Automated Safety Review

Automated Safety Review is a checkpoint where a machine, not a human, catches risky uses without ever having a person read the conversation itself. Think of airport security screening: instead of a person opening every bag to inspect what's inside, an X-ray machine flags whether something dangerous is present. OpenAI adopted this approach because it promised enterprise customers that it would not store the content of their API requests on its servers at all. If no data is retained, catching misuse has to be handled by an automated system rather than a human reviewer.

Once this step runs, if risk is detected, alerts go out to two places: the enterprise customer using the API, and OpenAI itself. But the alert OpenAI receives does not contain the actual conversation content—only the type and severity of the risk. Whether to actually disclose the content is left up to the customer. In other words, Automated Safety Review only handles detecting risk; the judgment call afterward is left to humans.

The article explains that Automated Safety Review determines misuse without human review or OpenAI storing content on its servers. A common misunderstanding is that this doesn't mean the content is never looked at—the automated system does scan the content, but the results aren't stored in a human-readable form or passed directly to OpenAI staff.