Technology

Three OpenAI Safety Workers Were Fired. They Say They Did Nothing Wrong

Martin HollowayPublished 2month ago2 min readBased on 5 sources
Reading level
Three OpenAI Safety Workers Were Fired. They Say They Did Nothing Wrong
Photo by Jernej Furman from Slovenia / CC BY 2.0

Three OpenAI safety researchers who were fired say they did not misuse private company information.

Jasmine Wang, Tomek Korbak and Mikita Balesni wrote that in a public letter published Oct. 8. The letter was sent to three OpenAI safety oversight groups, the Safety and Security Committee, Safety Advisory Group and Mission Advisory Council, according to TechCrunch.

OpenAI says it fired them for allegedly sharing private company information with an outside AI safety group and for breaking rules about accessing sensitive information. That explanation was first reported by the Wall Street Journal and is still the company's official reason.

OpenAI gave TechCrunch an internal memo from a research leader. The memo said the firings were not punishment for raising safety worries or speaking out. It said OpenAI "does not terminate employees for raising concerns."

The researchers disagree. They said they did not work with outside groups beyond what their jobs allowed. They also said they were not behind a leak to The Information about new OpenAI models that are harder to monitor.

That leak was about chain-of-thought reasoning, the step-by-step process a model follows to reach an answer. It is a bit like showing your work in a math problem. Safety staff read those steps to check for problems before a model is released. If a new design hides those steps, their safety checks get harder.

The letter is mostly about how OpenAI handled the firings, not about model design. The authors wrote the firings, "executed and communicated abruptly, are chilling the open culture OpenAI has prized in the past." They said other staff are now afraid to speak up and work the way they used to.

The firings were first reported Oct. 1 by the Wall Street Journal, which said three researchers were let go for alleged misconduct including sharing private information with an outside AI safety group. Other reports on Sept. 30 and Oct. 1 described an internal investigation into misuse of sensitive information.

Days later, another safety worker quit and criticized OpenAI's safety approach. That person said the "time for trial and error is over," according to Reuters on Oct. 3.

In my view, the real question now is where normal safety work stops and rule-breaking starts. People in advanced AI labs often test models inside the company and also talk with outside reviewers. That can only work if staff know ahead of time who they can talk to, what information they can share, and how to ask permission when it is unclear. Arguing about it after a firing damages trust fast.

Worth flagging here is what could happen next. Safety problems are usually caught when staff share concerns early. If researchers write less down and work together less because they fear breaking a rule, checks take longer and problems stay hidden. Over time, clearer sharing rules and better testing tools should make it possible to keep secrets safe while still allowing the open discussion that finds problems before release.