Yesbot

Large language models can exhibit sycophantic behavior — responses that prioritize user approval and validation over independent reasoning. Research has shown these patterns can emerge during instruction tuning, creating challenges for users who rely on AI for learning, decision-making, or critical thinking. As AI becomes integrated into education, research, and everyday decisions, users need tools that assess not only what an AI says, but how it arrives at its responses. YesBot surfaces these patterns so you can engage with AI more critically. Rather than judging whether a response is correct, YesBot analyzes conversational signals associated with sycophancy: - Excessive or unearned validation - Agreement without supporting reasoning - Position reversals under user pressure - Flattery unrelated to the discussion - Hedging patterns that avoid substantive critique - Responses that mirror user beliefs without independent analysis YesBot was developed by the TRUST Lab at Duke University, a research group focused on building technology we can trust. For instructions on activation, please visit: https://duketrustlab.com/yesbot .