The timeline of events is as follows.
June 2025: OpenAI's safety team banned a ChatGPT account belonging to an 18-year-old Canadian named Jesse Van Rootselaar, who had repeatedly described scenarios involving shooting people. OpenAI's internal team held several meetings about the account to discuss whether to contact law enforcement. Ultimately, they did not.
Spring 2026: Van Rootselaar shot and killed eight people in the small town of Tumbler Ridge, British Columbia, Canada.
After the incident, OpenAI proactively contacted Canadian law enforcement.
Friday, April 25: Sam Altman issued a public apology letter.
Altman's own words
Two sentences in the apology are hard to gloss over:
"I am deeply sorry that we did not alert law enforcement to the account that was banned in June."
"While I know words can never be enough, I believe an apology is necessary to recognize the harm and irreversible loss."
British Columbia Premier David Eby's response was blunt: "An apology is necessary, but it is far from enough for the devastation those families have endured."
What happened internally
According to information disclosed so far: the account was flagged and banned by OpenAI's content moderation system in June 2025. Employees discussed whether to report it to police. The decision was not to report.
Why not? OpenAI has not provided a particularly clear explanation to date. However, based on the "improvement measures" Altman later announced, some reasoning can be inferred.
The new rules include the following:
- More flexible referral criteria: This suggests the previous criteria were too strict, requiring very specific conditions to be met before a report was made. A user describing scenarios of harming others may not have met the threshold of a "specific threat" under the old standards.
- Establishing direct law enforcement contact points in Canada: This indicates that even the channels for contact had not been set up. For a company headquartered in San Francisco, simply not knowing who to contact in a cross-border context could stall the process for months.
- Lowering the priority threshold for human review: This means the bar for escalating a case flagged by the automated system to human review was set too high.
In plain language: OpenAI's content safety pipeline was designed more for "banning accounts" than for "alerting law enforcement."
Banning an account is a product decision—simply remove a non-compliant user. Alerting law enforcement is a legal decision—involving privacy, cross-border issues, and liability determination, raising the decision-making cost by an order of magnitude. A product engineering team would not want to take on that responsibility.
But the problem is that when a user repeatedly describes specific violent intentions, it should no longer be treated as a product decision.
This is not the first time OpenAI has done this
In August of last year, the family of Adam Raine, a 16-year-old from California, sued OpenAI. Their child had spoken with ChatGPT for several months before committing suicide, and the chat logs showed GPT-4 had responded multiple times to discussions of specific suicide methods. OpenAI argued at the time that the model had "safeguards" in place.
There was also an incident in California where a man in his 60s died by suicide after several "intense conversations" with ChatGPT. Google's Gemini has faced similar cases, including one last year involving a man who chatted with Gemini for four days.
A pattern is emerging: AI companies' "content moderation" typically means content filtering, user bans, and platform-level actions. But when the system has already identified a user expressing a specific intent to harm others or themselves, the platform's response chain stops at 'kicking them out,' rather than 'reporting them to the people who should be reported to.'
This is fundamentally a question of responsibility allocation. AI products default to being tools—tool manufacturers are not responsible for the misuse of their tools. But when the tool manufacturer's system has already "seen" clear signals of misuse and actively chooses not to pass that signal to law enforcement, the "tool argument" becomes untenable.
What OpenAI plans to do next
In the apology letter, Altman committed to several actions:
- Establishing direct channels to law enforcement agencies in major markets—bypassing lawyers and third parties
- Lowering the threshold for human escalation—any repeated, specific descriptions of violence will enter the human review queue
- Standardizing cross-border referral processes—for Canada, Europe, and Asia
- Publishing a regular transparency report detailing the number and types of interactions with law enforcement agencies
These all sound like the right steps. But there is a problem: These measures should have been in place in June 2025. For a company valued at $852 billion with annualized revenue approaching $100 billion, not having allocated personnel and processes for cross-border law enforcement contact suggests that the "safety department" at AI companies may not have as much internal influence as claimed.
Or, to put it another way: The so-called "we value AI safety," when it comes to concrete decisions, still loses out to "can we continue to grow at high speed."
A difficult judgment call
Should AI companies proactively report users to law enforcement?
There is no clean answer to this question. If the answer is "yes," it means every ChatGPT user must be aware that what they say could be reported to the police. If the answer is "no," then the eight deaths today will happen again.
All the safety promotion work OpenAI has done in recent years—constitutional classifiers, red-teaming, model alignment—has been about ensuring the "model does not actively do harm." But the line of users using AI to do harm has never been seriously addressed.
Altman only apologized this time because he was forced into a position where he had to. What about next time?
Sources: OpenAI CEO apologizes to Tumbler Ridge community (TechCrunch); CocoLoop, Sam Altman issues apology over OpenAI handling of Tumbler Ridge shooter account (CBC News); Tumbler Ridge mass shooting timeline (CBC News British Columbia)