A proposed US ban on ‘superintelligence’ runs into a measurement problem before it reaches Congress

Sen. Bernie Sanders and Rep. Greg Casar are introducing legislation that would ban “AI superintelligence,” prohibit recursive self-improvement and create a Cabinet-level agency to supervise artificial intelligence. Casar argues that recent agent-security incidents make narrower proposals inadequate. The bill is unlikely to advance in the lame-duck session, but its sponsors want the issue embedded in the 2028 Democratic primary.

A ban needs an observable threshold. “Exceeds human capabilities” could describe a model that is superhuman at coding, persuasion or molecular design while remaining weak at many ordinary tasks. Recursive self-improvement is also not a binary property: laboratories already use models to write training code, generate synthetic data and optimize evaluations. The legally relevant question is when that assistance becomes an autonomous improvement loop with material capability gains and insufficient external control.

An agency could define tests, require notification of large training runs and restrict deployment after specific evaluations. Yet a broad statutory label risks either capturing routine research or waiting until the dangerous capability has already appeared. Compute thresholds are easier to audit, but increasingly poor proxies as algorithms and chips become more efficient. Capability thresholds are closer to the harm, but can be gamed and may not transfer across domains.

A separate proposal from Sen. Mark Kelly would make technology companies fund workforce development as AI changes employment. That is a useful contrast: transition policy addresses measurable present costs, while a superintelligence ban targets uncertain catastrophic risk. Congress may need both, but they require different institutions, evidence and enforcement rather than one elastic definition of AI danger.