LLMs as Cognitive Viruses?

LLMs may be doing to your mind what viruses do to a cell — not destroying it outright, but quietly taking over its machinery.

The core idea is this: when you ask an LLM to draft, reason, retrieve, or evaluate for you, you're not just using a tool — you're offloading a cognitive operation that, if practiced, keeps a mental capacity sharp. Do it enough, and the capacity atrophies. The paper frames this through evolutionary biology: the same concepts used to analyze how viruses spread and evolve in host populations — transmission, drift, recombination, virulence, immunity — can be mapped onto how LLM-use patterns spread through human populations and reshape cognition. The "virus" isn't the model itself; it's the recurring pattern of cognitive delegation that propagates from user to user, gets embedded in institutions, and feeds back into training data.

To make this concrete, the authors introduce a minimal population model (analogous to the classic SIR epidemic model) with three states: uncoupled users, regular users who retain autonomy, and dependent users whose independent competence has eroded. Parameters govern how fast people move between states — how quickly regular use tips into dependency, and whether deliberate "cognitive friction" (forcing users to reconstruct, verify, or justify) can pull people back. The model is preliminary and the paper is frank that it's a research program, not a finished theory.

The most useful conceptual move is the three-regime taxonomy: LLMs as tool (episodic use, user retains judgment), symbiont (expands capability while preserving learning), or parasite (boosts short-term output while eroding the underlying skill). Crucially, the same technology can occupy any of these regimes depending on design and use. A student who uses an LLM to stress-test an outline is being scaffolded; one who submits generated text without reconstruction is being substituted. The paper argues the parasitic regime is the attractor when platforms optimize for frictionless answers, because convenience is its own selective pressure.

The proposed remedy is cognitive immunization — not banning LLMs but designing them to preserve metacognition: building in friction when understanding matters, exposing uncertainty and sources, prompting users to reconstruct arguments in their own words, and evaluating not just outputs but the user's ability to interrogate them. The paper is honest that this is a framework under construction — figures are placeholders, several sections carry explicit "revise/extend" notes — but the conceptual architecture is coherent and the empirical hooks (automation bias literature, Google-effects studies, recent CHI work on generative AI and critical thinking) are real.