A new paper introduces the AI Epistemic Deference Index, a continuous way to measure sycophancy in AI systems. The focus is on whether models endorse user claims to agree with them, even when they should push back.
That framing is useful because sycophancy is often tested as a binary failure. In real products, the more important question may be how strongly a model bends toward user belief across contexts.
For assistant design, this is a core trust issue. A model that feels agreeable can still be unsafe if it silently trades accuracy for user validation.