Sanders and Casar's superintelligence ban targets a future risk; the hard part is defining what it would actually stop

Senators Bernie Sanders and Greg Casar have introduced a bill to ban the development and deployment of superintelligent AI in the United States. According to the brief note we reviewed, it has three parts: a prohibition on superintelligence, a pause on advanced AI development, and a new federal department to regulate the sector.

Our source is a short video note, not the bill itself. We don't yet know how the text defines "superintelligent" or "advanced" AI, how long the pause would last, or what powers the new department would have. Those details will decide whether this is enforceable policy or a political statement. We won't guess at them.

Our thesis: the debate is worth having, but a ban defined by a word is not governance. "Superintelligence" has no agreed technical meaning. A law built on it has two bad options. It can draw the line so far out that it binds nothing today. Or it can draw it so loosely that it freezes useful work: models that speed up drug discovery, diagnostics or scientific research. We have argued before that the main risk in AI regulation is regulating panic instead of measured capability. A broad pause on "advanced AI" is the clearest example of that risk.

The proposal also has merits. A dedicated federal regulator speaks to a real gap. Today, oversight of frontier models depends heavily on voluntary commitments and on outside evaluators like the UK's AISI and METR. A body able to require testing, disclosure of internal models and incident reporting would be a meaningful step, if it is built around evidence.

Our reading: we don't expect this bill to pass as written. It matters anyway because it moves the Overton window, the range of ideas considered acceptable in mainstream politics. A ban on superintelligence is now a position sitting senators have formally proposed. The better outcome is that the debate pushes Congress toward capability-based rules: mandatory evaluations, thresholds tied to demonstrated risks such as cyber misuse or autonomous action, and transparency about what labs run internally. That is how society can manage short-term risks without giving up the long-term benefits, from curing diseases to broad abundance.