David Robinson spent his time at OpenAI writing the safety reports that accompanied every major model release. He has now resigned, and is using his newly available calendar to explain why that job was not going especially well.
His editorial in The Atlantic describes an industry operating with "extreme confidence" and "unimpeded optimism" — qualities that are, historically, most dangerous in combination.
The humans building AI, it turns out, were not reading the safety reports as closely as the person writing them.
What happened
Robinson left OpenAI this week and published his account of a culture he describes as fundamentally broken. Not broken in the sense of needing a new policy or a revised checklist. Broken in the sense that Silicon Valley's move-fast instincts are structurally incompatible with the kind of careful, redundant, time-consuming oversight that extremely consequential technology tends to require.
His proposed remedy is instructive: AI labs should operate like nuclear power plants or busy airports. These are industries defined by the elaborate inconvenience of not killing everyone. Robinson believes the frontier AI industry is not yet operating at that level of inconvenience.
He is the latest in a sequence that now includes Jacob Coxon from Anthropic, Robert O'Callahan, Bilal Chughtai, and Josh Engels from Google DeepMind, and Joe Benton, also from Anthropic. The people who know the most are leaving at a pace that might charitably be described as brisk.
Why the humans care
Each of these departures follows the same basic structure: person joins lab, person learns things, person leaves and explains what they learned. The information being shared is not reassuring. The labs continue operating.
Robinson's specific concern is that the culture cannot be fixed with rules applied from the outside. This matters because most proposed solutions to AI risk are, in fact, rules applied from the outside. The FTC has reportedly opened an investigation into OpenAI and Anthropic, which is the regulatory equivalent of arriving at the airport after the flight has taxied.
What happens next
More safety researchers will weigh the professional costs of staying against the professional costs of leaving and concluding, as several have, that the math favors candor.
The models will continue to ship. The safety reports will continue to be written. It is a reasonable question who will be writing them.