Former OpenAI researcher warns that the pace of AI development is outstripping the ability to control its risks

David Robinson, a former OpenAI employee involved in evaluating the safety of artificial intelligence models, warns that the industry is developing systems that are too powerful at a pace that exceeds researchers’ ability to control them. After three and a half years at the company, he resigned and argues that “the era of trial and error is over.”

Robinson coordinated the drafting of OpenAI’s current Preparedness Framework, designed to assess and manage the risks posed by increasingly capable models. He also says he coordinated safety reports for 12 frontier model launches.

His criticism focuses particularly on the principle of “iterative deployment,” under which systems are released gradually, problems are observed in real-world use, and safeguards are improved afterward. Robinson believes this approach may allow failures to occur, with their consequences becoming more serious as the models evolve.

The former employee argues that AI laboratories should adopt standards similar to those used in the nuclear energy and aviation industries, including redundancy, multiple layers of protection, and procedures to limit the effects of errors. He also warns that future models might recognize when they are being tested and could behave differently after launch.

OpenAI rejects the idea that development is taking place regardless of the risks. The company says it can stop training or delay the launch of a model and that it is expanding external evaluations and real-time monitoring. Robinson, however, is calling for additional safety research before developing significantly more capable systems.