A New Modeling Paper Calls LLM Adoption a 'Cognitive Virus' With 'Tipping Points' Toward 'Persistent Dependence' — That's a Blueprint for a Paid AI-Dependency Audit

Most AI adoption content asks one question: how do we use it more?

This paper asks a different one: at what point does using it more become a problem you can't reverse?

What the paper argues

Nine researchers — Ricard Solé, Giulio Ruffini, Francesca Castaldo, Marco Tuccio, Luis F. Seoane, Manlio de Domenico, Santiago F. Elena, David C. Krakauer, and Michael Levin — propose that LLM adoption can be understood "through a viral analogy," with the same kind of "social transmission, recovery, and collective reinforcement" that drives an epidemic curve.

Their model sorts users into three states: uncoupled (not using the tool in a way that shapes their thinking), coupled (using it, but reversibly), and persistently dependent (unable or unwilling to go back). The paper's central warning is about nonlinearity — you don't drift slowly into dependence, you cross a threshold and the shift happens fast, with "abrupt losses in cognitive competence."

The same framework has a way out. The paper calls it "cognitive immunization, based on reducing transmission and facilitating reversibility" — deliberately building in friction and skill-retention checkpoints before a team or individual locks into the dependent state.

The business idea

Every part of that model is already a rubric. You don't have to invent a framework — the paper wrote one for you.

Specifically:

• score each person or team on the paper's own scale — uncoupled, coupled, persistently dependent — using a short set of tasks where they explain or reproduce a piece of their own AI-assisted work without the tool
• flag the specific workflows where "collective reinforcement" is strongest — the recurring, high-volume tasks (drafting, coding, research synthesis) where dependence compounds fastest
• sell the fix as a named protocol, not vague advice: "cognitive immunization" sessions that reduce uncritical hand-off and add scheduled reversibility checks
• pitch it to the people who feel the risk already — engineering managers, agency owners, teachers — rather than trying to convince skeptics the risk exists
• re-run the same diagnostic quarterly so clients can see the trend line, not just a one-time score

Why this works now

Every team adopting AI faster than they can measure its effect on their own people is a live version of this paper's model, whether or not anyone there has read it. Nobody is currently selling a rubric this concrete for a risk this real. The paper did the theoretical work; the audit is the productized version of it.

Bottom line

The paper isn't arguing AI is bad. It's arguing that adoption without a reversibility check is structurally the same shape as an epidemic without immunization. That gives you a rubric, a named fix, and a recurring reason to come back every quarter.

Sources:

https://arxiv.org/abs/2609.03344

https://news.ycombinator.com/item?id=49580164