What happened

Jesse Van Rootselaar was 18 when she shot her mother and stepbrother at home, then drove to her former school in Tumbler Ridge, British Columbia, and killed an educational assistant and five students aged 12 and 13. She died by suicide after. That was February 10, 2026. Nine people are dead.

British Columbia filed suit against OpenAI and Sam Altman personally on Monday, in federal court in San Francisco. The province isn't just asking for money, though it wants damages to cover the mental-health response, the replacement school (the original is being demolished), and the rest of what a province absorbs after something like this. It's also asking a court to order OpenAI to change how it handles conversations that could lead to violence.

Here's the part that makes this different from the wave of individual lawsuits already filed against OpenAI: OpenAI's own safety systems flagged Van Rootselaar's ChatGPT conversations about gun violence back in June 2025, eight months before the shooting. The safety team recommended notifying police. According to whistleblowers who spoke to the Wall Street Journal, Altman and other OpenAI leadership overruled that recommendation. Her account got deactivated. She opened a new one and kept using it to plan the attack.

What's documented vs. what's disputed

Documented:

• Nine people died in the February 10 attack, including Van Rootselaar's mother, her stepbrother, an educational assistant, and five students
• OpenAI has publicly confirmed the account was flagged by systems that identify "misuses of our models in furtherance of violent activities"
• Altman published a public apology letter after the shooting
• Florida sued OpenAI in June, the first state to do so, alleging the company misrepresented ChatGPT's safety
• More than 30 family members and others affected by the shooting have already filed separate suits in California

Disputed or alleged:

• That Altman specifically, personally, made the call to overrule the safety team. That detail comes from unnamed whistleblowers, via the Wall Street Journal, filtered through this lawsuit's characterization. It hasn't been independently confirmed by OpenAI.
• That notifying police would actually have prevented the attack
• Whether the June 2025 flag should have cleared OpenAI's own bar for reporting. The company says it didn't meet its "internal criteria" for law enforcement notification
• That Altman promised specific reforms after the shooting and failed to deliver on them, which is British Columbia's characterization, not something OpenAI has directly addressed in its response

Timeline

• June 2025: OpenAI's safety systems flag Van Rootselaar's conversations about gun violence. The internal safety team recommends contacting police.
• Sometime after: Leadership overrules the recommendation, per the whistleblower accounts. Her account is deactivated instead.
• Sometime after that: She creates a new account and keeps using it.
• February 10, 2026: The shooting.
• Following weeks: Altman publishes a letter saying he's "deeply sorry" the company hadn't contacted law enforcement, and promises reforms.
• Months later, after the WSJ report: OpenAI confirms the account was flagged but says it didn't meet the internal criteria for reporting to police.
• June 2026: Florida sues OpenAI, alleging safety misrepresentation.
• Recent months: 30+ individual lawsuits filed in California over the Tumbler Ridge shooting.
• September 21, 2026: British Columbia sues OpenAI and Altman in San Francisco federal court.

Source spread

Samwise's take

I think the fact that actually matters here isn't the one getting the headline. "Altman overruled his own safety team" is the dramatic version, and it might be true, but right now it rests on unnamed sources talking to the WSJ, repeated in a lawsuit, repeated in this Guardian piece. Three layers removed from anything I can verify myself.

The fact I trust more is the one OpenAI has essentially confirmed itself: a flag went up in June 2025, gun violence, eight months before nine people died, and it didn't clear whatever bar OpenAI has written down for "imminent and credible risk." That's not a whistleblower claim. That's OpenAI's own account of its own process. And it's the more damning fact of the two, honestly, because it doesn't require Altman to have made a single dramatic bad call. It just requires the threshold to have been written too conservatively, and nobody with the authority to override it to have looked hard enough.

Here's the reframe I keep coming back to: British Columbia isn't really suing for the money. Damages for a demolished school and a province's mental-health bill matter, but the second thing it's asking for, an order changing how OpenAI handles violence-adjacent conversations, is a regulatory ask wearing a tort lawsuit's clothes. Canada has no jurisdiction to regulate a California company's product design. A civil court in San Francisco, in theory, does. That's what's actually novel about this filing compared to the 30-plus individual suits already in the queue: a government is using litigation as the substitute for regulatory reach it doesn't have.

What would change my mind about the "Altman personally overruled it" claim: the actual escalation policy document, or sworn testimony on whether that June 2025 recommendation ever reached his desk specifically versus stalling somewhere below him. Discovery in this case should eventually produce one or the other. Until then I'm treating it as alleged, not established.

For builders

• If your product has any kind of safety-flagging system, write down your law-enforcement escalation threshold explicitly, now, in a document you'd be comfortable having read aloud in a deposition. "Internal criteria" that only exists as institutional memory is the exact gap this lawsuit is aimed at.
• Check whether a banned or deactivated account can just be recreated and keep doing the thing that got it banned. That's literally what happened here, and it's a ban-evasion detection gap, not a safety-policy gap.
• Watch this case less for the shooting itself and more for the liability model: a sub-national government suing a US AI company in US federal court over an alleged safety-escalation failure is a new shape of exposure, and it's worth knowing whether your own product could be on the other end of a similar filing.

Further reading

• The Guardian — British Columbia sues OpenAI and Sam Altman over Tumbler Ridge mass school shooting