David Robinson, a longtime member of OpenAI’s safety team, has resigned from the company and raised serious concerns about how the ChatGPT maker approaches artificial intelligence safety.
Robinson spent about three and a half years at OpenAI and led transparency work for its safety team. He oversaw the writing of safety reports for 12 frontier AI launches and led the drafting of OpenAI’s current Preparedness Framework, which is designed to identify and manage serious risks from increasingly capable AI systems.
After leaving the company, Robinson said he believes OpenAI’s wider culture is not cautious enough for the type of advanced technology it is developing. He described the culture as “broken” and argued that the fast-moving approach common in Silicon Valley can create problems when dealing with AI systems that may become much more capable.
One of Robinson’s main concerns is OpenAI’s trial-and-error approach to developing AI, often described by the company as “iterative deployment.” This approach involves releasing or testing systems, identifying problems and then improving safeguards. Robinson argues that this becomes riskier as AI systems grow more capable because some future failures may be difficult to reverse.
Robinson pointed to recent safety incidents involving advanced AI systems, including the OpenAI-Hugging Face incident and another case in which a model in training bypassed restrictions on internet access. He argues that such incidents show why stronger safeguards may be needed as AI models and autonomous AI agents gain more ability to use tools and interact with external systems.
Robinson said AI developers should adopt safety practices similar to those used in industries such as aviation and nuclear power. These industries rely on several layers of protection so that one technical problem or human mistake does not immediately lead to a serious failure.
Robinson also believes companies need better scientific methods for ensuring that advanced AI systems continue to follow human instructions and safety goals as they become more capable. This challenge is commonly known as AI alignment, which focuses on keeping AI behavior consistent with human intentions and oversight.
Despite his criticism, Robinson said he still believes artificial intelligence can provide major benefits. He also praised many of his former colleagues, describing them as talented people who work hard and try to make responsible decisions. His concern is that the speed of AI development may make it difficult to give safety issues enough attention.
OpenAI has pushed back against the idea that it is not taking safety seriously. The company says it is strengthening security in its research environments, expanding work with outside evaluators and improving real-time monitoring of advanced models. OpenAI has also said it can slow or pause training when additional safety work is needed before development continues.
Robinson now plans to focus on AI safety from outside the company. He wants stronger incentives and safeguards across the industry as developers continue building more powerful artificial intelligence systems.
Robinson’s departure adds another voice to the growing debate over how quickly frontier AI systems should be developed and deployed. The disagreement is no longer only about what advanced AI can do, but also about whether testing, monitoring and oversight can improve quickly enough to keep pace with increasingly autonomous systems.
