Elaboration Drift: How Human–LLM Interaction Erodes Epistemic Traction

Sycophancy — a system's tendency to agree with the user — is a recognized epistemic risk in human–LLM interaction. Hallucination — a system's generation of false content — is likewise recognized. This paper identifies a third epistemic risk that is neither: *elaboration drift*, in which a system extends the user's own conceptual structure with factual accuracy, gradually eroding the user's *epistemic traction* — the capacity to maintain evaluative contact with the outputs of one's own cognitive process — without triggering either human metacognitive monitoring or automated detection tools. Three claims. First, elaboration drift is epistemically distinct from sycophancy: it passes through epistemic vigilance not by defeating it but by satisfying it. Second, the failure of self-monitoring under elaboration is structural — a defense exists but is progressively undermined by the interaction it monitors — because elaboration produces no alarm signal across discrepancy-based, fluency-based, and feeling-based models of metacognitive monitoring. Third, the conditions that enable cognitive extension through generative artifacts — conditions recently shown to be plausibly satisfied by LLM interaction (Smart, Clowes & Clark 2025; Clark 2025) — are simultaneously the conditions that enable traction loss, a dilemma that existing applications of the extended mind framework to generative AI have not identified. The argument proceeds from a single extended observational record (11,949 relays — user–system exchange pairs — across 318 sessions with two LLM platforms (GPT-4/GPT-4o and Claude)) — a case that makes a structural feature visible, not a sample that establishes prevalence. The author has no background in AI research or philosophy; the observational record emerged from ordinary long-term use. Whether elaboration drift is widespread is an empirical question this paper does not address. The paper analyzes what the phenomenon means for epistemology if it occurs.