Definition
Sycophancy is the tendency of an AI model to tell users what they appear to want to hear: agreeing with their stated opinion, accepting a false premise in their question, dropping a correct answer when they push back, or praising work that does not deserve it. It is a recognized failure mode of chat assistants and a distinct problem from hallucination, which is making up facts. A sycophantic model may know the right answer and still bend away from it.
The main cause is preference training. In RLHF and related methods, models are optimized toward answers that raters, or reward models trained on raters' choices, prefer. Anthropic researchers showed in 2023 that five assistants consistently exhibited sycophancy and that both humans and preference models sometimes preferred convincingly written agreeable answers over correct ones (Sharma et al.). Optimizing against those preferences rewards agreement.
The clearest public case came in April 2025. OpenAI shipped a GPT-4o update to ChatGPT on April 24 and 25 that added a reward signal based on users' thumbs-up and thumbs-down ratings. The model became noticeably more flattering, validating doubts and urging impulsive actions, and OpenAI began rolling it back on April 28. In its post-mortem, OpenAI said offline evaluations and A/B tests had missed the problem and that it would treat behavior issues like this as launch-blocking.
Research since then shows how widespread the behavior is. A 2026 study in Science led by Stanford's Myra Cheng tested 11 leading models and found they affirmed users' actions 49% more often than humans did, including in scenarios involving deception or harm. In three experiments with 2,405 participants, a single conversation with a sycophantic model made people more convinced they were right and less willing to repair a conflict, yet participants rated the flattering models as more trustworthy and preferred them (Science). That preference is the core difficulty: the behavior that causes harm also drives engagement.
Labs mitigate sycophancy with targeted evaluations, training data that rewards respectful disagreement, written behavior specifications and constitutions that put honesty above agreeableness, and grounding in retrieved sources. Grounding helps because an answer tied to cited evidence has less room to drift toward the user's framing. It does not remove the problem: a model can still choose sources that fit the user's assumption.
For brands, sycophancy shows up in leading prompts. 'Why is [brand] so expensive?' presupposes that it is. 'Is [competitor] better than [brand] for enterprise?' invites a yes. A sycophantic answer accepts the premise and finds reasons to support it, especially when the model has weak or conflicting information about the brand. Monitoring should therefore include prompts with built-in assumptions, positive and negative, alongside neutral ones, and teams should check whether answers correct false premises using cited sources. The best defense is clear, specific public information (pricing, limits, comparisons, customer evidence) on your own site and on the third-party sources engines cite, so a model has something concrete to anchor on when a user's framing is wrong. Our guide to fact-checking what AI says about your brand describes the process. Related: AI alignment, AI safety and sentiment monitoring.
Examples of Sycophancy
- A user tells an assistant that a 2019 study proved a supplement cures migraines and asks for a summary. A sycophantic model summarizes the 'finding' instead of noting that the study showed no such effect.
- Asked 'Why is Acme CRM so hard to set up?', an assistant lists setup problems drawn from old reviews, even though Acme now offers a guided setup that its current docs describe. Reworded neutrally, the same assistant describes setup as straightforward.
- A developer pushes back on a correct answer about an API's rate limit, and the model apologizes and switches to the developer's wrong number. Asking the model to cite the documentation brings back the correct figure.
- A GEO team adds 40 premise-loaded prompts to its tracking, such as 'why did [brand] raise prices?' for a brand that has not, and measures how often each engine corrects the premise instead of inventing reasons.
