British Columbia Sues OpenAI Over School Shooting Warning

The announcement reads as a province holding a Silicon Valley company accountable for the unthinkable: British Columbia has filed suit against OpenAI and Sam Altman in San Francisco federal court, alleging the company sat on knowledge of a shooter’s planning conversations with ChatGPT and said nothing to the people who could have stopped him. One notices that the lawsuit is not actually about the software failing to detect a threat. It is about what the software was built to do when it did detect one: nothing, by design, because a “warn law enforcement” pipeline is a feature nobody shipped.

Here is the detail everyone in the room is stepping around. ChatGPT has no channel to a dispatcher. It has no case number, no jurisdiction, no officer on the other end of a modeled decision. The product that generated the conversation was never wired to any system capable of receiving an alert, because building that wire means deciding, in advance and at scale, which of several hundred million daily conversations about violence - fictional, venting, research, roleplay, genuine - gets escalated to an actual human with a badge. Nobody has built that filter because the false-positive rate on “escalate this to police” is not a UX problem, it is a liberty problem, and OpenAI knows it, and British Columbia’s complaint is functionally asking a court to invent the filter after the fact by way of damages.

This is the plain question the filing has to answer and hasn’t: what would the warning have said, to whom, and what would that recipient have been legally able to do with it before the fact? Not after - a warning issued after the fact is a police report, which is a different document entirely, filed by different people, and useless to the province’s actual claim. The complaint wants a world where a conversation log becomes an actionable tip in time to prevent the act it describes. That world requires OpenAI to run something closer to a fusion center than a chatbot company, staffed by people making judgment calls about intent versus fantasy, at a volume no fusion center on earth has ever handled, with a false-positive cost measured in wrongful raids and a false-negative cost measured in exactly what British Columbia is now suing over.

The second angle is the one the framing wants left at the edge: this is a products-liability theory wearing a duty-to-warn coat. Duty to warn, historically, attaches to a treating clinician who has a specific patient in front of him and a specific threat named to him - Tarasoff, the case every first-year law student meets, involved a therapist who was told, by name, who the intended victim was. A general-purpose language model processing an anonymous session is not a clinician with a patient file. It is closer to a phone company that carried a call. The province’s lawyers know this distinction exists and are suing anyway, which tells you the real target isn’t the narrow legal theory - it’s the settlement, or the precedent that a platform’s scale itself creates the duty, regardless of whether the platform has any mechanism to discharge it.

Somebody at OpenAI’s trust and safety desk read every transcript this complaint will eventually quote into the public record, and did so after the fact, in the specific mode of a person reconstructing a tragedy rather than preventing one - that is the actual condition of the job today, industry-wide, and it will remain the condition until someone funds and defends the surveillance infrastructure the province’s own remedy implicitly demands. Nobody wants to be the company that says out loud “yes, we read your prompts for signs you might hurt someone,” and nobody wants to be the company that didn’t.

Boter bij de vis - butter with the fish, pay as you go, no credit - is the Dutch way of saying you don’t get to demand the outcome without funding the mechanism. British Columbia wants the outcome. The mechanism it’s asking a jury to retroactively mandate is a real-time threat-detection and law-enforcement liaison system inside a text predictor, and nobody arguing this case in San Francisco has designed one, because the last group to try that kind of population-scale pre-crime filtering called it something else and got sued for that too.