A Canadian province is suing OpenAI over a school shooting

Chronology of the event and its context (FACT unless labelled):

1. June 2025 — OpenAI's automated systems flagged Van Rootselaar's ChatGPT account for "gun violence activity and planning"; OpenAI banned the account but did not notify Canadian law enforcement (OpenAI later said the content did not meet its internal criteria for reporting; the company says it was unaware of a second account until after the shooting). The plaintiff families allege OpenAI's Intelligence and Investigations ("safety team") staff recommended contacting police and that leadership — including Altman — overruled the recommendation, citing whistleblower accounts reported by the Wall Street Journal. (COMPANY CLAIM / EARLY RESEARCH — contested by OpenAI; not adjudicated.)
2. 2026-02-10 — The shooting: eight victims killed, 25+ wounded; shooter died by suicide. OpenAI alerted the RCMP to the shooter's ChatGPT activity only after the shooting. (CONFIRMED)
3. 2026-04-23/24 — Altman published a letter of apology to the community of Tumbler Ridge (published by Tumbler RidgeLines and posted by Premier David Eby): "I am deeply sorry that we did not alert law enforcement to the account that was banned in June." (CONFIRMED)
4. 2026-04-29 — Seven victim families sued OpenAI and Altman in San Francisco federal court (Rice Parsons Leoni & Elliott LLP); more than 30 complaints total by September 2 (teachers, students, a principal; Maya Gebala and her family among them). (CONFIRMED)
5. 2026-07-07 — BC announced it had retained CFM Lawyers (Vancouver) and Stranch, Jennings & Garvey (California) to pursue legal action, coordinating with the families. (CONFIRMED)
6. 2026-09-21 — BC Government + SD59 filed the provincial lawsuit in N.D. Cal., announced at 11:45 AM via news.gov.bc.ca release 2026AG0067-001105 and at Sharma's Vancouver news conference. The complaint is described as ~39 pages (Chosun). (CONFIRMED)
7. 2026-09-21/22 — OpenAI response: spokesperson Drew Pusateri called the shooting "an unspeakable tragedy" and said OpenAI "remains committed to working collaboratively with government and law enforcement officials, and continuing to advance our ongoing safety work." (CONFIRMED — company statement)

• A sovereign sub-national government became a civil plaintiff against a frontier AI lab in U.S. federal court (not in its own courts). BC is the first Canadian province to sue OpenAI; globally this is a small club (Florida was the first U.S. state, June 2026).
• The legal theory moved from "foreseeable harm" to "foreseeable and known harm." The complaint pleads compensation for "the not just foreseeable but known harm OpenAI inflicted" and treats Altman's April apology as "an admission that OpenAI identified the risk and failed to act on it."
• Forum strategy: BC deliberately chose California (N.D. Cal.) — Sharma's stated reason: the decision not to report the chats was made in California, and OpenAI/Altman are subject to personal jurisdiction there. This sidesteps Canadian damages caps and BC forum questions in the same way the family lawsuits did.
• Parallel tracks: civil lawsuits (families, province), an ongoing RCMP criminal investigation, a coroner's inquest, and Sharma's letter to federal officials asking for Canada Criminal Code changes so AI companies "can be held accountable for criminal conduct arising from their technologies" — regulation-by-litigation-plus-legislation.
• The window's other stories (S04 California EO N-9-26 kill-switch; S05 UN panel brief on losing human control; S18 OpenAI urging global standards) show the same week converging on accountability/control; S06 is the litigation vector.

• Before: OpenAI's ChatGPT flagged a user for gun-violence planning in June 2025, banned the account, but did not alert police; the user returned on a second account; the shooting occurred in February 2026; families sued; BC retained US/Canadian counsel in July 2026; no government plaintiff existed.
• Change (2026-09-21): BC (with SD59) filed in N.D. Cal. naming OpenAI and Altman personally, pleading failure to warn, unsafe product design, negligence/product liability, damages and injunctive relief, and citing Altman's apology as an admission — while Sharma simultaneously pushed for Criminal Code amendments.
• After: OpenAI faces a multi-front docket: >30 victim/family/teacher/student complaints, the provincial government's case, a state-plaintiff precedent set by Florida, a possible BC-influenced federal law change, and an open criminal investigation. The "did the lab have a duty to call the RCMP" question is now formally a federal-court dispute that will generate discovery inside OpenAI's safety pipeline.

