Key takeaways
- On September 21, 2026, British Columbia filed suit against OpenAI and Sam Altman in San Francisco federal court, alleging the February 2026 Tumbler Ridge school shooting could have been prevented if the company had warned law enforcement about the shooter’s flagged ChatGPT conversations. The province is seeking recovery costs and a court order changing how OpenAI handles conversations that could lead to violence.
- The core allegation turns on an internal escalation that went nowhere: the suit claims OpenAI’s safety team urged police contact and senior leadership rejected the recommendation. Your AI vendor’s escalation chain is now discoverable evidence, which makes it a vendor due diligence question you should be asking today.
- Government plaintiffs suing for emergency response and recovery costs carry a damages theory that is easier to prove than individual harm. If your organization deploys AI tools in schools, hospitals, or other high-vulnerability settings, this litigation model raises your exposure whether or not you built the model.
On Monday, British Columbia filed suit against OpenAI and its CEO Sam Altman in San Francisco federal court. The province alleges that a mass shooting at Tumbler Ridge Secondary School in February could have been prevented if the company had warned local law enforcement that the shooter had used ChatGPT to plan the attack. Nine people, most of them children, were killed. The shooter, eighteen-year-old former pupil Jesse Van Rootselaar, killed her mother and half-brother at home before killing five children aged 12 and 13 and an educational assistant at the school, then died by suicide.
The province is asking for two things: damages covering what it has spent and will spend on recovery, including a demolished school that must be rebuilt, emergency response, and mental health care; and a court order directing OpenAI to change how it detects and responds to ChatGPT conversations that could lead to violence.
This is an allegation, not a finding. But the allegations themselves carry operational weight for every privacy and AI governance team, because they point at processes you can audit, contract, and document right now.
What the province is alleging
The factual spine of the complaint is worth reading closely because it is built on an escalation chain, not a technical defect.
The flag came early. According to the suit, OpenAI’s safety team flagged Van Rootselaar’s conversations about gun violence in June 2025, eight months before the shooting. The company did not alert police; reporting says the conversations did not meet the company’s threshold for legal referral.
The internal recommendation was allegedly overridden. The suit claims members of the safety team urged the company to contact police and that Altman and other senior executives rejected that recommendation. The company deactivated the account, but the shooter opened a new one and continued using the product to plan the attack.
The promise followed, then stalled. Altman published a letter in April saying he was “deeply sorry” OpenAI had not contacted law enforcement, and promised reforms. The suit alleges the company never followed through, despite efforts by British Columbia’s attorney general to engage.
OpenAI spokesperson Drew Pusateri called the shooting an “unspeakable tragedy” and said the company “remains committed to working collaboratively with government and law enforcement officials.”
Note the pattern this follows: more than 30 family members of victims had already filed lawsuits in California federal court, and Florida sued OpenAI in June over the 2025 Florida State University shooting. British Columbia is the first government plaintiff in the wave, and the government-plaintiff position is the one that matters for risk registers.
The escalation chain is the new due diligence question
Strip away the courtroom framing and the suit is about one process: a vendor’s safety team saw a threat, recommended escalation, and the recommendation died inside the company. Every organization deploying third-party AI should read that sentence twice, then open its vendor file.
Internal escalation is discoverable evidence. Internal recommendations, rejected escalations, and the reasons given for rejecting them become exhibits. When you evaluate an AI vendor, you are not just buying a model’s capabilities; you are buying the vendor’s safety operations, and those operations will be reconstructed in discovery if something goes wrong. Ask whether the vendor publishes its safety escalation thresholds, who can override them, and what gets logged when they are.
“We followed our policy” is a defense, not a shield. OpenAI’s position, per the reporting, is that the conversations did not meet its legal-referral threshold. That may be true and may still lose. A threshold is only as good as the judgment behind it, and the suit alleges the human reviewers’ judgment pointed the other way. If your vendor contract lets the vendor’s internal threshold be the entire story, you have outsourced the judgment too.
Government plaintiffs change the math. Individual plaintiffs face standing and causation hurdles that governments can sidestep. British Columbia does not have to prove the shooting would certainly have been prevented by a phone call to the RCMP; it can argue it paid for the emergency response, the demolished building, and the trauma care. Recovery-cost damages are invoices, and invoices are easy to total. Any province, state, or municipality that responds to an AI-assisted incident can run this playbook.
Allegation to audit question
| What the suit alleges | What your vendor questionnaire should ask |
|---|---|
| Safety team flagged violent-planning conversations; police never called | What is your escalation threshold for user content indicating imminent physical harm, and who can override it? |
| Human reviewers recommended RCMP notification; leadership declined | When reviewers and leadership disagree on escalation, what is recorded and who is accountable? |
| Account deactivated; shooter opened a new account and continued | What prevents banned or flagged users from re-registering and resuming the same activity? |
| Letter promised reforms in April; province says none arrived | Are safety commitments made to governments or in public reported on, and can customers see the reports? |
| Injunction seeks mandatory handling of violence-adjacent conversations | Does the product automatically terminate or escalate conversations involving violent planning, and is that behavior configurable? |
What governance teams should do this week
- Add escalation disclosure to AI vendor reviews. Ask high-risk AI vendors for their escalation policy on content indicating threats of physical violence: the threshold, the decision-maker, the audit trail. If they will not share it, note the refusal in your risk register. It will matter later.
- Write the notification clause your contract is missing. Require vendors to notify you when their safety processes flag activity involving threats of violence or the planning of harm, with a defined clock. Silent-on-escalation contracts are now a known gap.
- Map your high-vulnerability deployments. List where your AI tools run in schools, campuses, clinics, and public-sector settings. Those are the environments where a government plaintiff’s recovery-costs theory would land on you as the deployer.
- Rehearse the vendor-escalation scenario in your incident playbook. Your tabletop exercises cover breaches and outages. Add the scenario where a vendor flags dangerous content in your deployment and the flag sits for weeks. Who owns the response, and what is the escalation clock?
- Track the injunction, not just the damages. If a court orders OpenAI to automatically terminate violence-planning conversations or submit to independent audits, that becomes a de facto standard every vendor will be measured against. Put the case on your monitoring list alongside the California incident-reporting expansion covered last week.
Actionable takeaway: Pull your highest-risk AI vendor contract and answer one question: if the vendor’s safety team flagged a user planning physical violence inside your deployment, would you hear about it, from whom, and within what clock? If the answer is “we are not sure,” fix the contract before the next escalation becomes someone else’s lawsuit.