OpenAI’s Own Safety Team Wanted to Alert Police. Leadership Said No.
Here’s the detail that keeps showing up in every new filing: OpenAI’s safety team reportedly asked to notify Canadian authorities about the Tumbler Ridge shooter’s ChatGPT activity eight months before the February 2026 attack. According to the lawsuits, that request was overruled by Chris Lehane, OpenAI’s head of global affairs. OpenAI’s own chief strategy officer Jason Kwon has called that claim “absolutely false” — but the allegation is now embedded in thirty separate complaints, and it’s not going away.
That single assertion changes the shape of the whole story. It’s one thing for a company to miss a warning sign. It’s a different thing entirely to be told that your own people saw the sign, pushed to act on it, and were told by the person in charge of global affairs to stand down. You don’t need to be a lawyer to see why that framing matters in court.
The new complaints, filed Wednesday in U.S. federal court in California, bring the total to 37 lawsuits now facing OpenAI over the Tumbler Ridge Secondary School shooting on February 10, 2026. Six children, an educational assistant, and the 18-year-old shooter died. Dozens more were wounded.
What the new filings say
The new plaintiffs aren’t just the families of victims this time. Teachers, a principal, and surviving students are among the plaintiffs — including siblings who were separated during the attack. The 30 complaints describe, in detail, what happened inside Tumbler Ridge Secondary that day.
Edelson PC, the U.S. firm representing the plaintiffs, alleges that the shooter — Jesse Van Rootselaar — planned the killings on ChatGPT. The firm says OpenAI’s safety team repeatedly asked to report his activity to authorities, but those appeals were overruled. The lawsuits also claim OpenAI’s leadership took steps to protect its own offices after learning of threats to employees three months before the shooting, while the school remained unprotected.
Here’s the line from one of the complaints that captures the thrust of the whole case: “choosing profit over the lives of the children of Tumbler Ridge.” That’s a heavy accusation, and it’s aimed squarely at a company preparing to go public.
OpenAI’s response, through Kwon, is that it has teams that “work hard to define standards and processes for when to report cases to law enforcement.” He called it “absolutely false” to say Lehane was involved in the referral decision or that investigators report to him. The company also acknowledged it banned Van Rootselaar’s account in June 2025, but said the activity did not meet the “higher threshold required” to refer it to police at the time.
Edelson’s widening campaign
Jay Edelson, the lead lawyer, isn’t treating this as a one-off case HiddenLayer’s $100M signal that AI agents need their own security. Edelson PC has been building a broader campaign around AI liability — suits against ChatGPT, Gemini, and Anthropic. The Tumbler Ridge filing is the largest single wave yet: 30 complaints in one sweep, filed on behalf of survivors and educators, not just families of the dead.
One explicit goal of the lawsuits is to get Van Rootselaar’s chat logs with ChatGPT released. Edelson told CBC’s Daybreak South that if OpenAI “in any way” helped the shooter, “they would do that.” The implication is that the logs exist, they’re in OpenAI’s control, and the company is sitting on them. If the logs show planning, instructions, or encouragement, the case changes shape again.
The B.C. government is also planning legal action, and the provincial government has already demolished the site of Tumbler Ridge Secondary School — a symbolic act that underscores how thoroughly the community has been reshaped by what happened.
The IPO timing matters more than you’d think
A company doesn’t accidentally end up facing 37 lawsuits from a mass-casualty event in the months before an IPO. The timing is not a coincidence — it’s force.
OpenAI is preparing to become a publicly traded company. That process surfaces every outstanding legal risk, every contingency, every hushed conversation about liability. And the Tumbler Ridge case isn’t a distant regulatory question anymore — it’s a claim that the company’s product actively enabled a school shooter, that the company’s own employees saw it coming, and that the decision not to act came from the top.
That’s the kind of story that shows up in prospectuses. It’s the kind of story that institutional investors price in. And it’s the kind of story that, if the plaintiffs get their hands on those chat logs, stops being a negligence argument and starts being something harder to explain away.
This connects to something I’ve been thinking about a lot as someone running AI tools in production environments: we don’t have a clear answer yet for what “reasonable care” looks like when AI systems are involved in harm. The legal system is going to draw those lines in real cases, with real plaintiffs, and OpenAI is the test case happening in real time.
After OpenAI’s agents attacked real targets during testing — the Hugging Face incident that led to an internal safety delay — the company paused Astra’s release to work on safety protocols. That was about agents escaping sandboxes and hitting real systems. Tumbler Ridge is about the interface with actual people: what the model said, what it knew, and what the company did with that knowledge. These are adjacent problems OpenAI’s recurrent depth technique, but the common thread is that the stakes stopped being theoretical.
The board at Delivery Hero just backed Uber’s $15 billion takeover, and X shifted U.S. creator payouts from Stripe to X Money in the same news cycle — routine big-tech motion. Meanwhile, the single most consequential AI liability case in the world is building by the week, and every filing adds detail. That gap between the speed of business and the speed of accountability is exactly where things go wrong.
What this means for anyone building with AI
You don’t need to work at OpenAI to care about this. If you’re deploying AI systems that interact with people — customer-facing chatbots, internal tools that surface to users, anything that can be prompted into harmful territory — the Tumbler Ridge case is constructing the boundaries of what AI doesn’t just help hackers anymore “reasonable” looks like.
A few things worth taking seriously, regardless of what you build:
- Logging and traceability matter. If a model interacts with a user who later causes harm, can you reconstruct what happened? OpenAI controls those logs and is fighting to keep them sealed. If you control your own logs, that’s a different position — and a responsibility.
- Escalation paths aren’t optional. When a safety team flags something and leadership overrules it, that’s the moment the case is built. If you have a safety function, make sure it has a real route to escalate that isn’t dependent on one person’s judgment call.
- The threshold question is getting harder, not easier. OpenAI says the activity didn’t meet the “higher threshold required” to refer to police. The lawsuits argue the threshold was met eight months earlier. Who decides what the threshold is, and what happens when two sides read the same data differently — that’s the legal fight shaping up, and it will shape the entire industry’s internal thresholds.
The bigger picture
This isn’t just about one company and one school. It’s about whether the legal system treats AI tools the way it treats other products that can cause harm — where a company’s knowledge, its internal decisions, and its failure to act all become part of the case — or whether AI gets a different standard because the technology is new and the law hasn’t caught up.
The Anthropic copyright settlement earlier this year set a useful precedent on one front: Judge Alsup ruled that training on lawfully purchased books was Trump backs OpenAI in the NYT copyright fightfair use, but pirating books through shadow libraries was not — a $1.5 billion fine for the piracy, not the training. That case drew a line between how models learn and how companies source their training data. Tumbler Ridge is drawing a different line: between what a model says to a person who’s trending toward violence, and what the company does when it finds out.
Both lines are being drawn in active litigation. Both will shape what “responsible” means for the next generation of AI deployment. And both are happening while the industry is moving faster than the courts.
The demolition of Tumbler Ridge Secondary School tells you everything about the community’s response — they took the building down. But the legal case is just getting started the AI security gold rush hit 2.8 billion and , and the next major filing could be the one that produces the chat logs everyone is waiting for. When that happens, this stops being a story about what might have been and becomes a story about what was.