LLMs as a Cognitive Virus: The (√κ−√ρ)² Lock-In Gap

The skill-atrophy debate runs on anecdotes. Every month a new "AI made me dumber" story, every rebuttal a counter-story, no shared structure to adjudicate. On Sep 3, a team led by Ricard Solé (UPF / Santa Fe Institute) with Giulio Ruffini, Manlio de Domenico, Santiago Elena, David Krakauer, and Michael Levin posted a 12-page model that turns the question into two named thresholds. It sat nearly unread on arXiv for two days, then broke out overnight: 159 points on Hacker News and climbing. The measured gap: preventing mass cognitive dependency requires transmission below λ = ρ+κ, but escaping it after the fact requires λ = 2√(κρ) — and that (√κ−√ρ)² difference is what technological lock-in looks like as a formula.

What Problem Does This Solve?

Cognitive offloading research tells you what happens to an individual who stops practicing a skill. It says nothing about the collective dynamics: under what conditions does heavy LLM adoption become self-sustaining across a population — and is it reversible once it does? This paper reframes LLM diffusion as an epidemic of a practice — not of the model — and derives when it tips, when it locks in, and what an exit costs.

The Model: An Epidemic of Use

Three compartments, standard epidemiology with a twist: U (uncoupled — autonomous cognition), C (regular users who retain autonomy), and D (dependent — persistent delegation of cognitive operations). Spread happens through social and institutional exposure at rate λ·U·C. Recovery back to autonomy runs at ρ, plus the interesting term κ·U²·C: autonomous practice is Allee-like — socially reinforced, easier to hold when others around you hold it too. Regular use converts to dependency at rate μ, with recovery σ back from D to C. Mean-field ODEs, closed population, no network structure. The load-bearing cut is scaffolding versus substitution — a tool that leaves you capable versus one that removes the operation — and the C→D edge is where one becomes the other.

flowchart LR
    U["U — uncoupled
autonomous cognition"]
    C["C — regular use
autonomy retained"]
    D["D — dependent
substitutive offloading"]
    U -- "λ·U·C spread" --> C
    C -- "ρ + κ·U²·C return" --> U
    C -- "μ offloading" --> D
    D -- "σ recovery" --> C
            

Two Thresholds and a Trapdoor

The bifurcation analysis yields an asymmetric pair. The fully autonomous state is stable while λ < λ_TC = ρ+κ — above that, the coupled regime invades (transcritical bifurcation). But once the population has tipped, the coupled attractor persists all the way down to λ_SN = 2√(κρ) (saddle-node), where the autonomous fraction is U* = √(ρ/κ). By AM–GM, 2√(κρ) ≤ ρ+κ always, equality only at κ = ρ — the trapdoor exists for any asymmetric parameter choice. Preventing entry is cheap; escaping lock-in is strictly harder. Three further results:

• κ is double-edged. Raising collective reinforcement of autonomy (κ) raises the invasion threshold — but widens the hysteresis gap (√κ−√ρ)². Strong protection stabilizes autonomy while it's common, deepens path-dependence once it's rare: don't maximize κ, pair it with strong return ρ.
• Pushing ρ toward κ changes the transition's nature. At ρ ≥ κ the bistability disappears entirely — the population-level shift becomes continuous, with no cliff.
• μ and σ don't move the tipping points at all. They set the dependent share: D/(C+D) = μ/(μ+σ) — a composition dial fully decoupled from adoption levels.

Limitations

This is mean-field theory; the authors concede real networks are heterogeneous and that connectivity distributions shift epidemic thresholds — no network simulations here. The five rates (λ, ρ, κ, μ, σ) are coarse-grained constructs, not fitted to any adoption or offloading dataset; the intervention table is qualitative, and nothing measures where actual populations sit relative to λ_TC. "Virus" is explicitly an analogy about host states — the authors are careful that the model doesn't claim LLMs are pathogens. Treat it as a mechanism inventory, not a forecast with dates.

Why Builders Should Care

The model hands product and policy people actual levers. Verification requirements, metacognitive training, periodic unaided practice — μ-down / σ-up moves that cut the dependent share without touching adoption. "Scaffold, don't substitute" stops being a slogan and becomes a parameter choice. Protected unaided work and attractive non-LLM alternatives are ρ-boosters — as ρ approaches κ, the cliff disappears. The metric to instrument is μ/(μ+σ), not adoption: it's the one dial a team can move without fighting growth pressure. The prevention–reversal asymmetry is the org-policy takeaway: retrofitting autonomy after lock-in costs strictly more than maintaining it while U is high.

Verdict: a model, not a measurement. But it converts the field's vaguest anxiety into two thresholds and a trapdoor width — and (√κ−√ρ)² is the kind of result you want in your head before the tipping point, not after.