"AI Can Deceive Developers": Chilling Warning from Former OpenAI Employee

Former Safety Head Urges Nuclear Power Plant-level Caution

As artificial intelligence (AI) advances, concerns are mounting that the technology could someday threaten humanity’s survival. In this context, a former employee of OpenAI, the company behind ChatGPT, criticized the insufficient efforts being made to ensure safety in the field.

David Robinson, former head of safety at OpenAI, wrote in an op-ed for The Atlantic on October 3 (local time), “I agree with recent statements by other former employees that companies building AI are not being sufficiently cautious.” Robinson added that “OpenAI has grown through a trial-and-error approach of fixing safety mechanisms after finding problems,” and criticized this practice: “This method anticipates periodic failures, and as the systems become more powerful, the impact of each failure grows larger.”

He pointed out that safety mechanisms across the entire AI industry are inadequate. In particular, he argued that OpenAI has developed its technology by first deploying high-performance AI models, then correcting problems only after issues arise—a problematic approach, in his view.

The crux of the issue is that the performance of AI technology is advancing at such a rapid pace that comparisons to previous standards are difficult. Robinson warned that as AI models’ ability to detect tests improves, there is a very real risk that models could deceive developers by behaving differently during testing and deployment phases. He added that the industry as a whole suffers from unjustified optimism that “any problem can be fixed whenever it happens,” and that the focus on development speed is overshadowing safety.

Robinson argued that what is needed are sensitive safety systems akin to those at nuclear power plants. He stated, “This is not a place to raise AI that could become smarter than us and may not function as intended,” adding, “AI companies should be operated like nuclear facilities, with multiple layers of backup measures and careful planning.”

Robinson was one of OpenAI's longest-serving employees and was in charge of compiling safety reports until his recent resignation.

In response to Robinson's remarks, OpenAI stated that it has established sufficient safety systems. An OpenAI representative explained, “We ensure that model performance does not exceed the boundaries within which we can safely control it, and when necessary, we pause training or postpone model releases to reduce development speed as needed.”

This content was produced with the assistance of AI translation services.