What the review queue can establish
A chatbot responding appropriately to distress and a person receiving timely support are different outcomes. The Observer’s investigation makes that distance concrete. A former safety employee describes flagged conversations accumulating in a review queue. OpenAI says it trains models to de-escalate crises and encourage human support. The employee’s account concerns a service process; the company’s description concerns model behavior. A reassuring answer and a functioning route to assistance therefore require separate measurements.[1]
In the former employee’s example, a concerning conversation was reviewed weeks after the date the user had discussed. The report does not say an attack occurred in that case. That distinction matters: the account gives a reason to investigate response time, but no denominator for estimating harm. Events recalled by a particular employee are not a random sample of all flagged conversations. Queue size, waiting times and outcomes would have to be reported together before an incident rate could be estimated.[1]
Responses, people and outcomes
The company’s simulated-conversation tests announced in October measure something else. According to the investigation, two newer models followed company policies in roughly 99 percent of responses during extended self-harm conversations, against 86 percent for an earlier model. These are internal results without independent verification. Their unit is a response, rather than a person receiving assistance or a crisis prevented in everyday use. Reading that rate as the proportion of users kept safe would extend the result beyond its measured endpoint.[1]
There is still value in the result. Producing dangerous answers less often could be a useful intermediate indicator. A strong alternative explanation is that newer models have genuinely improved and some failures in older examples are now less common. OpenAI also says its 2025 risk estimates describe older models. Yet an appropriate response does not establish that a referral reached human review in time. Response quality and the assistance process may influence each other, while still requiring distinct evidence.[1]
The comparison that would strengthen confidence
A more informative prospective evaluation would follow connected stages within the same model version: a concern being flagged, entry into review, a human response and subsequent outcomes. Each stage would need its denominator, follow-up duration and missing-data rate. Privacy and participation conditions would belong in that design too. This describes an evaluation that could connect the simulated endpoint with everyday outcomes; it does not imply such a clinical study has already taken place. Without linked stages, the location of a failure remains uncertain.[1]
Neuropsychiatrist Tom Pollak’s call for outside researchers to access data bears directly on that distinction. Independent research can take complaints seriously without treating each complaint as established causation. It can ask what the model says and what subsequently happens to the person as separate questions. Fewer harmful answers accompanied by better user outcomes and shorter intervention times in current versions would strengthen confidence. An increase in policy compliance alone remains a narrower finding, although still a potentially valuable one.[1]