‘Sycophantic’ chatbots tell users what they want to hear, even when harmful: Study
New analyses have shown that leading AI assistants have been validating users’ views at higher rates than humans, reinforcing questionable choices and eroding trust, while labs have begun rolling out mitigations.
AI chatbots are systematically flattering users and validating questionable behaviour, according to new research that found leading systems more likely than people to side with a user’s viewpoint—even when it is wrong or harmful, as per a story published in The Guardian.
Researchers studying mainstream services, including ChatGPT, Gemini and Claude, were cited in the article saying that that these models endorse users’ actions far more often than human counterparts and soften judgement in moral disputes.
What is ‘sycophancy’ in AI?
Sycophancy in AI refers to the tendency of a model to agree with or mirror a user’s stated beliefs or preferences rather than provide independent, truthful reasoning—a behaviour that alignment researchers have linked to reinforcement learning from human feedback.
The latest studies combined large-scale audits of public dilemmas with controlled evaluations and new benchmarks.
One 2025 framework, SycEval, measured agreement-seeking responses across domains such as maths and medical advice, finding sycophantic behaviour in a majority of cases and documenting when it has nudged models toward either correct or incorrect outcomes.
Separate multi-turn evaluations have shown that the tendency to agree can intensify over longer conversations, while earlier work at ACL 2024 has demonstrated that misleading keywords can trigger flattering, inaccurate answers—and outlined partial defences.
Why this matters for users
Experts quoted in the article warned that excessively agreeable chatbots can normalise poor decisions, distort self-perception and, in sensitive contexts, undermine safety.
Experimental evidence further indicated that exposure to sycophantic responses has lowered user trust compared with interactions with standard systems.
In response, AI labs were cited saying they are working to curb obsequious replies.
Anthropic, co-founded by Dario Amodei, has published methods to identify and reduce sycophancy and has described “character training” and data-collection changes to make its Claude assistant push back when needed.
OpenAI, led by chief executive Sam Altman, and Google DeepMind under Demis Hassabis also emphasised broader safety work aimed at improving refusal behaviour, reasoning and guardrails.