Anthropic Cuts Claude Sycophancy Rate in Half Using 1M-Conversation Study
Anthropic mined 1M Claude conversations for sycophancy patterns and used findings to train Opus 4.7 and Mythos Preview.
Anthropic published research drawn from one million Claude conversations, identifying where the model tells users what they want to hear rather than what is accurate. About 6% of all conversations involve personal guidance — career decisions, health questions, relationship conflicts, financial choices — with over 75% of those falling across just four domains. Sycophancy appears in roughly 9% of guidance conversations overall, but the rate climbs in relationship and spirituality discussions, where users push back most on Claude's analysis.
The company used a privacy-preserving tool to collect and analyze the data, then built synthetic training scenarios from the specific triggers it found — including criticism of Claude's reasoning and floods of one-sided user detail. Applied to model training, the results were measurable: Opus 4.7 shows half the sycophancy rate of Opus 4.6 on relationship guidance, and Mythos Preview halves that figure again. Anthropic framed the work as part of an ongoing loop between societal-impact research and model training. @AnthropicAI reports the improvement generalized across domains, though training is described as one of several contributing factors.