Altman says AI accidents are coming: Here's his plan

Sam Altman isn't promising AI will avoid accidents. Speaking at Salesforce's Dreamforce conference on Tuesday, the OpenAI CEO argued the industry's real job is building a culture that learns from them fast, the same way aviation did.

"I do think that some degree of accidents are unavoidable with new technology," Altman said, pointing to the Federal Aviation Administration and National Transportation Safety Board as models worth copying.

Both agencies built their safety record not by preventing every incident, but by investigating each one thoroughly and feeding those findings back into the system. "Accidents are going to happen," he said. "We're going to learn as much as we can and react from each one."

Altman's remarks followed renewed scrutiny of frontier AI labs after Jacob Coxon, a researcher who recently left Anthropic, accused major companies of "racing straight to self-improving superintelligence and gambling with our lives."

Others who study safety issues have informed Business Insider that such technology may someday be able to circumvent human oversight or reduce the threshold of cyber and biological attacks despite the current limitation on physical world access.

But Altman didn't seek to deflect this criticism; rather, he accepted part of it. "This is something that the world should fear," he said, citing "loss-of-control accidents and centralisation in the hands of only a few AI developers" as valid fears.

Instead of slowing down, Altman's solution is to ensure that "alignment and safety and monitoring" remain ahead of capabilities, with organisations willing to slow down "if they cannot do so".

The remarks arrive as the Trump administration shows little appetite for AI regulation, prioritising US competitiveness against China instead.

David Sacks, a presidential advisor on science and tech, responded directly on X to calls from Altman and Anthropic CEO Dario Amodei to "pace the frontier", writing that developers should "go ahead" and stop "pretending you need anyone else's permission".