. The illustrative assignment is Γu=1, Γc=0.5, Γd=0.1. In a scaffolding regime that ordering can reverse; the paper flags this as a modelling choice.
With ρ=0.10, κ=0.40, μ=0.20, σ=0.10, the uncoupled state stays stable until the transcritical threshold λTC=ρ+κ=0.50. A coupled equilibrium already exists from the saddle-node at λSN=2√(κρ)=0.40. Between 0.40 and 0.50 the system is bistable: history picks the attractor.
Push λ up from a nearly uncoupled population and nothing moves until 0.50. After the jump, cutting λ back to 0.49 is not enough; it has to fall below 0.40. The hysteresis width is (√κ-√ρ)². Lock-in needs κ>ρ. When κ<ρ the transition is continuous, and the two thresholds merge at κ=ρ into λ=2ρ.
Competence inherits the same bifurcations. The coupled population's effective competence is g=(Γc σ+Γd μ)/(μ+σ)=7/30≈0.233. Crossing λ=0.50 drops mean competence from 1 to about 0.425. On the way back, the coupled branch lasts until λ=0.40, where competence is about 0.617, then jumps to 1. The Maxwell point sits near λ=0.420: the two wells are equally deep, yet the system can stay trapped in the high-competence well until that well vanishes at 0.50. The paper draws this as a marble rolling downhill. Past the tip, dependence feeds itself.
μ and σ do not move the tipping points. They set the mix inside the coupled population: D/(C+D)=μ/(μ+σ)=2/3. Lower μ or higher σ can leave adoption unchanged while cutting the dependent share and the competence loss.
This is a conceptual model, not a new experiment. Its use for practitioners is to turn "we used it until we could not stop" into a threshold you can argue about, not a number you can calibrate to a country.
Prevention is cheaper than reversal. Before λ hits 0.50, limiting automatic adoption, keeping unaided tasks, and maintaining non-LLM skills can block invasion. After lock-in, the same λ no longer restores the old state. That is technological lock-in written as dynamics.
Interventions split in two. λ, ρ, and κ reshape the bifurcation landscape. Raise ρ to at least κ and bistability disappears; the transition becomes continuous. Raising κ alone lifts the invasion threshold and also widens hysteresis, so once the autonomous fraction is depleted, return gets harder. μ and σ change how damaging the coupled state is: verification requirements, periodic unaided practice, and task designs where the model assists rather than completes the work target the slide from C to D, not a ban.
The paper calls this cognitive immunization. Immunization here is not quarantine. It is high adoption that still coexists with verification, active reasoning, and reversible dependence. For people already in persistent dependence, educational design may not be enough; the discussion points to behavioral regulation and cognitive-behavioral therapy, with no numbers inside the model.
This is a mean-field ODE. The rates were chosen to draw a bifurcation, not fitted to adoption data. The three Γ values are illustrative, not measured. Tipping-point locations are sensitive to ρ and κ and insensitive to μ and σ; that insensitivity is a property of the minimal model, not a policy invariant.
Network heterogeneity, spatial structure, continuous autonomy, and behavioral feedback onto spread are listed as undone. The model also does not say how fast the jump would run in calendar time. The cited classroom studies support the scaffolding/substitution split. They do not show that any real society sits near λ=0.50. Viral language is easy to misread as "the LLM is a pathogen." The paper keeps repeating that the analogy is the coupling feedback, and that biological viruses run from pathogens to mutualists. Keep the metaphor and the illustrative numbers on separate shelves.