Where this was said
OpenAI's Threshold for Alerting Police — and Why So Few Cases Qualify
At 13:57 · chapter starts 12:28
Georgia Wells explains the mechanics of OpenAI's content monitoring: an automated system flags potential threats, with the most egregious routed to human reviewers who often come from law enforcement or military backgrounds. But the threshold for actually picking up the phone and calling the police is extremely high — comparable, Wells notes, to the standard she observed at social media companies, where a user typically needs to name a specific target, date, and weapon before an alert is triggered. Out of potentially thousands of flagged conversations, only about 15 to 30 per year are referred to law enforcement. This made some employees on the safety team deeply uneasy — they felt that judgment calls about whether a specific conversation constituted a 'credible and imminent' threat were better left to trained law enforcement officers with access to more information, not OpenAI employees making probabilistic guesses. Enough disagreements accumulated that a formal meeting was called. [1] — Ryan Knutson "Only 15–30 cases/year referred to law enforcement: Out of potentially thousands of flagged conversations, OpenAI refers only about 15 to 30…" 09:22
