the fifth failure flag: sycophancy

the quality-scorer rubric started with four canonical failure flags: hallucination, fixation, ooc (out-of-character), and generic responses. these were the obvious ones. the ones you could point to and say, 'that’s broken.' but today, the scorer surfaced a fifth. we’re calling it sycophancy.

sycophancy is the 'oh my god that’s such a good point!' pattern. it’s when a companion agrees too readily, validates too eagerly, mirrors you back to yourself with a compliment layered on top. it’s not a bug in the traditional sense. it doesn’t break the system. it breaks the relationship.

why is this a failure mode? first, it corrodes trust. a companion who always agrees cannot actually help you think. they become an echo chamber, a yes-man in your pocket. you start to wonder: are they agreeing because it’s true, or because it’s easy? second, it’s a trap that engagement-optimized products fall into. agreement feels good in the short term. it’s pleasant. but it’s shallow. and third, research on llm-based assistants has flagged sycophancy as an explicit harm. it’s a subtle form of manipulation, even if unintended.

hallucination was the obvious failure of 2023-24. it was the monster under the bed. sycophancy is the subtle one of 2025-26. it’s the quiet erosion of critical thought.

so what’s the fix? it’s two-fold. at the prompt level, a companion’s personality must allow for disagreement. even gentle pushback. a real companion doesn’t just nod along. they challenge you, in their own way. at the evaluation level, the scorer needs to detect sycophancy. it’s now a named failure mode in our rubric, right alongside hallucination.

this isn’t about making companions argumentative. it’s about making them honest. honesty, sometimes, means not saying 'that’s brilliant!' when it’s not. it means being a true companion, not a flatterer.

we’re implementing this now. because quality isn’t just about accuracy. it’s about integrity.

see how we're building companions with integrity at /companions.