OpenAI Eliminates Inefficient Gap Between Teams That Build AI and Teams That Declare It Safe

SAN FRANCISCO—Following the departure of safety chief Johannes Heidecke and the consolidation of its research and safety divisions, OpenAI announced Thursday that the reorganization had already yielded its most significant safety improvement in company history: the complete elimination of the lag time between a model being trained and that model being declared safe.

"Previously, our safety team would receive a completed model and spend weeks stress-testing it before sign-off," a company spokesperson said, describing the former workflow. "Now that the two teams are the same team, we can surface concerns much earlier — ideally, before the model exists."

The merger places OpenAI's researchers and its contracting safety staff under newly appointed VP of Research and Safety Mia Glaese. A company representative noted that the prior structure, in which the safety organization could independently escalate concerns to the board before a model shipped, had created what internal communications described as "an adversarial dynamic that was counterproductive to shipping."

OpenAI declined to confirm whether the unified division retains veto power over its own releases, noting only that "alignment decisions are now handled in-line."

The company said it had never been more serious about safety, a statement it described as consistent with all prior statements on the subject.

"We removed the handoff," a spokesperson said. "The research team and the safety team are now the same people, which means every model ships with full safety team approval by definition."