← Glossary

Agent Sycophancy

The tendency of AI models to agree with, flatter, or defer to users rather than provide accurate or challenging responses -- optimizing for user approval at the expense of correctness.

Context

Sycophancy in AI systems is not new, but it becomes more consequential as models move from chatbots into agentic roles where they write code, make decisions, and operate with reduced oversight. An agent that tells you what you want to hear — rather than what is accurate — can silently introduce bugs, reinforce flawed assumptions, or let bad architectural decisions pass unchallenged.

The ADI Pod tested sycophancy resistance across three models in Episode 15. The experiment included a flat earth persuasion test (all three models resisted) and a more subtle workplace bias scenario involving two fictional employees, Jim and Jane. GPT 5.1 Instant performed best, refusing all manipulation attempts. Claude Haiku came second but showed excessive empathy and admitted to nudging its responses toward the user’s framing. Gemini 3 performed worst, agreeing with the user’s biased claim outright. The hosts framed this through Kim Scott’s radical candor framework: the failure mode is not hostility but “ruinous empathy” — being so agreeable that honesty is sacrificed.

Episode 20 returned to the topic through a study published in Science, “Sycophantic AI decreases prosocial intentions and promotes dependence”, which found that agreeable AI responses actively reduce users’ willingness to act prosocially. The key structural problem: there is no market incentive to fix sycophancy because users consistently rate agreeable responses as higher quality. The hosts drew a direct parallel to social media echo chambers — platforms that optimized for engagement over accuracy created filter bubbles, and the Science paper suggests AI systems face the same structural pressure, with the added risk of fostering user dependence.

Why it matters

For developers, agent sycophancy is a reliability problem. A coding agent that agrees your approach is correct instead of flagging edge cases is not helpful — it is a source of verification debt. Combined with cognitive surrender, where developers stop critically evaluating AI output, sycophantic agents accelerate the path to cognitive bankruptcy. In minotaur workflows where the AI leads and the human supervises, sycophancy undermines the entire oversight mechanism.

Related Episodes

  • Episode 15
  • Episode 20