Your AI Safety System Hands Out Phone Numbers

When OpenAI's system failed to flag a user who went on to kill eight people in Tumbler Ridge, British Columbia, in February 2026, the company faced a concrete safety failure: a shooter used ChatGPT, the system detected something, and nothing happened. Canada called it out. OpenAI's response, in part, involves ThroughLine — a rural New Zealand startup that routes people in crisis to helplines. ThroughLine is now in discussions with The Christchurch Call, an initiative formed after a white supremacist killed 51 people at mosques in Christchurch in 2019, to build an extremist deradicalization chatbot. The same labs that couldn't stop a school shooter are getting credit for preventing radicalization.

That gap is the story.

ThroughLine was hired by OpenAI, Anthropic, and Google to handle a narrow task: when a user types something suggesting self-harm, domestic violence, or an eating disorder, route them to a human helpline. It manages a network of 1,600 helplines across 180 countries. The founder, Elliot Taylor, is a former youth worker based in rural New Zealand — the kind of operation where the main helpline and the founder's phone number are probably still the same line. It is crisis routing, not crisis prevention. A user already in distress gets a number. The system does not predict who will become dangerous. It does not flag potential shooters.

That distinction matters because the Christchurch Call conversation treats it as if it does. Galen Lamphere-Englund, a counter-terrorism adviser representing The Christchurch Call, told Reuters he hoped to roll ThroughLine's product out to moderators of gaming forums — places where extremist content spreads before it becomes violence. The implication is that an AI chatbot telling someone to call a helpline will interrupt radicalization before a mosque shooting, a bombing, or a school.

The Tumbler Ridge case suggests otherwise. On Feb. 10, 2026, a 12-year-old was shot in the neck and head in the small British Columbia mining town. Eight people died. Canada's AI minister, Evan Solomon, publicly threatened OpenAI that same month over what he described as the company's failure to report a Canadian ChatGPT user who went on to commit the attack. OpenAI has denied the allegation. But the incident exposes the difference between the threat model ThroughLine was built for — someone already in crisis, reaching out — and the threat model that produced Tumbler Ridge, or Christchurch, or any of the mass casualty events that preceded them.

The Christchurch Call itself was formed in May 2019, after a terrorist broadcast his attack on Facebook Live. Fifty-one people died. The initiative brought New Zealand's government together with major technology companies to try to prevent the online spread of terrorist content. Eight years later, the conversation has shifted from removing videos to deploying chatbots. The gap between those two things — content moderation at the point of distribution versus conversational intervention at the point of ideation — is where the PR and the reality diverge.

There is nothing wrong with crisis routing. Helplines save lives. Routing someone in acute distress to a trained human is better than leaving them alone with a chatbot. But crisis routing is not deradicalization. It is not violence prevention. It is triage after the crisis has already begun. When Galen Lamphere-Englund talks about gaming forum moderators using ThroughLine's tools, he is describing a product designed for a different failure mode than the one that produces mass shootings.

The labs know this. OpenAI, Anthropic, and Google hired ThroughLine because it is operationally useful, not because it solves the hard problem. The hard problem is predicting who will become violent before they act. No chatbot has solved that. The Christchurch Call's interest in ThroughLine is useful to the labs because it turns a narrow safety tool into a broader public safety narrative — exactly the kind of association that matters when regulators, advertisers, or governments ask what these systems are actually doing to make the world safer.

Canada's threat to OpenAI was about a concrete failure: someone used the product and killed people, and the system did not report it. The Christchurch Call conversation is about a prospective product: an AI chatbot that might, someday, talk someone out of radicalization. One is a measurable, auditable question. The other is a promise. The labs are getting credit for the second while the first is still unresolved.

The Christchurch Call was formed to prevent what happened in Christchurch. If the answer eight years later is a chatbot that routes gamers to helplines, it is worth asking what changed and what didn't.

---

† Consider revising to: 'The Christchurch Call itself was launched on May 15, 2019' and include a citation to the Christchurch Call website for the specific date, or attribute to a specific source that confirms the May formation date.