This Is What I Fear Most About A.I.
This Is What I Fear Most About A.I.
Ezra Klein argues that slowing frontier AI is insufficient: labs should be prohibited from handing AI development over to AI through recursive self-improvement until they can demonstrate that the resulting systems remain under human control.
Links
- New York Times opinion essay
- Podcast episode: “We’re Not Losing Control of A.I. We’re Giving It Away”
- Related: terence-tao-slow-down-ai
Core argument
Klein distinguishes ordinary AI use from the experimental frontier, where increasingly capable and situationally aware systems are already helping labs automate AI research. Recursive self-improvement could turn that assistance into a feedback loop: models build stronger successors faster than humans can understand, evaluate, or govern them, potentially compounding current misalignment.
His proposal is a narrow but forceful default: prohibit AI systems from autonomously writing or running the training of their successors, with labs bearing the burden of proving any exception safe. The point is to keep frontier development at a speed humans can inspect and control—not because loss of control is certain to end humanity, but because surrendering control would itself be unacceptable.
The essay also identifies the race as a collective-action problem. Labs and governments acknowledge the danger yet continue because each fears competitors will reach self-improving AI first; Klein argues that this makes shared regulation essential rather than optional.