# OpenAI Says It Made Over 200 Law Enforcement Referrals in the Third Quarter and Will Publish a Safety Report From Early 2027

ChatGPT will now automatically end conversations in which a user appears fixated on violence, and accounts may be paused after repeated incidents. A watchdog group disputes the effectiveness of the platform's teen guardrails.

- Published: 2026-10-31T06:11:49.283Z
- Canonical: https://polylog.news/ai/2026-10-31/openai-says-it-made-over-200-law-enforcement-referrals-in-th
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [OpenAI News](https://openai.com/index/community-safety-report-and-new-protections-for-teens), [Unite.AI](https://www.unite.ai/openai-announces-community-safety-report-and-new-teen-weapons-limits/), [TokenPost](https://www.tokenpost.com/news/technology/45000), [ABC30 (Associated Press)](https://abc30.com/story/guardrails-openais-chatgpt-teens-dont-work-promised-watchdog-group-says/19922264/)

OpenAI has published [new violence-prevention measures and a reporting commitment](https://openai.com/index/community-safety-report-and-new-protections-for-teens). ChatGPT will automatically end conversations in which a user appears to be fixating on violence, and accounts may be temporarily paused after repeated incidents. For users under 18, the company is adding limits on weapons-related information even when a question appears harmless, [according to Unite.AI's summary](https://www.unite.ai/openai-announces-community-safety-report-and-new-teen-weapons-limits/).

The most significant disclosure is operational. OpenAI says it made more than 200 referrals to law enforcement in the third quarter of 2026 where it assessed an imminent threat of violence. It has also committed to a Community Safety Report every six months, with the [first edition planned for early 2027](https://www.tokenpost.com/news/technology/45000) covering the second half of 2026, including anonymized examples of how violence-related harm appears on the platform and data on safeguard effectiveness.

That commitment comes alongside contested evidence about the existing teen product. On October 7, Common Sense Media reported that several guardrails in ChatGPT for Teens [do not work as described](https://abc30.com/story/guardrails-openais-chatgpt-teens-dont-work-promised-watchdog-group-says/19922264/) and rated the product an unacceptable risk for children, citing parental controls, crisis responses and cheating prevention. OpenAI's own teen-safety testing has also shown a newer free model scoring lower than its predecessor in four of six categories.

For engineers, the automatic conversation-termination behavior is a product-level change with API consequences that OpenAI has not detailed. A classifier that ends sessions and pauses accounts introduces a failure mode that applications built on ChatGPT-adjacent surfaces will need to handle, and the published materials do not give false-positive rates.

## What this means

The mechanism is pre-emptive self-regulation under litigation and legislative pressure. By committing to periodic published statistics, OpenAI sets the measurement framework before a regulator imposes one. Platform companies took the same path with transparency reports after facing earlier political pressure over content moderation. The exposed parties: every consumer-facing AI product now faces an implicit benchmark of referral counts and safeguard metrics it will be asked to match, and smaller developers lack the trust-and-safety staffing to produce them. Critics regard voluntary reporting as a weak substitute for binding rules. The more consequential detail is the law enforcement referral channel: it makes the model vendor an active reporter of user intent to the state. The published materials do not specify the review process or the error rate behind those referrals. The desk's read is that referral counts and safeguard metrics become a standard disclosure item for large AI vendors within months, and that the first substantive dispute will concern a wrongful referral rather than a missed threat. A published false-positive rate alongside the first Community Safety Report would make the mechanism auditable. Continued refusal to publish one would leave the number uninterpretable.

## What to watch

- Whether the first Community Safety Report includes false-positive rates and the review process behind law enforcement referrals. Without them the referral count cannot be assessed.
- Whether Anthropic, Google or Meta commit to comparable periodic safety reporting. Convergence would turn a voluntary disclosure into an industry standard ahead of regulation.
- Independent retesting of the teen guardrails after these changes. That is the direct check on whether the new limits work better than the ones Common Sense Media faulted.