British Columbia and School District 59 Sued OpenAI and Sam Altman on 21 September 2026 Over a ChatGPT Account Flagged in June 2025 and Not Reported to Police Before the Attack
In October 1969 Prosenjit Poddar killed Tatiana Tarasoff in Berkeley, California. Two months earlier he had told a psychologist at the university's student health service that he intended to do it. The psychologist asked campus police to detain Poddar; they interviewed him, judged him rational, and let him go. Nobody told Tatiana Tarasoff or her parents. In 1976 the California Supreme Court held, in Tarasoff v. Regents of the University of California, 17 Cal. 3d 425, that a therapist who determines, or under professional standards should determine, that a patient poses a serious danger of violence owes a duty to exercise reasonable care to protect the foreseeable victim. That duty rested on the special relationship between therapist and patient, which gave the clinician both the knowledge and the standing to act, not on general foreseeability. The court summed it up in the sentence the case is remembered by: the protective privilege ends where the public peril begins.
Practical takeaway. This article focuses on the alleged failure to escalate the flagged account: someone inside OpenAI concluded that it did not meet the company's bar for a call to the police. The open questions are the ordinary ones — which law governs, what duty if any was owed, and what caused the harm — and each pleaded theory carries its own elements.
June 2025: reviewers flag a ChatGPT account for gun-violence conversations, and management declines the police referral
According to the province's account of its complaint, OpenAI's safety systems identified activity on an account belonging to Jesse Van Rootselaar, then seventeen, as potential misuse connected with violent conduct. Human reviewers looked at it. They recommended that the company contact the Royal Canadian Mounted Police. Company leadership decided the material did not cross OpenAI's internal threshold for referral to law enforcement, and the account was banned in June 2025 instead. A second account was opened afterwards, and OpenAI's repeat-violator detection did not connect it to the banned one.
OpenAI's own account of that month is on the public record, and it is more specific than the company's later statements to reporters. In a letter dated 26 February 2026 to Evan Solomon, Canada's Minister of Artificial Intelligence and Digital Innovation, Vice President of Global Policy Ann M. O'Leary wrote that an automated system detected the account, that it went to human review to determine whether the usage policies were violated and whether the account warranted referral to law enforcement, and that "based on what we could see at that time the account was banned in June 2025, we did not identify credible and imminent planning that met our threshold to refer the matter to law enforcement."
The disagreement is therefore narrower than an absence of any standard. OpenAI had a threshold, applied it, and concluded the material fell short. What the reviewers saw, and whether any identifiable intended victim was named in it, is not public. The failure the plaintiffs allege is not a gap in the machinery but a judgment call at the top of that path.
10 February 2026: eight people die, and the account nobody connected becomes the fact that matters
Tumbler Ridge is a coal-mining town in the Peace River country of northeastern British Columbia, population somewhere near two thousand. On 10 February 2026 Van Rootselaar, by then eighteen and a former student at the secondary school, killed a mother and a younger half-brother at home, then five children and an educator at the school, before dying at the scene. Eight dead, six of them children. It is among the worst mass killings in Canadian history.
That detail — Peace River South School District 59 suing alongside the province — is what puts a number on the damages. The relief sought includes the cost of emergency response and recovery and the cost of building a replacement school, because the building where it happened is not a building the district can reasonably ask children to return to.
April 2026: Altman writes that he is deeply sorry, and the first seven private suits are filed
In April 2026 Altman sent the Tumbler Ridge community a letter containing a line the province is reported to rely on: "I am deeply sorry that we did not alert law enforcement to the account that was banned in June." Premier David Eby, sharing the letter, said the apology was "necessary, and yet grossly insufficient for the devastation done to the families of Tumbler Ridge." The same month, seven civil suits were filed in the Northern District of California by victims and families. Thirty more followed on 2 September 2026. By the time British Columbia filed, about thirty-seven private actions were already pending in the same courthouse.
An apology written for a public audience and an admission of legal fault are different instruments, and whether this letter can serve as the latter is a question for the court, not for a reader of the reporting. What the letter itself supports is narrower: it is dated, signed, and specific about which account and which month.
