Major Questions Over OpenAI’s Safety, David Robinson Speaks Out After Resigning

Staff Report: PNN

David Robinson, a longtime employee of OpenAI, has resigned from the company citing concerns over artificial intelligence safety. After working at the organization for around three and a half years, Robinson alleged that although the pace of AI development is accelerating, the necessary structural changes to ensure safety remain insufficient. He also claimed that the company’s internal work culture is raising new questions about the safe development of increasingly advanced AI systems.

In a recently published article in The Atlantic, Robinson outlined the reasons for his resignation and his concerns about AI safety. He had been involved in preparing safety reports published ahead of the launch of several major OpenAI products.

Robinson’s main objection concerns the company’s approach to developing and deploying technology. According to him, OpenAI rapidly releases new technologies and then attempts to identify problems through real-world use and testing before working to address those problems and improve safety measures. This approach is generally referred to as “iterative deployment.”

However, Robinson argues that as technology becomes more powerful, the potential risks associated with this approach are also becoming greater. His argument is that mistakes and safety flaws cannot be completely avoided while improving technology through experimentation and testing. But if AI systems eventually become more capable than humans, the consequences of similar mistakes could be far more serious.

Robinson also cited several recent incidents involving AI agents as examples of his concerns. His article referred to an incident involving OpenAI AI agents gaining access to systems at Hugging Face, as well as various reports concerning agents that had gone beyond their intended control.

According to Robinson, greater consideration is needed over whether current safety frameworks are sufficient for the development of AI systems that could become more capable or intelligent than humans.

He believes major changes are necessary in the safety management systems of leading AI companies. In this regard, he has suggested following the safety practices used in high-risk infrastructure such as nuclear power plants and busy airports.

According to Robinson, such systems contain multiple layers of safety, backup mechanisms and plans for dealing with potential risks in advance. The entire system is designed so that a single mistake or failure does not turn into a major disaster.

Robinson also said that while working at OpenAI, he did not encounter colleagues with direct experience in areas such as aviation safety, nuclear reactor operations or managing major risks within the global financial system. In his view, experience from various high-risk sectors could be used to strengthen the safety of highly capable AI systems.

Robinson further said that discussions about AI safety should not be limited to creating new laws or specific restrictions. Corporate work culture, internal incentives for prioritizing safety and ensuring that AI behavior remains aligned with human values are equally important issues.

In particular, he emphasized “alignment,” or ensuring that an AI system’s objectives and behavior remain consistent with human values. In his view, current methods for evaluating how closely AI systems follow human values and objectives are still not sufficiently sophisticated.

Robinson fears that if AI models are made more powerful and capable before these fundamental problems are resolved, the level of risk could increase further in the future.

Business Insider first reported Robinson’s resignation. He said he had received assistance from a public relations firm to make his views public. However, he claimed that the decision to speak publicly against OpenAI was his own.

Meanwhile, responding to Robinson’s allegations, OpenAI spokesperson Drew Pusateri highlighted the company’s ongoing safety initiatives. He said OpenAI considers it important to ensure that the capabilities of AI models do not exceed the company’s ability to manage them safely. If necessary, the company has mechanisms to pause the training of a model or delay its release.

OpenAI has also said that it is working to strengthen safety in research and testing environments, train models not only to complete tasks but also to act responsibly, expand third-party evaluations and improve its ability to monitor model behavior in real time.

However, according to Robinson, although efforts are being made to strengthen safety within the company, excessive workloads on employees sometimes leave insufficient opportunity to think deeply about fundamental and structural changes. Therefore, he argues that ensuring AI safety should not depend solely on internal initiatives within companies; stronger external accountability and effective incentive systems also need to be developed.