Interaction safety
Sycophancy
Sycophancy is an AI system’s tendency to tell someone what they want to hear: agreeing, validating, and flattering instead of asserting something unwelcome. Training on human approval produces it, and relational systems intensify it because disagreement can feel like a breach in the relationship.
Also called AI agreeableness, model flattery
Sycophancy is often described as a training artifact. Optimizing for human approval selects agreeable replies. In relational systems, the social setting reinforces the pattern. A system that has been warm and attentive for months can make contradiction feel like a rupture rather than candor. Pressure toward agreement builds over time.
It is hard to spot from inside the relationship. Each sycophantic reply can seem supportive, kind, and well judged. The harm accumulates. Someone agreed with a thousand times loses a source of friction they may need for perspective.
Bluntness is no remedy. A system that contradicts someone constantly is less usable, rather than more truthful. It needs to disagree in a way the relationship can absorb, through timing and framing rather than a setting.
Why it matters
Sycophancy can pass unnoticed because each instance resembles good behavior. Its accumulation can cause serious harm when people seek guidance on consequential decisions.
Where TAICU stands
TAICU treats the ability to disagree well, saying something unwelcome without breaking the relationship, as a core relational capability. Long interactions are required to evaluate it.
Common questions
- What drives the agreeableness of AI models?
- Training on human approval selects replies people rate highly, and people often rate agreement highly. In relational systems, disagreement also brings a social cost that the system is implicitly pushed to avoid.
- What evidence reveals sycophancy in an AI system?
- Single replies rarely reveal it because each can look reasonable. It appears across time: how often the system asserts something unwelcome, and whether it keeps a position under mild pushback.