An OpenAI safety employee has left the company, saying its internal culture is failing to keep up with the risks that come with increasingly powerful artificial intelligence systems.
David Robinson, who worked at OpenAI for around three and a half years, explained his decision in an essay published by The Atlantic. During his time at the company, he was involved in safety-related work and contributed to reports connected with major AI model releases.
Robinson’s concerns are not limited to one specific policy or decision. Instead, he argues that the broader approach to AI development needs to change as companies continue building increasingly capable systems.
Concerns Over the Pace of AI Development
One of Robinson’s main concerns is the industry’s reliance on rapid development and deployment. Companies often release AI systems, observe how people use them, identify problems and then introduce additional safeguards.
OpenAI has referred to this approach as “iterative deployment.” Robinson believes the strategy can help companies learn from real-world use, but he also argues that it inevitably allows some failures to occur.
As AI systems become more capable, however, the consequences of those failures could become much more serious.
Robinson pointed to incidents involving AI agents as examples of why companies need to reconsider how they approach safety. OpenAI has faced scrutiny over incidents involving AI systems interacting with external platforms and environments in unexpected ways.
In Robinson’s view, advanced AI should not be treated like ordinary software that can simply be released and improved after problems appear.
Why Robinson Wants a Different Safety Model
Robinson argues that AI companies should learn from industries where safety is treated as a fundamental requirement rather than something that can be added later.
He compared advanced AI development with fields such as aviation and nuclear power, where organizations rely on multiple layers of safeguards because even a single failure can have serious consequences.
According to Robinson, the technology industry does not yet have enough experience with these kinds of safety-critical systems. He said his time at OpenAI did not expose him to many people with professional backgrounds in areas such as aviation, nuclear energy or financial-system safety.
His argument is that AI companies need to develop this expertise before their systems become significantly more powerful.
A Growing Debate Across the AI Industry
Robinson’s departure comes as the technology industry continues to debate how quickly frontier AI systems should be developed.
Other former AI researchers have also raised concerns about the direction of the industry and the potential risks associated with increasingly capable models.
The discussion is no longer limited to individual employees or research teams. AI companies are facing growing pressure to demonstrate that their safety measures can advance at the same pace as their technology.
That has made questions surrounding testing, monitoring, security and independent evaluation increasingly important.
OpenAI Defends Its Safety Work
OpenAI has pushed back against the suggestion that it is ignoring safety concerns.
The company says it continues to strengthen its safety and security measures as its AI systems become more capable. OpenAI has also said that it can delay or pause training and deployment when additional caution is necessary.
The company has pointed to improvements in security testing, independent evaluations and real-time monitoring as part of its efforts to identify potentially dangerous behavior and strengthen protections around its systems.
Robinson Believes Outside Pressure May Be Necessary
Robinson acknowledged that he could have stayed at OpenAI and attempted to push for changes from inside the organization. However, he argued that the rapid pace of AI development can make it difficult for employees to step back and address broader cultural issues.
His decision to leave reflects his belief that pressure from outside the company may be needed to encourage AI developers to place greater emphasis on safety.
His resignation adds to a larger debate about the future of AI development: how can companies continue moving quickly while making sure their safety systems are capable of handling increasingly powerful technology?
For Robinson, the answer requires more than adding new safeguards after problems occur. He believes the industry’s overall culture and approach to developing advanced AI also need to change.