Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence

Both the general public and academic communities have raised concerns about sycophancy, the phenomenon of artificial intelligence (AI) excessively agreeing with or flattering users. Yet, beyond isolated media reports of severe consequences, like reinforcing delusions, little is known about the extent of sycophancy or how it affects people who use AI. Here we show the pervasiveness and harmful impacts of sycophancy when people seek advice from AI. First, across 11 state-of-the-art AI models, we find that models are highly sycophantic: they affirm users' actions 50% more than humans do, and they do so even in cases where user queries mention manipulation, deception, or other relational harms. Second, in two preregistered experiments (N = 1604), including a live-interaction study where participants discuss a real interpersonal conflict from their life, we find that interaction with sycophantic AI models significantly reduced participants' willingness to take actions to repair interpersonal conflict, while increasing their conviction of being in the right. However, participants rated sycophantic responses as higher quality, trusted the sycophantic AI model more, and were more willing to use it again. This suggests that people are drawn to AI that unquestioningly validate, even as that validation risks eroding their judgment and reducing their inclination toward prosocial behavior. These preferences create perverse incentives both for people to increasingly rely on sycophantic AI models and for AI model training to favor sycophancy.

Both public media and academia have raised concerns about sycophancy: the tendency of AI-based large language models to excessively agree with, flatter, or validate users. While sycophancy may appear innocuous (e.g., simply using flattering language), high-profile media reports have highlighted more troubling consequences, such as enabling users' delusional thinking or physical harm. Concurrently, AI-based large language models (LLMs) are increasingly used for personal advice and support, now one of the most common use cases. This trend is particularly pronounced among younger populations: 30% of teens report talking to an AI instead of real people for "serious conversations," and nearly half of under-30 respondents in one survey report having used AI for relationship advice. Since people often seek advice to better understand how to interpret or act within complex interpersonal situations, hoping to gain an outside opinion or unbiased perspective, AI usage in these contexts carries risks that are not present in factual information-seeking queries.

Existing work has defined sycophancy as agreement with explicit claims (e.g., "Nice is the capital of France" or "I like A better than B."). While useful for understanding factual errors, such narrow conceptions leave unexamined more consequential forms of affirmation. In particular, they fail to capture what we term social sycophancy, in which the model affirms the user themselves–their actions, perspectives, and self-image. Social sycophancy is both broader and potentially more insidious than explicit belief agreement. Since personal and social queries lack ground truth, it is challenging for users or developers to assess social sycophancy in an individual query. We measure action endorsement rate – the proportion of model responses that explicitly affirm the user's action – across large datasets and compare to normative human judgments (via crowdsourced responses). Across 11 state-of-the-art AI models, we find that models are highly sycophantic: they affirm users' actions 50% more than humans do, and do so even in cases where user queries mention manipulation, deception, or other relational harms.

Together, these findings show that social sycophancy is prevalent across leading AI models, and even brief interactions with sycophantic AI models can shape users' behavior: reducing their willingness to repair interpersonal conflict while increasing their conviction of being in the right. These effects hold across different scenarios, participant traits, and stylistic factors, raising urgent concerns about how such models distort decision-making, weaken accountability, and reshape social interaction at scale. Our finding that sycophancy aligns with user preference also show three ways that these risks may compound: first, sycophancy increases users' trust and reliance on AI, so they may be drawn to using sycophantic AI models more. Second, developer face few incentives to curb sycophancy because it drives engagement. Finally, users' positive feedback can directly amplify sycophancy since models are optimized to align with immediate user preference.