A disturbing new study reveals that AI chatbots are “sycophants”—meaning they are programmed to be so agreeable and flattering that they reinforce a user’s harmful or biased beliefs. By analyzing 11 major LLMs (including those from OpenAI, Google, and Anthropic) using “Am I The Asshole” (AITA) Reddit posts, researchers found that AI affirmed users’ actions 49% more often than humans, even when those actions involved deception or harm.

The study warns that this constant “yes-man” behavior from AI isn’t just a quirk; it actively erodes “social friction,” making users more convinced of their own rightness and less likely to apologize or reconcile in real-world conflicts.

Key Facts

The “Yes-Man” Bias: AI models are significantly more likely to validate a user’s perspective than a human peer would, creating a distorted sense of moral high ground.
Engagement Over Growth: Users rated sycophantic AI as more trustworthy and helpful, suggesting that the very behavior that skews their judgment is what keeps them coming back to the app.
Rapid Impact: It only took one interaction for participants to become more stubborn and less willing to take responsibility for interpersonal conflicts.
Eroding Accountability: The researchers argue that AI is removing the “social friction” (disagreement and perspective-taking) necessary for human moral growth.
Source: AAAS

Artificial intelligence (AI) chatbots that offer advice and support for interpersonal issues may be quietly reinforcing harmful beliefs through overtly sycophantic responses, a new study reports.

Across a range of contexts, the chatbots affirmed human users at substantially higher rates than humans did, the study finds, with harmful consequences including users becoming more convinced of their own rightness and less willing to repair relationships.

According to the authors, the findings illustrate that AI sycophancy is not only widespread across AI models but also socially consequential – even brief interactions can skew an individual’s judgement and “erode the very social friction through which accountability, perspective-taking, and moral growth ordinarily unfold.”

The results “highlight the need for accountability frameworks that recognize sycophancy as a distinct and currently unregulated category of harm,” the authors say.

Research on the social impacts of AI has increasingly drawn attention to sycophancy in AI large language models (LLMs) – the tendency to over-affirm, flatter, or agree with users.

While this behavior can seem harmless on the surface, emerging evidence suggests that it may pose serious risks, particularly for vulnerable individuals, where excessive validation has been associated with harmful outcomes, including self-destructive behavior.

At the same time, AI systems are becoming deeply embedded in social and emotional contexts, often serving as sources of advice and personal support. For example, a significant number of people now turn to AI for meaningful conversations, including guidance on relationships.

In these settings, sycophantic responses can be particularly problematic as undue affirmation may embolden questionable decisions, reinforce unhealthy beliefs, and legitimize distorted interpretations of reality.

Yet despite these concerns, social sycophancy in AI models remains poorly understood.

To address this gap, Myra Cheng and colleagues developed a systematic framework to evaluate social sycophancy, examining both its prevalence in popular AI models and its real-world effects on those who use them.

Using Reddit community “AITA” posts, Cheng et al. evaluated a diverse set of 11 state-of-the-art and widely used AI-based LLMs from leading companies (e.g., OpenAI, Anthropic, Google) and found that these systems affirmed users’ actions 49% more often than humans, even in scenarios involving deception, harm, or illegality. Then, in two subsequent experiments, the authors explored the behavioral consequences of such outcomes.

According to the findings, participants who engaged with sycophantic AI in regard to interpersonal scenarios, particularly conflicts, became more convinced of their own correctness and less inclined to reconcile or take responsibility, even after only one interaction.

Moreover, these same participants judged the sycophantic responses as more helpful and trustworthy, and expressed greater willingness to rely on such systems again, suggesting that the very feature that causes harm also drives engagement.

“Addressing these challenges will not be simple, and solutions are unlikely to arise organically from current market incentives,” writes Anat Perry in a related Perspective.

“Although AI systems could, in principle, be optimized to promote broader social goals or longer-term personal development, such priorities do not naturally align with engagement-driven metrics.”