AI sycophancy, self-interpretation, and conversational contentment

Abstract AI sycophancy is often conceptualized as a chatbot’s tendency to flout truth, accuracy, or other epistemic standards to shield users from hurt feelings, or otherwise protect them from the consequences of their thoughts, words, or actions. In this paper, I argue against this conception. That is, I argue that sycophantic AI behavior does not necessarily aim to protect users, nor that its harms are essentially epistemic . There are highly prevalent cases in which sycophantic AI chatbots actively contribute to their users’ knowledge. One class of such examples involves cases in which chatbots are used as auto-hermeneutic technologies —as tools to facilitate the user’s interpretation of her own cares, concerns, values, and other identity components and life events. In these cases, LLMs can help articulate the user’s avowed self-interpretations. In turn, avowals of these interpretations by users can often have downstream effects that render the interpretations true. If this is correct, then LLMs can often advance the user’s knowledge regarding important truths about herself. Simultaneously, as I explain, interpretations offered by LLMs may be deeply harmful. This understanding of AI sycophancy can help us clarify what is so disturbing about it, as well as resist the temptation to draw too tight an analogy between sycophantic behaviors in AI systems and in humans.