OpenAI has shut down the team known as Preparedness. Its remit was single and explicit: assess whether the company's own models could pose catastrophic risks. Its work has been parcelled out to existing groups inside the company, and several safety staff have left.

What the team did

Preparedness looked at a particular question within AI safety — not whether a model swears or states something false, but its capacity to cause harm at scale. Biological and chemical weapons knowledge, offensive cyber capability and self-replication fall under that heading.

There was a logic to keeping that assessment in a separate team: independence from the group charged with shipping the product. A finding that would delay a launch carries different weight when it comes from outside the team doing the launching.

The mood inside

According to The Decoder, unease is building within the company. One source describes it as a "burbling sense of responsibility and dread" that OpenAI is not doing enough on safety.

That phrasing matters because of where the concern originates: not an outside critique, but the assessment of the people doing the work.

Two defensible views

There are two readings of the move, each internally coherent:

  • Dissolving is right — when safety is one team's job, other teams stop treating it as their responsibility. Spreading the work into product groups can make safety everyone's concern.
  • Dissolving is wrong — the separate team's real value was its independence. An assessment embedded inside a product team loses when it collides with that product's launch schedule.

Which applies depends on how the work was distributed: do the groups inheriting it have the authority to stop a release, or not?

The timing

The move lands while OpenAI prepares for a public listing and senior departures continue. In that context, dissolving a safety team stops being a standalone reorganisation and reads as a signal about priorities.

The company's account stresses that the work has not disappeared, only moved. Whether that holds will only be visible at the next major model release — who performed its risk assessment, and how much of the findings were published, is the real answer to the question.

A pattern across the industry

This is not the first case of a safety team being restructured. Over the past two years more than one major lab has seen a similar movement: a safety unit set up separately is later either dissolved or loses its key people.

There is no single explanation for that repetition, but it points to a structural tension. A safety team's job is by definition to slow things down: if a finding never delays a product, the finding has no practical consequence. In a market where speed is a competitive advantage, carrying a unit that permanently slows you down is organisationally hard.

The conclusion is not that companies act in bad faith, but that internal oversight is difficult to sustain without an external requirement. That is why regulators have recently concentrated on independent evaluation.

What to watch

  • Who performs the risk assessment at the next major model release.
  • Whether the findings are published, and how much of them.
  • Whether the group doing the assessment has the authority to stop a release.