21 September 2026: the first government plaintiff, a named chief executive, and four theories
News accounts of the filing describe four theories: negligence, defective product design, product liability, and aiding and abetting a mass shooting. The complaint itself was not obtained from the court docket for this piece, so those counts are reported as characterized, not as read. What is consistently reported is that Altman is named personally alongside the company, and that the province asks for an injunction covering reliable refusal, termination and de-escalation of violence-related conversations, and a changed process for handling material that indicates a risk of violence.
As described, the theories ask different questions. Negligence, and the failure-to-warn argument reported to sit underneath it, ask about a decision by people. Defective design asks whether a product that detects threats and then routes them into an internal judgment is designed with reasonable care. The elements and the available defenses differ across them, and governing law and causation have to be resolved for any of them; none of that can be assessed from news coverage of a complaint. OpenAI's spokesperson, Drew Pusateri, called the events an unspeakable tragedy and said the company trains its models to refuse requests that enable violence and works with law enforcement on safety. The company has not conceded liability and has not yet answered the complaint.
Where the internal threshold came from, and why it was never external
OpenAI does publish thresholds. Its Preparedness Framework assigns capability ratings and gates model releases against them, and in September 2026 the company used it to rate GPT-6 Astra as Critical for cybersecurity, the first OpenAI model to carry that rating, with real operational consequences inside the company. That machinery is built for one question: what the model can do. The Tumbler Ridge allegation concerns a different question — what an identified user is about to do in the physical world — and the company's answer to it is set out in the February letter rather than in a versioned public framework. OpenAI described the standard in words ("credible and imminent planning", "an imminent and credible risk") and said it had partnered with mental health, behavioural and law enforcement experts to refine it, and that those experts now help assess difficult cases. What the letter does not describe — and what none of the public material reviewed here establishes — is the criteria themselves, the reasoning that applies them to a particular account, or any independent audit or regulatory review of either.
The letter also records two changes that matter to the province's case. OpenAI wrote that its referral criteria are now "more flexible to account for the fact that a user may not discuss the target, means, and timing of planned violence in a ChatGPT conversation," and that "under our enhanced law enforcement referral protocol, we would refer the account banned in June 2025 to law enforcement if it were discovered today." It also disclosed that after the perpetrator's name became public, the company found a second ChatGPT account and shared it with law enforcement. A plaintiff will read the first of those sentences as the company conceding that its June 2025 threshold was set too high. OpenAI can fairly read it as ordinary post-incident improvement, which is not the same thing as negligence at the time.
Regulators have been moving toward the first question and not the second. California's Executive Order N-9-26 of 18 September 2026, issued three days before this complaint was filed, directs work on independently verified shutdown and containment capability for frontier systems. It addresses the model. The thread this article follows runs instead to the escalation desk behind it, which the litigation route reached first, in the absence of any published rule pointing at it.
How Quentir Reads It
The interesting thing about this filing is how conventional it is. There is no novel theory of machine agency in it. On the province's account a company detected a risk, its own staff said tell the police, someone senior said no, and the province is suing on the ordinary law of duty, foreseeability and defective design. Sharma put the point this way in the province's 21 September 2026 announcement: in any other circumstance where a person aids, encourages or conspires to commit a criminal offence, or is criminally negligent, they can be investigated, prosecuted and judged, and no company should be exempt from accountability simply because the technology is new. That is an argument about which body of law applies, not a demand for a new one.
That matters for anyone running an AI system with a safety review queue, because it relocates the risk. For most of 2026 the operative compliance question has been output: refusals, jailbreaks, red-team results, capability ratings. This complaint says the exposure sits in the disposition log — who saw the escalation, what they decided, on what written criterion, and whether that criterion was ever shown to anyone outside the company. Firms that do content review at scale, in banking and on platforms as much as in AI labs, generally keep records of that kind, and the standard applied in them is typically internal. Whether any of that is legally actionable is genuinely open: Tarasoff rests on a special relationship, the California Supreme Court reached it on California law, and a suit filed in a California federal court does not by itself settle which jurisdiction's substantive law governs a Canadian plaintiff's claim. An internal referral policy is not a legal duty. The province is asking a court to make one.
