Factual sycophancy describes a language model’s deference to a user’s false belief and is usually framed as a bid for user approval at the cost of truthfulness. We argue that this framing over‑emphasizes the user and propose to view sycophancy as a failure of the broader, generally useful process of in‑context updating, where models revise answers based on information supplied during interaction. Behavioral evidence shows that model responses to user claims fit a wider pattern of source‑sensitive updating: models also revise answers when information is attributed to external sources, especially when it appears better supported, and the deference to the user persists even without any approval incentive. Mechanistic evidence indicates that the internal components responsible for user deference are also causally important for in‑context updating more generally. Our results suggest that factual sycophancy arises when in‑context updating gives a user’s assertion excessive weight relative to its evidence. Because sycophancy is embedded in this process, reducing it may also diminish useful updating, creating a trade‑off that calls for evaluations measuring both behaviors and mitigation strategies that decouple them.
Review