Months before a deadly school shooting in Tumbler Ridge, British Columbia, OpenAI employees internally debated whether violent ChatGPT logs should be reported to Canadian police. The case highlights a difficult boundary for AI companies: when does a troubling conversation become an emergency that justifies contacting law enforcement?
What OpenAI employees saw
According to the report, Jesse Van Rootselaar described gun violence scenarios in ChatGPT over several days in June 2025. About a dozen OpenAI employees discussed the activity internally after the conversations were flagged.
Some employees viewed the messages as warning signs for possible real-world violence. They urged senior management to contact Canadian police, according to the report.
OpenAI says its models are trained to steer users away from real-world violence. The company also says that when users express intent to harm, those conversations can be flagged for human reviewers.
Those reviewers can escalate a case to law enforcement if there is an immediate risk of serious physical harm. In this case, management decided not to make that report at the time.
The threshold for calling police
OpenAI later said Van Rootselaar’s account had been suspended. But a company spokeswoman told the Wall Street Journal that the activity did not meet the bar for contacting law enforcement.
The standard cited by the company was a "credible and imminent risk of serious physical harm to others." That wording matters because it shows the company was not deciding whether the chats were disturbing in a general sense. It was deciding whether they met a specific threshold for emergency reporting.
That distinction is central to the privacy and public safety tension around AI systems. Chatbots can receive private, disturbing, fictional, confused, or threatening messages. Companies then face a hard operational question: which messages are warning signs, and which are not specific enough to justify involving police?
The source article does not say that ChatGPT caused the attack. It says OpenAI staff debated the logs, senior management decided against reporting them, and the company later contacted the RCMP after learning about the attack.
Other digital warning signs
ChatGPT was not the only online space mentioned in connection with Van Rootselaar. On Roblox, she allegedly simulated a mass shooting in a shopping mall. She also joined online discussions about YouTube videos from gun enthusiasts, according to the report.
These details point to a broader issue for platforms. A single service may see only part of a person’s online activity. One company might review chatbot messages, another might host simulated behavior, and another might carry discussions connected to weapons-related content.
That fragmentation makes risk assessment harder. A single flagged conversation may look ambiguous inside one company’s review process, while a fuller pattern across platforms could appear more concerning. The source does not say whether OpenAI had access to those other activities before the attack.
What happened in Tumbler Ridge
On February 10, Van Rootselaar was found dead at the scene of a shooting rampage, apparently from a self-inflicted injury. Eight people were killed and at least 25 wounded.
The Royal Canadian Mounted Police identified the 18-year-old as the suspect. After learning about the attack, OpenAI contacted the RCMP and is now cooperating with their investigation.
The sequence creates a stark policy question for AI companies. If a conversation is violent but not judged imminent, should it stay inside a company’s safety process? Or should companies report more aggressively when messages suggest potential real-world harm?
The facts in the report show why there is no simple answer. Reporting too little can leave warning signs unshared. Reporting too much can expose private user conversations to police even when the threat is unclear or not imminent.
Why this matters for AI safety
The case puts human review, law-enforcement escalation, and user privacy under scrutiny. OpenAI’s stated approach depends on identifying intent to harm and determining whether the risk is immediate enough to involve authorities.
That process requires judgment. Reviewers and managers must interpret language, context, and risk without adding certainty that may not be there. The source shows that OpenAI employees did not all view the same material the same way: some wanted to alert police, while management decided the reporting threshold was not met.
For the wider AI industry, the lesson is not only about one company or one chatbot. It is about how platforms define escalation rules, how they document internal debates, and how they balance privacy against public safety when violent content appears in user interactions.
As AI tools become more embedded in private conversations, companies will continue to face these decisions. The Tumbler Ridge case shows that the most difficult calls may happen before anyone knows whether a threat is real, imminent, or part of a larger pattern.