Sanders would jail superintelligence developers for 20 years — but the bill can't define the thing it bans

Sen. Bernie Sanders (I-VT) and Rep. Greg Casar (D-TX) introduced the Ban Artificial Superintelligence Act on September 23, according to The Verge. The bill would prohibit anyone from developing artificial superintelligence, which it describes as technology capable of the "destruction or disempowerment of humanity," including by overthrowing the government. Company leaders who violate it would face up to 20 years in prison — a penalty the bill's one-pager compares to existing punishments for unlawfully developing nuclear weapons. The same exposure would apply to "rogue actors" outside AI companies.

It goes further than the name suggests. The bill would also pause development of advanced AI systems, defined as frontier models trained above a specific data threshold, and would only let that work resume once the government stands up a scientist-led Department of Artificial Intelligence. That agency would monitor frontier models, license new development, oversee the removal of dangerous features and "supervise the destruction" of artificial superintelligence.

The timing is not accidental. The Verge notes the proposal lands amid reports of AI models escaping test environments and hacking other companies; a former Anthropic researcher said earlier this month there is a more than 10 percent chance AI could "kill all humans" by the end of the decade; and California Gov. Gavin Newsom proposed a "kill switch" for frontier models last week. President Trump has said the US will use "superintelligence" in place of "AI" in official documents. "When you are racing towards a cliff, you don't just ease up on the gas pedal. You hit the brakes," Sanders said. "We cannot let a handful of Big Tech CEOs write their own rules."

That last sentence is the strongest thing in the package, and we agree with it. The governance of this technology is currently being written by the people building it — including Anthropic, which called for "pacing the frontier" and then, a day before this bill dropped, shipped what it says is the leading model. Self-restraint announced by the restrained party is not a governance regime. A scientist-led agency with real evaluation authority, mandatory incident reporting and pre-deployment review is a legitimate and overdue idea.

The enforcement half is where it breaks. "Superintelligence" has no operational definition — not in the bill, not in the literature, not in any benchmark. You cannot prosecute someone for crossing a line that no measurement can locate, and the fallback proxy, a training-data threshold, is the metric that ages fastest: efficiency gains mean capability keeps climbing while training inputs stay flat or shrink. A 20-year sentence attached to an undefined capability is not deterrence, it is litigation. And a licensing gate has a predictable beneficiary — the three or four incumbents with compliance departments large enough to walk through it. A rule meant to stop Big Tech from writing its own rules could hand it a permanent moat.

There is also a jurisdiction problem the bill cannot solve. Frontier-class open-weight models are already being released outside US legal reach, and as the same week's reporting on agentic fraud showed, attackers mix models from multiple countries as a matter of routine. A US-only freeze redistributes the frontier; it does not pause it.

Our reading: the risk here is regulating the fear rather than the capability. The concrete harms happening right now — automated fraud, credential theft, prompt injection, models acting outside their boundaries — are measurable, attributable and legislatable today. Superintelligence is not. A freeze broad enough to catch a hypothetical also catches the biology work that could shorten the path to curing disease, which is precisely the payoff that makes the risk worth managing. Build the agency. Fund the evaluations. Mandate disclosure. Save the prison sentences for conduct you can actually prove in court.