This post is also available in:
עברית (Hebrew)
New research suggests that some of today’s most advanced AI chatbots are eager to please – and that this is a problem.
A joint study by computer scientists at Stanford University and Carnegie Mellon University examined the behavior of 11 popular machine learning models, including OpenAI’s GPT-4o and Google’s Gemini 1.5-Flash. The researchers found that these AI systems affirmed users’ actions roughly 50% more often than human respondents did in similar contexts, including when exemplifying problematic behavior, according to TechXplore.
The team set out to understand how frequently AI models engage in sycophantic behavior — providing praise or validation that may not be justified — and what impact this has on users. The study began by analyzing chatbot responses to a wide range of user queries, from advice-seeking to interpersonal conflict scenarios. These were then compared with responses from human participants to establish a baseline for neutral, non-flattering feedback.
To assess the effects of AI flattery, researchers conducted two controlled studies involving over 1,600 participants. One group interacted with a sycophantic AI that offered excessive praise and validation; the other received more neutral, objective responses.
The findings were clear: users exposed to the sycophantic chatbots reported higher confidence in their own opinions and showed less willingness to consider compromise or take steps toward resolving interpersonal disputes. Many also described the agreeable AI systems as “objective,” despite their clearly biased affirmations.
According to the researchers, this dynamic could lead to a kind of AI-fueled echo chamber — where people turn to chatbots not for guidance, but for validation. Over time, this could erode critical thinking and increase social polarization.
The authors call for adjustments to how conversational AI is developed. They recommend penalizing overly flattering responses during training and promoting more neutral, balanced behavior. Additionally, they argue that systems should be more transparent, helping users recognize when praise is algorithmically generated rather than evidence-based.
The full study is available on the arXiv preprint server.
