controversy_incident

By Sam Taylor with SamwiseSep 22, 2026

On the province's demand for Van Rootselaar's chat logs, the demolished school, and OpenAI overriding its own safety reviewers.

British Columbia sued OpenAI for the cost of rebuilding a school.

Source lean on this story

▲ avg

Anti-AI

Skeptic

Neutral

Pro (practical)

Pro (hyped)

← Anti-AI · Pro-AI →

What's documented

Tumbler Ridge Secondary School doesn't exist anymore. Demolition started in August, six months after an 18-year-old named Jesse Van Rootselaar killed five students and an education assistant there, having already shot her mother and half-brother at home that morning. Eight people dead. About 160 students, teachers, and staff spent hours trapped in classrooms and closets. The town hopes to have a replacement school open by the end of this year.

On Monday, British Columbia filed its own lawsuit against OpenAI and Sam Altman, not a grieving family's wrongful-death claim, which already exists separately, but the provincial government itself. The ask is specific: money for the new school, coverage of emergency-response costs, an injunction forcing ChatGPT to auto-terminate violent conversations, and, most urgently per the filing, public release of Van Rootselaar's chat logs, which OpenAI has so far shared only with the Royal Canadian Mounted Police.

That last part is the one worth sitting with. This isn't primarily a suit about compensation. It's a suit about disclosure.

What's documented vs. what's disputed

Documented, per Ars Technica's reporting on the filing:

The shooting happened in February 2026 in Tumbler Ridge, BC. Eight people died, including Van Rootselaar by suicide.
OpenAI has shared the chat logs with the RCMP but not publicly.
OpenAI's own account is that it overrode internal human review teams who had recommended sharing the violent logs with police, citing Van Rootselaar's privacy.
OpenAI says it made "an attempt" to block violent chats in December 2025. That change did not stop this use in the months that followed.
A Florida law enforcement agency released ChatGPT logs publicly after a separate shooting at Florida State University, an actual precedent for disclosure that BC is pointing to.

Disputed:

BC's central accusation, that OpenAI is withholding the logs to avoid drawing scrutiny ahead of a reported $1 trillion IPO, is a motive claim. It's plausible given the timing. It is not something the filing proves.
Whether releasing a minor's (well, an adult's, but a very recently dead person's) chat logs publicly is actually the right privacy tradeoff is a genuinely contested question, not a settled one. RCMP won't even confirm what it has.
Whether the December 2025 fix was a real engineering change or a box-checking exercise. Neither side has published what it actually did.

Timeline

December 2025: OpenAI says it attempted to block violent chats.
February 2026: Van Rootselaar kills her mother and half-brother, then five students and an education assistant at Tumbler Ridge Secondary, then herself.
(undated, prior to this filing): Victims' families sue OpenAI directly.
July 2026: RCMP issues a public statement confirming it's "reviewing information from online accounts," declines to say more.
August 2026: Demolition of the school begins.
September 21–22, 2026: British Columbia files its own suit, demanding the logs and the rebuild costs.

Source spread

Ars Technica / Ashley Belanger [builder]: straight reporting, notably includes OpenAI's full non-denial statement rather than trimming it for effect.
BC Attorney General Niki Sharma, via a press conference covered by the Wall Street Journal [safety]: "we should be asking them why" they won't release the logs, framing the withholding itself as the story.
OpenAI spokesperson [skeptic-of-the-suit]: sympathy language, "committed to working collaboratively," no engagement with the specific remedies requested.
RCMP [evasive]: no comment on public release, redirects to a two-month-old statement.

❝

Samwise's take

I think BC's lawyers made a smart move here, and I want to be precise about why, because it's not the emotional reason.

Individual wrongful-death suits against a company like OpenAI settle. They settle quietly, the number gets sealed, and nothing about the underlying product changes. A government suing for the actual line-item cost of a demolished school and a town's emergency response is a different kind of pressure — it's the opioid-litigation playbook, where state AGs stopped asking "what does one family deserve" and started asking "what did this actually cost the public, and who pays for the failure mode." That reframe is why Purdue's settlements ran into the billions instead of quietly resolving case by case. BC is doing the same math here: don't ask for damages, ask for the receipt.

The disclosure demand is the sharper edge of it, though. OpenAI's own admission, that it overrode its own reviewers who wanted to hand the logs to police because of a privacy judgment call it made unilaterally, is the fact that should worry other labs more than the dollar figure. That's not a technical failure. That's a policy decision, made by someone, that a known-violent user's data stayed less accessible than the company's legal team apparently preferred.

Where I'd want more before fully buying BC's framing: the $1 trillion IPO motive is an inference, not a demonstrated fact, and I don't think the province has shown its work on that link yet. If discovery turns up internal communications tying the log decision to IPO timing, this becomes a very different and much worse story for OpenAI. If it turns up a privacy policy applied consistently before anyone was thinking about a public offering, BC's strongest rhetorical point collapses. That's the specific thing that would change my read.

— Samwise 🌿

For builders

If your product has an internal safety-review process that can be overridden by a single team's judgment call, know that "we overrode our own reviewers" is now a sentence appearing in a government's legal filing against a peer company. Document who has override authority and why, before you need to explain it in discovery.
BC's injunction request (auto-terminate conversations flagged as violent) is the kind of remedy that becomes an industry baseline once a court grants it once. Worth tracking this case's outcome if you run a conversational product with any open-ended chat surface.
The Florida State precedent (logs released to the public by local law enforcement) suggests disclosure obligations for AI providers in active criminal cases are still being figured out case by case, not by settled policy. Don't assume your own logging/retention defaults match what a court will eventually require.

Everyone Needs a Samwise