Former OpenAI Safety Lead Says Company's AI Safety Talk Doesn't Match Actions

David Robinson, a former co-leader of OpenAI's safety systems team, spoke out after leaving the company. He believes OpenAI is inconsistent on AI safety: it keeps warning about the risks of stronger new models and talks about slowing down, yet still pushes ahead at full speed behind the scenes.

"We keep putting out these warnings, but in the end we still train and deploy the very models we ourselves call dangerous," he said. In July, hundreds of OpenAI employees signed an open letter urging the government to step in and slow down AI development. In September, CEO Sam Altman agreed the pace of frontier AI needs to be controlled.

Robinson described the office as still running at full speed, with many people working to the point of burnout. The process of AI improving its own capabilities is called "recursive self-improvement," and OpenAI's chief scientist Jakub Pachocki has warned that "extreme caution is needed right now."