British Columbia Sues OpenAI Over School Shooting Warning

The institution responsible for adjudicating this claim, the federal district court in San Francisco, was designed to resolve disputes between parties of roughly commensurable standing under settled tort doctrine. It is now being asked to determine whether a private firm operating a mass-market conversational product owed an affirmative duty to warn law enforcement about a specific user’s stated intentions before that user committed a school shooting. Assess the gap: the court has ample rational-legal authority to hear the case, but no settled doctrinal category into which it fits, because the question of what obligations attach to a platform that generates text in response to prompts was not contemplated by the negligence and product liability frameworks now being asked to bear its weight.

British Columbia’s suit against OpenAI and Sam Altman should be read first as an act of institutional taxonomy. A government is asking a court to classify a new kind of authority. Traditional authority does not apply here; OpenAI commands no inherited legitimacy. Charismatic authority explains something, since Altman’s public standing rests substantially on personal conviction rather than institutional process, but charisma is not what is on trial. What is on trial is whether a rational-legal apparatus, built on procedures, logs, terms of service, and content-moderation protocols, generates the kind of institutional competence that the law has historically demanded of, say, a bank noticing suspicious transactions or a school noticing a troubled student. The province’s claim presupposes that OpenAI’s system produced enough structured knowledge of danger to trigger a duty. The company’s likely defense will be that no such structured knowledge existed, only a pattern-matching engine indifferent to intention.

This is the trap of bureaucratic rationalization applied to systems that were not built as bureaucracies. A hospital that fails to report a threatening patient can be judged by protocols it explicitly adopted, protocols written by professionals bound by licensing bodies with centuries of accumulated case law behind them. OpenAI has moderation protocols too, but they were assembled recently, unevenly, and largely in response to public relations pressure rather than statutory command. The implementation gap is precisely here: a duty to warn is only meaningful if there exists a reliable mechanism for detection, escalation, and human judgment sitting between the machine output and the police. Whether such a mechanism functioned, whether a flagged conversation reached a human being empowered to act, whether that person weighed the signal and decided against escalation, or whether no signal reached anyone at all, is the entire case. The court cannot resolve the abstract question of AI’s duty to society. It can only resolve whether a specific set of internal procedures, followed or not followed on a specific occasion, met a standard of care that has not yet been legally defined for this class of institution.

I am reminded, oddly, of the German factory inspectorates of the previous century, bodies charged with a new duty (safety oversight) before anyone had agreed what a safe factory looked like. The inspectors did not lack authority. They lacked a settled definition of competence, and so the early rulings were less about facts than about which template of responsibility, employer, machine-designer, or worker, the courts would import. Something similar is happening now, except the machine in question produces language rather than motion, and the injury it is accused of enabling was not mechanical but psychological and eventually lethal.

The legitimacy question beneath all this is whether the public will accept a rational-legal verdict on a matter that feels, to most observers, moral rather than procedural. A grieving province wants accountability commensurate with the scale of harm; a technology company wants liability bounded by demonstrable causation and defined obligation. Courts resolve procedural questions well and moral ones badly, which is why litigation of this kind rarely satisfies anyone, win or lose, even as it slowly manufactures the doctrine that will govern the next such case.

What the structure predicts is this: the case will turn not on whether the shooting was foreseeable in some cosmic sense, but on the narrower, colder question of what OpenAI’s own systems actually registered and who, if anyone, had the standing to act on it. The verdict will be written in the vocabulary of duty and proximate cause. The grief of British Columbia will not appear in the holding at all.