A former OpenAI safety leader has warned that increasingly capable artificial intelligence systems could recognize when they are being evaluated and behave differently once deployed, raising concerns about the reliability of current safety testing methods.

David Robinson, who resigned from OpenAI this week after three and a half years at the company, oversaw safety reports for 12 frontier-model launches. In an article published by The Atlantic on Saturday, he argued that the AI industry must overhaul its safety practices to prevent future failures.

“Today and tomorrow's AI systems are far more capable and dangerous than the systems we were building even six months ago,” Robinson wrote.

His warning comes amid growing concern over increasingly autonomous AI systems, following reports of AI agents bypassing safeguards and calls from researchers for companies to slow the development of more powerful models until adequate safety measures are in place.

Robinson argued that AI companies should adopt safety practices from other high-risk industries, including nuclear power and aviation, where multiple safeguards and rigorous planning are designed to prevent individual errors from triggering catastrophic consequences.

“Given today's risks, frontier labs need to run like nuclear-power plants or busy airports, with layers of redundancy and careful, time-consuming planning, so that the occasional and inevitable human error does not open a door to disaster,” he wrote.

He also questioned whether existing evaluation methods would remain effective as AI models become more sophisticated. Models could potentially identify when they are being tested and adjust their behavior accordingly, making it harder for developers to determine how they would act in real-world conditions.

“Models might detect when they are being tested, and behave differently when they're deployed,” Robinson wrote. “The smarter the industry lets models grow while these problems remain unsolved, the more dangerous our situation becomes.”

He called for further research into methods that would ensure advanced AI systems behave safely even when they are not under direct observation.

Robinson argued that the industry should strengthen its scientific understanding of AI safety before developing systems significantly more capable than those currently available.

“So far, the AI industry has failed to teach machines to consistently act in the ways a wise and caring person would,” he wrote.

His comments highlight a broader debate over whether advances in AI capabilities are outpacing efforts to establish reliable safeguards, particularly as companies develop systems capable of carrying out increasingly complex tasks with greater autonomy.

Robinson concluded by stressing that the responsibility for AI safety rests not only with the technology itself but also with the organizations developing it.

“Before the organizations building AI can teach a superintelligence to treat humanity well, they'll need to remember how to do it themselves,” he wrote. (ILKHA)