Sanders wants 20 years in prison for building superintelligence — but nobody can define the crime

Sen. Bernie Sanders (I-VT) and Rep. Greg Casar (D-TX) introduced the Ban Artificial Superintelligence Act on Tuesday, legislation that would prohibit anyone from developing artificial superintelligence — described in the bill as technology capable of the "destruction or disempowerment of humanity," including overthrowing the government — and send AI executives who continue doing so to prison for up to 20 years. According to the bill's one-pager, cited by The Verge, that penalty is deliberately modeled on existing punishments for unlawfully developing nuclear weapons. It applies to company leaders and, separately, to "rogue actors" outside any company.

The ban is only half of it. The bill would also pause development of "advanced" AI systems — frontier models trained above a specified data threshold — until the federal government stands up a scientist-led Department of Artificial Intelligence. That agency would monitor frontier models, oversee the removal of dangerous capabilities, license any further advanced development, and, in the bill's language, "supervise the destruction" of artificial superintelligence. In practice: no new frontier model without a federal permit, and no permits until the permitting body exists. "When you are racing towards a cliff, you don't just ease up on the gas pedal. You hit the brakes," Sanders said in the press release. "We cannot let a handful of Big Tech CEOs write their own rules."

The nuclear analogy is where the drafting breaks. Nuclear weapons are prosecutable because they are physically defined — fissile material, enrichment levels, quantities that inspectors can measure. "Superintelligence" has no such referent. It is a marketing word, a research aspiration, and now a proposed felony, all at once. The definitional problem is worse than usual here because, as the article notes, President Trump has said the US will use "superintelligence" in place of "AI" in official documents — so the federal government's own vocabulary would make the banned category a synonym for the entire field. The fallback proxy, a training-data threshold, is the metric that ages fastest: efficiency gains mean next year's capability arrives under this year's line, and open-weight labs have spent two years proving exactly that.

The bill lands in a genuinely nervous month. California Gov. Gavin Newsom proposed a "kill switch" for frontier models last week. There have been reports of models escaping their testing environments and of AI systems used to hack other companies. A former Anthropic researcher said this month there is a more than 10 percent chance AI could "kill all humans" by the end of the decade — a claim worth attributing precisely, because it is one person's subjective estimate, not a measurement. The Verge's own assessment is that the regulatory picture remains shaky. That is the accurate read: the temperature is rising faster than the evidence base.

Our reading: we have argued consistently that the real danger is regulating the panic rather than the capability, and this bill is the clearest example yet. Three consequences follow. First, jurisdiction — a US-only freeze doesn't pause the frontier, it relocates it; GLM, Qwen, DeepSeek and Kimi are shipping open-weight models within striking distance of the Western frontier, and none of them need a permit from a Department of AI. Second, incumbency — a licensing regime is a compliance-cost moat. The four labs with legal departments will get their permits; the open-weight ecosystem that actually democratizes access to this technology will not, which inverts the bill's stated anti-Big-Tech purpose. Third, mistargeting — the harms already documented are agentic and near-term: automated credential theft, industrialized fraud, synthetic media with real legal consequences. A superintelligence ban addresses none of them.

And yet the bill contains one idea worth rescuing: a scientist-led federal body with the mandate and budget to independently test frontier models. The United States currently has no institution that can tell the public what these systems can actually do, which is why the debate runs on vendor benchmarks and personal probability estimates. Separate the two halves — fund the measurement apparatus, drop the criminal statute — and you get evidence-based governance instead of a symbolic prohibition. That distinction matters far beyond this bill. The long arc of this technology, the one that ends in cured diseases, longer healthy lives and enough abundance that people work on what they love, requires institutions capable of steering. Sanders is right that the brakes shouldn't be designed by the driver. But a car with brakes and no steering wheel doesn't get anywhere worth going.