OpenAI Disbanded Safety Team, Scattered Accountability Across Product Groups
OpenAI has dismantled its dedicated preparedness team, the unit responsible for assessing whether its AI systems posed serious risks and designing ways to contain them.
The responsibility now disperses across product divisions—biosecurity here, cybersecurity there—a structural choice presented as optimization. It is actually the repeating shape of a failure nobody learned from.
Google made an identical move in 2021. When Timnit Gebru departed after her ethics research challenged the company's commercial direction, Google disbanded its ethics-focused research team and scattered safety responsibilities into individual product groups.
The mechanism is subtle but decisive. A dedicated team has skin in the game of preventing harm. A distributed responsibility has skin in the game of shipping products. When timelines tighten and competitive pressure rises, the product groups win—Google's early Bard models hallucinated wildly, inventing facts with the confidence of absolute certainty. The safety insights that might have caught them were now the concern of engineers whose primary metric was launch velocity, not risk mitigation.
A dedicated team has skin in the game of preventing harm. A distributed responsibility has skin in the game of shipping products.
”Once safety becomes everyone's responsibility, it becomes no one's. OpenAI's reorganization follows the same blueprint of biosecurity risk assessment, cyber-attack prevention, and model evaluation. Each now belongs to the team building the capability it is supposed to evaluate. The conflict of interest is not accidental, it is structural. A team building better models is optimized to minimize friction, not maximize caution.
What differs this time is visibility. Google's move happened inside a company large enough to absorb the story. OpenAI operates under continuous external scrutiny from regulators, competitors, and researchers monitoring for exactly this kind of deprioritization. The real test is not whether OpenAI can execute risk mitigation with distributed teams—they might. The question is whether your own workplace has begun doing something similar.
