Pacing the frontier isn't enough, Ezra Klein argues. Someone needs to stop it.
In a New York Times opinion piece published September 20, Klein calls on AI labs to halt recursive self-improvement. That's the process where AI systems train and build their own successors. Klein says it shouldn't continue until it's proven safe. "Walking quickly off a cliff is only marginally better than sprinting off one," he writes.
The Case for Stopping, Not Slowing
Klein's argument leans on Anthropic's own data. In February 2025, almost none of the code added to Anthropic's codebase was written by Claude. By May 2026, that figure passed 80%. Anthropic now says Claude leads 26% of the company's research and development tasks outright, not just assists with them.
"Misalignment present in today's models could compound as the models build their successors, growing more frequent but less understood until we lose control of them," Anthropic has warned in its own research.
Klein's point: the labs most worried about losing control are also racing hardest to give it up.
OpenAI, for its part, says it has already built an automated AI "research intern."
A fully automated AI "researcher" is expected by March 2028. Yet the company has also said plainly it doesn't yet know how to safely reach full recursive self-improvement.
Why Miami Researchers Say Alignment Gets Harder, Not Easier
Or Cohen-Sasson directs the Miami Law and AI Lab at the University of Miami. His research examines how AI alignment breaks down as models get more capable.
"The smarter the model, the harder it is to make it align with our instructions," Cohen-Sasson has said of his lab's work.
That's the exact mechanism at the center of Klein's argument, capability and controllability moving in opposite directions.
Cohen-Sasson's lab has also studied a regulatory approach Klein's essay doesn't directly address. It's one Congress would need to answer if it acted on his call. Current proposals often set safety thresholds based on how much computing power went into training a model. Cohen-Sasson's research found that approach can be circumvented. Smaller models working together can outperform a single larger model while staying under the threshold entirely.
The Regulatory Gap
Klein's core objection isn't philosophical. He argues no comparable oversight exists for AI labs that exists for ordinary construction.
"You cannot build an eight-story apartment building without an agonizing public review process," he writes, "and yet somehow it is possible for these labs to unleash a swarm of 40,000 AI agents to build a society-altering superintelligence without so much as a hearing."
For a Miami-based lab already probing how regulatory thresholds fail in practice, that gap isn't an abstract policy complaint. It's the specific research question Cohen-Sasson's work is already trying to answer, well before Congress takes up anything resembling Klein's proposal.