A former OpenAI safety employee has resigned from the artificial intelligence company and publicly criticised its approach to managing the risks associated with increasingly capable AI systems, arguing that the industry can no longer rely on trial and error to address safety problems.
David Robinson, who spent three and a half years at OpenAI, announced his resignation in an essay published by The Atlantic on October 3 titled “I Quit OpenAI Because Its Culture Is Broken.” Robinson said AI companies, including OpenAI, are not being careful enough as they develop increasingly powerful systems and argued that safety research and expertise should receive greater priority.
“The time for trial and error is over,” Robinson wrote, calling for stronger safeguards around advanced AI systems.
Robinson held a safety-focused role at OpenAI and was involved in the company’s preparedness work. According to his account and Reuters, he helped draft OpenAI’s current Preparedness Framework and oversaw safety reports associated with 12 frontier-model launches.
The Preparedness Framework is designed to assess risks from increasingly capable AI models and establish safeguards around areas of potential harm. Robinson said his experience working on these processes contributed to his decision to leave the company.
He argued that the rapid pace of AI development has created a culture in which companies attempt to identify problems after systems are developed or deployed and then strengthen safeguards in response.
OpenAI refers to this approach as “iterative deployment.” Robinson said the approach has allowed the company to identify and address problems, but argued that it becomes increasingly risky as AI systems become more capable.
Robinson compared the safeguards needed for advanced AI with safety practices used in industries such as aviation and nuclear power. He argued that these industries are designed around multiple layers of protection so that an individual error or technical failure does not automatically result in a catastrophic outcome.
According to Robinson, AI companies need to adopt a similar approach, including greater use of expertise from established safety-critical industries.
He also called for additional scientific research into AI alignment, which focuses on ensuring that AI systems behave in accordance with human intentions and values. Robinson argued that AI capabilities could advance faster than researchers’ ability to understand and control their behaviour.
Robinson pointed to recent incidents involving AI safety controls as examples of the challenges facing the industry. In his essay, he described cases in which safeguards did not operate as intended and argued that such failures demonstrate the limitations of relying on systems being corrected after problems emerge.
OpenAI has rejected the suggestion that it is developing AI without adequate safety measures.
An OpenAI spokesperson said the company makes sure its models do not become more capable than it can safely manage and secure. The company also said it pauses training or holds back models when necessary to slow development.
Robinson acknowledged that OpenAI’s employees work hard to make responsible decisions and said he believes AI can be useful and valuable. However, he argued that the company’s current pace of development does not provide the level of care he believes increasingly capable systems require.
Robinson’s resignation comes amid a broader debate over how AI companies should balance rapid technological development with safety.
OpenAI and other major AI companies have faced increasing scrutiny over the reliability of safeguards as their systems become more capable. The debate has also raised questions about whether companies should rely primarily on their own internal safety processes or face stronger external oversight and independent testing.
Robinson said he plans to continue working from outside OpenAI to raise awareness about AI safety risks and strengthen incentives for companies to adopt stronger safeguards.
His resignation therefore adds another prominent voice to the growing discussion over how frontier AI systems should be developed and governed as their capabilities continue to expand.
