A former OpenAI reviewer alleges that flagged conversations waited weeks for attention, raising questions about the handling of urgent warnings. OpenAI says it is strengthening safeguards and cooperation with authorities. Its newer simulated-conversation results remain internally measured, while outside researchers are seeking information on what happens to vulnerable users over time.
An OpenAI employee who formerly reviewed concerning ChatGPT conversations alleges that warnings accumulated faster than staff could examine them. Sarah, a pseudonym used by The Observer, described delays lasting weeks. She also recalled carrying an emergency phone overnight, rather than belonging to a specialist response unit staffed around the clock. The investigation did not confirm the team's precise headcount.[1]
Reviewers have powers to close accounts and send imminent threats to police. OpenAI offered sympathy to the affected families, pledging further work with authorities and improvements to its safeguards. The company describes training intended to identify psychological distress, calm crises and steer people toward support from other people. Families bringing lawsuits allege that product design and safety practices contributed to harm; those causal claims remain disputed allegations.[1]
Neuropsychiatrist Tom Pollak wants researchers outside the company to obtain data for studying vulnerable users over time. OpenAI's latest safety exercises use simulated conversations, and outside researchers have not independently verified their results. Such exercises do not measure whether harmful incidents among actual users have become less frequent. The company says its previously published prevalence estimates describe older models; replacement estimates have not been released.[1]