• Legal claim structure (as pleaded — COMPANY CLAIM until adjudicated):
  • Failure to warn / duty of care: OpenAI knew (via its own June 2025 flag and internal review) of credible threats of serious violence and did not notify law enforcement; the complaint quotes Sharma: "in any other circumstance, where a person aids, encourages, conspires to commit a criminal offence or is criminally negligent, they can be investigated, prosecuted and judged. There is no AI exemption to those criminal law principles."
  • Unsafe product / product liability: ChatGPT is characterized as "an exceedingly dangerous product" whose deployment "facilitated the mental instability of the shooter"; the attack "would not have occurred but for OpenAI's deployment" of it.
  • Aiding/abetting framing appears in companion family filings and is echoed in coverage (Chosun: violations of product-liability law and aiding/abetting a mass shooting).
  • Admission argument: Altman's April 23 letter ("I am deeply sorry that we did not alert law enforcement...") is pleaded as an admission that the risk was identified and acted on inadequately.
  • Remedies: damages for public costs (new school — the old one was demolished; federal government committed $200M toward a new school and health centre; mental health, victim services, police, educators, clinicians), injunctive relief ordering changes to how OpenAI handles conversations that could lead to violence, punitive damages under California law, interest and costs. Amounts "to be proven at trial"; complaint states damages exceed $75,000 (federal diversity threshold).
• Jurisdiction: N.D. Cal. — OpenAI is headquartered in the district; Altman resides there; the reporting decision was made there (per Sharma).
• Context of prior practice: BC has a record of public-entity litigation — tobacco ($3.7B/18-year settlement with provinces) and opioid class actions — which Sharma explicitly invoked ("This is no different").

• Tests whether a frontier lab owes a legal duty to warn third parties when its safety systems observe violence-planning in user conversations — the "AI duty-to-warn" question, now in front of a U.S. federal court with a government plaintiff and an apology letter in the record.
• Multiplier risk: a sub-national government winning damages/injunctive relief in California over conduct in Canada creates a template. Other provinces, US states (beyond Florida), and municipalities watching the OpenAI docket could file similar actions — in the same week California signed a kill-switch EO (S04) and the UN issued its agent-control brief (S05).
• Corporate accountability signal: it converts "we train models to refuse and we notify on imminent/credible risk" (OpenAI's policy language) into a justiciable standard: what counts as "imminent and credible," and who decides?
• Forum economics: choosing N.D. Cal. means US discovery rules, US damages rules, and potential jury exposure for OpenAI on a Canadian tragedy — and it keeps all the Tumbler Ridge litigation in one court.

• A government-coordinated discovery window into OpenAI's threat-reporting decisions, safety-team processes, and leadership escalation (the province demands the ChatGPT chats OpenAI has refused to share).
• Precedent-setting rulings on AI product liability and duty to warn that would bind or guide the rest of the industry.
• Canadian statutory change: Sharma's Criminal Code letter adds a legislative track running parallel to the civil track.
• Consolidated litigation strategy: SD59 as co-plaintiff lets school-district recovery ride the province's claim, and the family cases can be coordinated (transfers, common discovery, bellwether).

1. Will OpenAI produce the ChatGPT transcripts and internal escalation records, and on what timeline?
2. What exactly did the June 2025 review conclude, and who made/reversed the referral decision?
3. Will the case survive a motion to dismiss (forum/pleading), and how will the court treat the apology as an admission?
4. Does SD59's co-plaintiff status add recoverable school-district claims beyond province costs?
5. What will the RCMP criminal investigation and coroner's inquest conclude, and do they interact with the civil docket?
6. Will Canada's Criminal Code be amended in response to Sharma's letter — and would that make Canadian forums more attractive next time?
7. Does the "unsafe product" characterization survive for an LLM — is ChatGPT a "product" for products-liability purposes?
8. Which other provinces/states file next, and does BC's case consolidate with the >30 family complaints?

• Weeks (Sep–Oct 2026): OpenAI's response/motion-to-dismiss filings in N.D. Cal.; possible consolidation or coordination with the >30 family complaints; disclosures/negotiations over the chats; RCMP investigation updates; BC's Criminal Code letter percolating in Ottawa.
• Months: discovery fights over internal safety deliberations (executive privilege vs plaintiffs' need); motions practice on "product" status and duty to warn; Florida case and other state actions continue in parallel; Canadian federal AI legislation responds to the Tumbler Ridge context.
• Outlook (PREDICTION, low-to-mid confidence): the case is unlikely to be tried quickly; expect 1–3 years of motions and discovery. The likeliest early inflection is a ruling on whether the apology letter and internal flags survive dismissal — that will set settlement pressure on both sides. An injunctive-relief outcome (or a settlement mandating law-enforcement referral protocols) would be the industry-wide event to watch.

This is the week's clearest signal that AI accountability is no longer a policy debate — it is a courtroom fixture. A Canadian provincial government walked into a San Francisco federal court the same week California mandated kill switches (S04), the UN warned about losing human control of agents (S05), and OpenAI itself asked Washington to lead global safety standards (S18). The Tumbler Ridge case forces the most concrete question of all: when a model's own safety systems see planning for mass violence, whose job is it to call the police — and what happens to a company whose answer was "no one's"? Whatever the court decides, "the AI should have told someone" is now a legally operative claim with a government plaintiff, an apology letter, and eight names on it.

Evidence labels used: CONFIRMED (independently corroborated events), COMPANY CLAIM (party/company assertions, unadjudicated), INDEPENDENT EVIDENCE (media/court-derived verification), EARLY RESEARCH (reported but contested/undetermined), INTERPRETATION/PREDICTION (analyst views). Discovery-record claims contradicted by verified reporting are corrected in Section 1.