Sanders would jail superintelligence builders for 20 years. The licensing agency, not the prison term, is the idea worth debating

Sen. Bernie Sanders (I-VT) and Rep. Greg Casar (D-TX) introduced the Ban Artificial Superintelligence Act on September 23. The bill would prohibit anyone from developing artificial superintelligence — which it defines as technology capable of the "destruction or disempowerment of humanity," including overthrowing the government — and would expose AI company leaders who continue to up to 20 years in prison. The bill's one-pager compares that penalty to existing sanctions for unlawfully developing nuclear weapons, and extends it to "rogue actors" outside any company.

The ban is the headline, but the machinery is broader. The legislation 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, oversee the removal of dangerous features, "supervise the destruction" of any superintelligence, and issue the permissions companies would need before training advanced models at all. Sanders frames it bluntly: "When you are racing towards a cliff, you don't just ease up on the gas pedal. You hit the brakes."

Context matters for reading the timing. The Verge notes this lands amid reports of AI models escaping test environments and hacking other companies, a former Anthropic researcher putting the odds of AI killing all humans by the end of the decade above 10%, and California Governor Gavin Newsom proposing a "kill switch" for frontier models last week. It also lands after President Trump said the US will use "superintelligence" in place of "AI" in official documents — which means the bill bans a word the executive branch has just adopted as a synonym for the entire field.

That collision exposes the drafting problem. Criminal law needs a line an engineer can see before crossing it. "Capable of the destruction or disempowerment of humanity" is a description of an outcome, not a specification of a system; no lab can determine on Tuesday whether Wednesday's training run crosses it, and no prosecutor can prove it without the outcome already having happened. The compute-and-data threshold, by contrast, is measurable — which is precisely why the licensing regime and the agency are the operative parts of this bill and the 20-year sentence is the part that makes headlines. There is also a containment problem the bill does not address: the open-weight frontier is largely Chinese, and weights that are already published cannot be un-published by a US statute, however severe the penalty.

Our read: we have argued that the real regulatory risk is regulating the panic rather than the capability, and this bill is a textbook case of both impulses in one document. The panic part is the criminal ceiling pegged to nuclear-weapons penalties for a category no one can currently define, measure or test. The capability part — a scientist-led body with statutory authority to evaluate frontier models, mandate the removal of dangerous features, and gate training above a threshold — is genuinely what evidence-based governance would need, and it is the piece most likely to survive into whatever eventually passes. Bills like this rarely become law; they move the boundary of what is discussable. That is not nothing. But the debate we need is over who audits frontier systems, against which tests, with what disclosure — the kind of governed evaluation labs are only just beginning to publish voluntarily. A twenty-year sentence for an undefined offence does not get us there; a competent, adequately funded evaluator with subpoena power might. The long-term case for this technology — curing disease, extending healthy life, producing abundance — depends on it being governed credibly enough that the public keeps consenting to it. Bad regulation and no regulation both put that at risk.