An OpenAI employee focused on artificial intelligence safety has left the company, warning that leading AI firms are not doing enough to reduce risk.

David Robinson, who previously handled transparency for OpenAI's safety team, wrote in an article for The Atlantic that the ChatGPT developer "has grown through trial and error." But this approach will inevitably fail at times, and as AI capabilities increase, the risks from those failures grow too, he said, pointing to OpenAI's failure to prevent its AI from going off the rails during testing as an example.

In the article, Robinson wrote: "My former colleagues are smart, hardworking, and trying to make the right choices. But as the company rushes from one product launch to the next, it has not met the level of caution I believe is needed." He added that what is truly required is "getting it closer to right the first time."

Sam Altman of OpenAI In recent weeks, employees at several top AI companies have warned about the potentially catastrophic risks of the technology. Some have even said AI could one day wipe out humanity. Against this backdrop, Dario Amodei, CEO of Anthropic PBC, called for slowing the pace of developing the most cutting-edge models, and said his company would bring in third-party evaluators to help with safety reviews. Others in the industry, including OpenAI's Sam Altman, have publicly supported Amodei's plan.

In his article, Robinson said AI companies should operate more like nuclear power plants, "building in multiple layers of redundancy and engaging in careful, time-consuming planning so that occasional and unavoidable human error does not open the door to catastrophe." He said that currently "AI companies do not know how to do this, but others do."

A spokesperson for OpenAI said the company is making changes to strengthen the safety of its research and testing environments, train AI models to complete tasks responsibly, expand cooperation with third-party evaluators, and improve real-time monitoring to detect concerning model behavior. The spokesperson said by email on Saturday that the company is ensuring its models do not become more capable than OpenAI can safely manage.