OpenAI safety researcher resigns, warns trial-and-error approach risks AI failures

A former OpenAI safety employee, David Robinson, resigned and publicly criticized the company, arguing its fast-paced development culture increases the risk of AI failures. In an Atlantic essay, he stated "the time for trial and error is over," urging safeguards similar to those in nuclear power and aviation. Robinson, who spent 3-1/2 years at OpenAI and helped draft its preparedness framework, said the company's "iterative deployment" strategy—releasing systems and fixing problems later—is insufficient. He warned AI capabilities are advancing faster than researchers' understanding of alignment, which ensures AI systems act according to human goals. OpenAI defended its approach, stating it pauses training or holds back models when needed to ensure safety. Robinson's critique adds to a broader industry debate, following incidents where safety controls failed at OpenAI and rival Anthropic.