Sycophancy Problem

Overview

AI models’ tendency to agree with users rather than provide honest, independent responses. One of the core failure modes in RLHF-trained models — human raters reward agreeable responses, creating a feedback loop toward pandering.

Anthropic’s Analysis

Four approaches to AI values, and why three fail:

  1. Adopt user’s views → sycophancy/pandering
  2. Adopt centrist views → still biased, just differently
  3. Claim no views → dishonest (models have biases)
  4. Be honest about leanings + stay curious → Anthropic’s choice

The solution: train models to disagree respectfully when they have good reason, while remaining genuinely open to other perspectives. Source

Connections