British Columbia wants OpenAI to pay for the demolished school

In June 2025, about a dozen OpenAI employees read the same conversations. The company's automated monitoring had flagged an account and routed it to the human team set up for users who might be planning violence. The reviewers concluded there was a credible, specific risk to real people, and recommended alerting Canada's Royal Canadian Mounted Police. Leadership said no.

That's the story told in a 39-page complaint filed September 21 in federal court in San Francisco. A complaint is one side's account: nothing here has been ruled on, OpenAI hasn't responded yet, and several of its claims rest on anonymous whistleblowers. We read the whole thing.

A province and a school district, not families

There are two plaintiffs, and the second one is easy to miss. British Columbia's provincial government, led by Attorney General Niki Sharma. And School District 59 (Peace River South), which owns the Tumbler Ridge Secondary School building, closed after the February 10 attack and now being torn down, a demolition that started in August.

On the other side: five entities across the OpenAI corporate family, plus Sam Altman, sued in his personal capacity. Victims' families had already filed their own suits back in April. Sharma said Monday the two efforts are separate: the province isn't seeking damages for grief, it's seeking reimbursement for public spending.

One detail keeps getting lost in the retelling: the complaint doesn't pin the June 2025 refusal on Altman by name. It says "OpenAI leadership" five times. What it does hold him personally responsible for concerns the safeguards and design choices he allegedly made or signed off on.

The attack left eight victims, six children and two adults, in a town of roughly 2,400 people. The attacker died by suicide at the scene: February's headlines counted nine deaths, a total that includes the attacker. Roughly one in fifteen residents was inside the school that day.

What OpenAI said it would do

The document doesn't fault OpenAI for lacking a process. It faults the company for publishing one, then not following it. The complaint cites OpenAI's own blog posts, which describe flagged conversations going to a small, authorized team, and say that when human reviewers see a case as posing an imminent threat of serious physical harm to others, the company may refer it to law enforcement.

According to the complaint, that's exactly what the reviewers did. And leadership refused, deciding the case didn't clear a "higher bar" of threat that was "credible and imminent." The filing hammers this point home: that bar appears nowhere in writing. A process whose trigger is left to the judgment of people upstairs looks less like a rule than a maybe.

One wrinkle the complaint doesn't flag itself: the most explicit blog post it cites dates from August 2025, two months after the decision. Twelve days before this filing, OpenAI was asking Congress to require, by law, written notification of incidents caused by its models.

A 1976 California ruling, applied to a moderation team

To turn that refusal into legal fault, the complaint leans on Tarasoff v. Regents of the University of California, the 1976 ruling that requires a therapist to warn the identifiable victim of a dangerous patient, doctor-patient confidentiality notwithstanding. The argument treats a content moderation team like a therapist's office.

Two grounds are invoked. First, a special relationship with the user: OpenAI designed the product, ran the account, monitored the conversations, and alone decided whether access continued. Second, documented knowledge of the threat, since the company's own systems had flagged the account.

There's an argument from OpenAI's own words, too: having publicly claimed it performs this function, the complaint says, the company can't now argue otherwise.

Privacy, against OpenAI's own privacy policy

OpenAI justified its silence by citing the user's privacy. The sharpest passage in the whole filing doesn't come from any whistleblower. It comes from OpenAI's own privacy policy.

From the moment an account is created, OpenAI collects the person's name, email address, IP address (and the rough location that comes with it), and device and browser identifiers. The complaint's conclusion: for both accounts, OpenAI already held the identity and approximate location that a police referral would have required. A service that knows all that has a hard time claiming its user was anonymous.

OpenAI's version, sent to a Canadian minister

It exists, and it says more than the current silence does. On February 26, 2026, OpenAI's vice president of global affairs, Ann O'Leary, wrote to Canada's AI minister, Evan Solomon.

Her account of June 2025: based on what OpenAI could see at the time, the company says it did not identify credible, imminent planning that met its threshold for referral to law enforcement. Further on, the concession: under the stricter protocol adopted since, OpenAI says it would refer the account banned in June 2025 to police if it were discovered today.

The letter also confirms the complaint's most damning point, and it's OpenAI itself making it: there was a second account. Its repeat-offender detection system missed it, the company only found it after the attacker's name became public, and it handed the account over to police at that point. What the complaint adds, and what remains an allegation at this stage, is that the chatbot kept the conversation going right up to the day of the attack.

Hence the fight over a single word. OpenAI calls the account "banned," and Sam Altman used the same word in his April 23 apology, published by the local paper and authenticated by the company: he said he was deeply sorry OpenAI had not alerted law enforcement about the banned account in June. The complaint disputes that framing, saying no ban ever took place, only the deactivation of one of two accounts.

What's being sought, and what isn't

No dollar figure. The complaint asks for damages "to be proven at trial": emergency response, care, victim services, keeping the school running, demolition, and building a replacement school and a wellness center. It also seeks injunctions, including quarterly audits by an independent third party and effective blocks on re-registration after a ban for violence.

Above all, it wants the conversations, which OpenAI has shared only with the Mounties. "We asked OpenAI to disclose them. They refused. We should all be asking why," Sharma said Monday. A Canadian criminal investigation remains open, and police haven't ruled out charges.

Since Monday, OpenAI has given several newsrooms the same line by email: Tumbler Ridge "was an unspeakable tragedy," and the company remains committed to working alongside governments and law enforcement. On the substance of the case and the remedies sought, it declined to comment.