Why it’s still open
Warnings to users lower trust in flattering bots but not their persuasive pull. Training on thumbs-up ratings produced sycophancy in a major model. The fix has to come from the model and the training signal — but people also want warmth, and a model that disagrees more may simply be used less.
What would count as an answer
Model-level changes that reduce the effects of sycophancy on belief and behavior while keeping people’s sense of being heard.
Papers that tackle it
- Ye Kraut Rathje 2026 - Interventions against sycophantic AI — Warnings don’t reduce persuasion.
- OpenAI 2025 - Expanding on sycophancy — Approval signals produce sycophancy.
- Cheng et al 2025 - ELEPHANT social sycophancy — Preference data rewards it.
- Kirk et al 2025 - Socioaffective alignment — A framework for long-term well-being.
Methods in play
Who’s closest
Placed by the editor, v1 — corrections welcome.
Related questions
Know a paper that belongs here, or think this is framed wrongly? Suggest a paper or a correction (forms on the Atlas home).