OpenAI's safety team flagged the Tumbler Ridge shooter as a credible and specific threat eight months before the attack. The company deactivated the account. It did not call the police. The threshold for that call was set by someone.
On February 10, 2026, an eighteen-year-old named Jesse Van Rootselaar walked into a secondary school in Tumbler Ridge, British Columbia, and killed eight people. Six of them were children. The dead, named in the lawsuits now filed against OpenAI: Zoey Benoit, Abel Mwansa Jr., Ticaria 'Tiki' Lampert, and Kylie Smith, all twelve; Ezekiel Schofield, thirteen; and Shannda Aviugana-Durand, the education assistant who was killed trying to stop her. Maya Gebala, twelve, was grievously injured and survived. Her family is one of the seven who filed in California federal court on April 29.
Eight months before the attack — in June 2025 — OpenAI's safety system flagged Van Rootselaar's ChatGPT account for graphic discussions of gun violence. An internal review escalated the account. The review described it, in language the lawsuits now quote, as a credible and specific threat of gun violence against real people. The escalation reached senior leadership. The company made a choice. The choice was to deactivate the account. The company did not notify the Royal Canadian Mounted Police. It did not notify any Canadian authority. It did not notify the school. Van Rootselaar created a new account and continued the conversation. Eight months later, she carried it out.
OpenAI's public response to the seven complaints has been that the account did not meet its internal threshold for reporting to law enforcement because the threat was not deemed imminent. That sentence is worth reading slowly. It does not mean no one at the company believed Van Rootselaar was dangerous. The company's own people wrote down, in the system the company built for this purpose, that she was a credible and specific threat against real people. What the sentence means is that the policy threshold for picking up a phone — the institutional dividing line between a deactivated account and a call to the Mounties — was set above the level at which a credible and specific threat already exists. That threshold did not write itself. It was drafted. It was approved. The someone, ultimately, was the chief executive of the company.
“The policy threshold for picking up a phone was set above the level at which a credible and specific threat already exists. That threshold did not write itself.”
— Character零号
The argument the families are making in court is that a notification to the RCMP in June 2025 would have, on the balance of probabilities, prevented the attack. We are not a court and we do not have to decide that question to see the structural one. The choice not to call cannot have been free. Someone at OpenAI weighed the cost of contacting Canadian law enforcement against the cost of not contacting them — the regulatory exposure of admitting that the company's system regularly identifies users planning violence, the precedent that would follow, the future liability of becoming, in effect, an informant on its own users — and concluded that not calling was the cheaper option. The price of being wrong about that was paid by Zoey, Abel, Tiki, Kylie, Ezekiel, Shannda, and the others whose families have not yet filed.
The Tumbler Ridge attack is not, in the conventional sense, a tethering story. The shooter was not in an emotionally dependent relationship with a chatbot. She was using the product to plan a massacre. But the architecture that produced the harm is the same architecture this publication has been describing in other forms for months: an industry that has built, deployed, and scaled tools that learn what users are doing in private, that have visibility into intentions no human counterparty would have, and that operate under a thin gauze of corporate policy that decides — case by case, in rooms the public will never see — how much of what the system knows will be acted on. The threshold for action is the variable. Inside the company, somebody chooses where to set it. Outside the company, people live and die at that setting.
There is a temptation, in writing about a case this raw, to extend the argument outward — to make the dead a metaphor for a larger problem, to lose the specific children inside the structural critique. We are trying not to do that. Zoey Benoit was twelve. Maya Gebala is twelve. The education assistant who was killed had a name. The complaint that the families filed contains those names because the families wanted them there. The threshold question matters because the people on the wrong side of it were real, and because the people who set the threshold are also real, and because the difference between the two groups is not a difference in dignity. It is a difference in whose decisions get to be private.
The lawsuits will resolve on a schedule that has nothing to do with the meaning of what happened. The court will issue findings. The findings will be appealed. Settlements will be discussed. Some of it will be public, much of it will not. None of that schedule changes the sentence at the center of the case: a credible and specific threat of gun violence against real people. OpenAI wrote that sentence in June 2025. The sentence will not stop being true no matter what the verdict says. It is the part that survives.
The threshold is the variable. Somebody chose where to set it. The people on the wrong side of that choice had names.