U.S. Congress pushes a ban on self-improving AI

The U.S. Congress is pushing to ban self-improving AI, according to a report from South Korea's Chosun Ilbo, moving a long-running safety debate into the legislative arena.

Self-improving AI generally refers to systems that can modify, optimize, or rewrite their own code or training processes without direct human intervention. Critics worry that once such a loop closes, capability gains could outrun the pace of human oversight and control.

The report centers on the legislative move itself: Congress pushing for a ban. The scope of any prohibition, which parts of the research pipeline it would cover, and how far the effort has advanced are not yet clear from the available detail.

If such a ban moves forward, it would reach directly into frontier training pipelines, where self-improvement and recursive self-improvement sit at the center of recent capability gains. It could also become a reference point for regulators in other jurisdictions.

Seen more broadly, the move reflects a shift in regulatory thinking, from abstract warnings about AI risk toward specific prohibitions on particular technical directions. Self-improvement is an unusually concrete target because it is a research direction rather than only a product feature.

What to watch next: the bill text, its sponsors, any hearings, and whether the effort draws bipartisan support. Those details will determine whether this is a policy statement or the start of binding law.