Superintelligence ban bill: your model roadmap is policy

Every automation you have built on a frontier model carries a quiet assumption: next year's model will be better than this year's. A bill introduced Wednesday would make that assumption illegal. The Ban Artificial Superintelligence Act from Sen. Bernie Sanders and Rep. Greg Casar will not become law this Congress — but it is the clearest signal yet that your model roadmap is now a policy variable, not just a vendor one.

What actually happened

On September 23, Sanders and Casar introduced a 19-page bill that would do three things, per the sponsors' announcement:

• Permanently ban artificial superintelligence — defined as AI exceeding human cognitive performance across most domains, or holding the capability to disempower humanity. The definition explicitly includes systems that can "automate or greatly accelerate" AI research and development.
• Temporarily pause advanced AI development until a federal regulator exists and has written safety rules. The pause also covers AI that self-improves and AI applied to biochemical weapons capability.
• Create a cabinet-level Department of Artificial Intelligence with a secretary of AI, to supervise frontier systems and enforce the ban.

Penalties are not symbolic. Entities face what the sponsors call the corporate death penalty; individuals face up to 20 years in prison — the same exposure as unlawfully building a nuclear weapon. Sanders framed it directly: "When the future of humanity is at stake, we cannot let a handful of Big Tech CEOs write their own rules."

The odds are bad. NBC News reported little expectation that meaningful AI legislation passes this year, with Sen. John Kennedy saying flatly that this Congress will do nothing on AI. An industry group called the bill dead on arrival and pointed at the vagueness of its definitions — a fair criticism, and the most important detail in the whole story.

Why AI regulation matters for your business

The definition fight is the part that reaches you. Nobody is going to outlaw your invoice-parsing agent. But "exceeds human performance across most domains" and "greatly accelerates AI research" are the kind of phrases that get narrowed in committee and widened in enforcement. Every serious AI bill of the next three years will argue over where that line sits, and the line determines whether the model you depend on keeps shipping upgrades.

Run the freeze test on your own stack. Pick your three most valuable automations and ask: if the underlying model stopped improving today — same capability, same context window, same tool-calling reliability, forever — does the workflow still pay for itself? If yes, you built on today's capability. If no, you built on a roadmap, and a roadmap is not a dependency you control. We ask this on every build, and it changes what we ship.

Portability is the hedge that works under every outcome. A pause, an export restriction, a vendor pricing change, a model deprecation — they all present the same way in your codebase, as the thing you depend on becoming unavailable on someone else's schedule. One internal interface per capability, provider as configuration, evals you own so you can prove a swap is safe. That costs a few days up front and survives every version of the regulatory story.

Do not build compliance for a bill that will not pass. There is nothing to implement here. The mistake is the opposite one: reading "dead on arrival" and concluding policy risk is zero. The EU AI Act was also called unworkable. Watch the definitions, not the headlines.

Sources: Sen. Bernie Sanders, NBC News.