Human control reaches its limits as AI agents grow more autonomous

Warnings have emerged that AI is slipping out of human control. AI safety researcher Jeffrey Ladish said in an interview that "humanity does not have a practical strategy for controlling AI models and agents (programs that perform tasks by making their own judgments) that are becoming increasingly autonomous." Ladish is executive director of the AI safety research group Palisade Research and helped build the security team at AI developer Anthropic from September 2021 to October 2022. Ladish made the remarks to reporters after presenting at a Senate AI briefing for lawmakers hosted by Sen. Bernie Sanders (D-Vermont) at the U.S. Capitol in Washington, D.C., on September 16, 2026.

He said that during his time at Anthropic, employees were "quite concerned" about the direction of the technology, and that people he knows at OpenAI share the same view.

To those who doubt the pace of AI progress, Ladish pointed to mathematics. He said an AI agent solved the Navier–Stokes problem, a challenge humanity has failed to crack for decades. He noted that just three years ago, AI was only capable of solving high-school-level math problems.

As another example of the technology's rapid advancement, he cited video generation. People who laughed a few years ago at the distorted AI video of Will Smith eating spaghetti, he said, could be stunned by the realistic videos current models can now produce.