OpenAI safety reviewers faced months-long backlog

OpenAI's team that reviews flagged ChatGPT conversations for threats of violence faced a backlog that stretched back months, a former member of the team told an investigation by The Observer and Tortoise. Two current employees said the intelligence and investigations team is still small, between 12 and 60 people, though The Observer could not confirm its exact size. ChatGPT is used by 1.2 billion people every week.

The former reviewer, who spoke anonymously, described reading months of chat logs from a teenager who told ChatGPT he wanted to shoot classmates who bullied him and named a date for the attack. By the time she reached the case, that date had passed. As far as she is aware, he never carried out the shooting.

Reviewers can ban accounts and refer imminent threats to the authorities. The former employee said the team did not run a round-the-clock emergency rota; instead she had a phone she was expected to answer at night if an emergency arose.

The investigation also revisits the February 2026 Tumbler Ridge school shooting in Canada. OpenAI's monitoring had flagged the shooter's conversations eight months earlier, but senior leadership decided the threat did not meet its threshold for alerting police and banned the account instead, according to the Wall Street Journal. Sam Altman later apologised, and OpenAI told the Canadian government that under a revised protocol it would refer that account to police.

OpenAI says it trains ChatGPT to recognise mental distress, de-escalate crises and encourage real-world support. It also says newer models handle mental health conversations better in its internal tests, which have not been independently verified.