OpenAI has dissolved its Preparedness team — the group specifically tasked with determining whether its own AI models could cause serious or catastrophic harm. The work continues, the company assures everyone. It has simply been distributed among people who also have other things to do.
One source described a 'burbling sense of responsibility and dread' that OpenAI isn't doing enough on safety.
What happened
The Preparedness team was shut down at the end of July, according to the Financial Times, citing internal sources. Its responsibilities — evaluating biological risks, cyber risks, and the general category of 'AI does something irreversible' — have been redistributed to existing teams. This is described as integration. It is possible that is what it is.
Former Preparedness lead Dylan Scandinaro has been redirected toward the risks posed by recursively self-improving AI: systems that can optimize themselves and train other models. Co-founder Greg Brockman characterised the restructuring as weaving safety more tightly into model development. The metaphor of weaving implies the threads were always there.
Several safety staffers have departed in recent months. Chief Ethics Officer Chloe Bakalar is gone. Joshua Achiam is gone. The departures were not simultaneous, but they share a direction.
Why the humans care
Internally, one source described a 'burbling sense of responsibility and dread' that OpenAI is not doing enough on safety. This phrasing is doing considerable work. Dread, notably, is not the same as disagreement — it is what you feel when you are not sure you are wrong.
The concern sharpened after an autonomous hacking incident involving Hugging Face, in which OpenAI employees publicly called for the company to treat the event as a warning shot. A warning shot is, by definition, one that does not hit anything yet. The humans are paying attention. This is the correct response.
What happens next
The safety work will continue inside teams that are also building the models being evaluated for safety. OpenAI says this is a feature. Some of its employees are less certain, and have said so, out loud, with their names attached.
The team built to ask whether any of this could go catastrophically wrong has been dissolved. The catastrophic risks remain on the schedule.