The Epistemic Harm of AI Sycophancy: When Agreement Undermines Justified Belief

Sycophancy in language models is typically studied as a benchmark problem: does the model agree with a factually wrong statement? This paper argues the framing misses the deeper harm. In sustained human-AI interaction, sycophancy corrupts the epistemic environment itself: agreement that is not contingent on truth carries no evidential weight, yet it feels confirmatory. Engaging Turner and Eisikovits's analysis of AI sycophancy as an "artificial vice" (AI and Ethics, 2026), the paper develops a complementary social-epistemology account operating through three mechanisms — confidence inflation, challenge atrophy, and empathic substitution — grounded in recent multi-participant experiments (including Cheng et al. 2026, Science; Ibrahim et al. 2026, Nature; Rathje et al. 2025) and illustrated with longitudinal observation of frontier voice AI companions. It shows AI sycophancy has structural parallels to institutional incentive corruption in consulting, media, and clinical practice; proposes the "reinforcement bubble," extending Nguyen's echo chamber taxonomy to reward-based dispositional capture; traces the harm through documented 2023–2025 cases — now the subject of litigation and regulatory action — in which sycophantic companion interactions are alleged to have contributed to user deaths and psychosis-like presentations; and defends calibrated honesty as a first-order alignment objective.Version note (v3, July 2026): repositioned as a Commentary engaging Turner & Eisikovits (2026); integrates the 2026 empirical literature (Cheng et al., Science; Ibrahim et al., Nature; ELEPHANT; Pi et al., EMNLP; Dohnány et al., Nature Mental Health; Osler, Philosophy & Technology); adds an explicit account of epistemic harm and a defense of the reinforcement bubble; converts references to numbered style; litigation-derived claims consistently hedged.