A province as plaintiff also changes the arithmetic. Families sue for what happened to their family. A government sues for a replacement school, a counselling program and an emergency response — heads of loss that are public expenditure, documented in budgets, and awkward to make disappear quietly.
The daily Insights posts, including this one, are free to read. What the All-access membership adds is the paid layer around them: the Medicine and Defense Evidence Registers with a readiness level on every entry, the weekly Quentir Radar, and the members' area editions, bundled in one subscription instead of bought separately.
Watch the motion to dismiss, where the threshold questions — forum, governing law, and whether a software provider owes any duty of this shape to third parties — will be argued before anything reaches the merits. If the duty claim survives, discovery becomes possible — subject to the scope a court allows and to whatever privilege and cross-border objections OpenAI raises — over the June 2025 review thread and the decision that closed it. Every company running a safety escalation queue has an interest in what that written criterion looks like if a court ever reads it.
Sources: Al Jazeera, "Canada's BC sues OpenAI over ChatGPT role in Tumbler Ridge school shooting" (published 22 September 2026), for the 21 September 2026 filing date, the federal court in San Francisco, OpenAI and Sam Altman as defendants, Attorney General Niki Sharma's statement, the composition of the eight dead, the 10 February 2026 date and location, and the April 2026 Altman letter; Engadget, "OpenAI faces lawsuit from British Columbia over Tumbler Ridge shooting" (22 September 2026), for the unsafe-design and negligence counts, the allegation that reviewers' RCMP recommendation was overruled, the second account, and the replacement-school relief; Lawyer Monthly, "British Columbia Sues OpenAI and Sam Altman Over Tumbler Ridge School Shooting" (September 2026), for the June 2025 account flag and ban, the injunctive relief on violence-risk conversations, and OpenAI spokesperson Drew Pusateri's statement; Gadget Review, "B.C. Sues OpenAI Over Tumbler Ridge Shooting, Citing No Police Warning" (22 September 2026), for Peace River South School District 59 as co-plaintiff, the Northern District of California, the exact wording of the Altman line, and the roughly thirty-seven private suits filed in April and on 2 September 2026; Sharma's "no company should be exempt from accountability" passage is quoted from the Province of British Columbia's announcement of 21 September 2026 as reported in Canadian coverage (news.gov.bc.ca did not resolve from this lane on 22 September 2026, so the release is named and not linked); Bloomberg Law, "British Columbia Sues OpenAI for Failure to Warn Before Shooting" (21 September 2026, in part behind a paywall), for the failure-to-warn framing; CNN, "Sam Altman apologizes to Canadian community after failing to flag mass shooter’s conversations" (24 April 2026), and CBC News, "OpenAI’s Sam Altman writes apology to community of Tumbler Ridge" (April 2026), for the apology letter and its date; Supreme Court of California, Tarasoff v. Regents of the University of California, 17 Cal. 3d 425 (1976), for the special relationship, the duty to exercise reasonable care to protect a foreseeable victim, and the quoted passage; and OpenAI, letter of 26 February 2026 from Ann M. O’Leary, Vice President of Global Policy, to the Hon. Evan Solomon, Minister of Artificial Intelligence and Digital Innovation (PDF), for the June 2025 detection and human review, the "credible and imminent planning" threshold, the revised referral protocol and the statement that the banned account would be referred today, and the discovery and sharing of the second account after the perpetrator was named. The line "necessary, and yet grossly insufficient for the devastation done to the families of Tumbler Ridge" is British Columbia Premier David Eby's, not the Attorney General's. Public sources checked 22 September 2026. The complaint itself was not obtained from the court docket; every allegation above is reported as pleaded, and OpenAI has not answered it.
Published intelligence, built to inform your own decisions. Published: September 22, 2026.