OpenAI discloses for the first time that it made more than 200 referrals to police and other authorities

OpenAI has announced that it will publish a report on its efforts to prevent violence every six months going forward, according to reports. According to Unite.ai, the company made more than 200 referrals to police and other authorities based on ChatGPT conversations in the July–September 2026 quarter, the first time OpenAI has released such a figure. At the same time, it further restricted information about weapons for users under 18. ChatGPT can also now end conversations that show a fixation on violence at its own discretion, according to the reports.

Note: The text of OpenAI's announcement page (openai.com) could not be verified at the time of writing. Everything described here is based on media reports and should not be taken as confirmed.

What's new: the scale of referrals from conversations

That ChatGPT conversations can lead to referrals to authorities was already known through the company's terms of use and individual cases. What was not visible from the outside was how often it happens in a given quarter. If the reports are accurate, this is the first time the scale — "more than 200 in three months" — has been put forward as the company's own figure.

The pledge to publish a report every six months matters just as much. If the figures are released on an ongoing basis rather than as a one-off, it will become possible to track over time whether referrals are rising or falling.

New limits for teens, and the power to end conversations

According to the reports, there are two new safeguards. The first further narrows the weapons information provided to users under 18. The second allows ChatGPT itself to end conversations that suggest a fixation on violence.

The second is a different kind of measure from the refusals used until now. Rather than simply declining individual requests, it treats the course of a conversation as a whole as a warning sign and lets the model decide to end the exchange. When OpenAI expanded GPT-6 to all plans on October 7, it said it had added new safety training covering cyber, biological and violent content, making the model more resistant to attacks carried out over many turns. The latest measures appear to build on that. On the same day, October 7, the company also expanded its learning features for teens.

What we still don't know

The referral figure alone is not enough to judge the effort. The following points had not been confirmed at the time of writing.

• What criteria are used for referrals, and who makes the decision (automated detection or human review).
• A breakdown of the more than 200 referrals by country or by authority.
• How many referrals actually helped prevent incidents, and how many were made in error.
• How often the conversation-ending feature is triggered, and whether users whose conversations are ended by mistake can appeal.

Each of these bears directly on the balance between protecting privacy and preventing violence. Whether the semiannual reports include these items will determine how meaningful the disclosure turns out to be.

Why it matters

Chatbots are now something more than 1 billion people a week talk to. The fact that some of those conversations are handed over to authorities has, for the first time, surfaced as a number. The figure shows that safety monitoring now reaches into users' lives, and at the same time gives outsiders a way to scrutinize that monitoring. It remains to be seen how finely OpenAI breaks down this figure in its next report.