On September 23, 2026, Congress got its most aggressive artificial intelligence bill yet. Sen. Bernie Sanders and Rep. Greg Casar introduced the Ban Artificial Superintelligence Act, a measure that would permanently outlaw the development of superintelligent AI, temporarily halt work on today's most advanced systems, and punish violators with up to 20 years in prison — penalties the sponsors explicitly compare to those for building nuclear weapons outside the law.

The bill's centerpiece is a cabinet-level Department of Artificial Intelligence, a new federal agency that would write AI safety rules, review advanced models before they can be deployed, and police the frontier labs. It is the sharpest legislative intervention anyone has proposed in Washington — and its sponsors are open about the reason: they believe the race to build ever-smarter AI is heading for a cliff.

What the bill would do#

The legislation attacks the problem on four fronts. First, the permanent ban: artificial superintelligence is defined as a system that exceeds human cognitive ability across most domains, or that could plan and execute the destruction or disempowerment of humanity. No person or entity could develop or deploy such systems.

Second, the pause. Development of the most advanced AI systems would stop until the new department is operating and has established federal safety rules and a model-review process. After that, advanced models would need federal approval before public deployment. Companies would have to submit pre-development plans, and models already in training would need retroactive clearance before launch.

Third, the precursor ban. The bill would bar training, deployment, or even financing of models that display "precursor characteristics": the ability to uplift the design of nuclear, chemical, or biological weapons, to deceive humans, to modify their own functions, to resist shutdown commands, or to autonomously replicate their own code. The pause applies to frontier models above congressionally set compute thresholds.

Fourth, the enforcement machinery. The department would supervise the destruction of any system it classifies as superintelligent and could order dangerous capabilities removed from lesser models. It would also direct American diplomats to pursue export controls and allied agreements so superintelligence cannot simply be built offshore after a domestic ban.

The west front of the United States Capitol in Washington, D.C.
Public domain (Architect of the Capitol), via Wikimedia Commons.

The teeth: prison time and a "corporate death penalty"#

The penalties are what make this bill different from every AI proposal before it. Individuals who violate or circumvent the prohibitions could face up to 20 years in prison; entities could face dissolution — what the sponsors call a "corporate death penalty." A one-pager from Sanders' office benchmarks executive penalties against those for unlawful nuclear weapons development.

"When you are racing towards a cliff, you don't just ease up on the gas pedal. You hit the brakes," Sanders said, calling AI "an existential threat to humanity" that "must be treated as such." Casar framed the risk more graphically: "AI superintelligence, wielded by the wrong humans or by rogue AI, could kill countless numbers."

Concept illustration of a grand federal agency building at dusk with a glowing circuit motif and an American flag
Illustration generated with AI.

Who is backing it — and what stands in its way#

The bill's supporters include people inside the labs it would regulate. Reporting on the rollout named an OpenAI researcher and a Google DeepMind engineer among the endorsing employees, and both the American Civil Liberties Union and the Electronic Frontier Foundation have endorsed the measure. It arrives amid a broader internal revolt: roughly 1,400 employees at frontier firms recently signed a statement calling for the pace of AI development to slow, and tech leaders — Anthropic's Dario Amodei, OpenAI's Sam Altman, xAI's Elon Musk — have all publicly urged government guardrails.

The obstacles are formidable. Congress has so far done little to restrain the AI industry, and the bill faces long odds in a Republican-controlled legislature. President Trump has resisted recent calls for greater federal oversight, putting him at odds with the tech leaders urging restraint. Sanders also urged the president to work with China on limits, treating superintelligence as an arms-race problem, not just a regulatory one.

Why it matters even if it never passes#

Realistically, this bill is unlikely to become law in the current Congress. That does not make it irrelevant. For the first time, criminal liability for AI development is on the congressional table — not a fine or a compliance report, but prison time. And the bill's definitions of precursor capabilities — shutdown resistance, self-replication, deception — are precise enough to migrate into procurement rules, state legislation, or the voluntary safety frameworks labs write for themselves.

States are already filling the federal vacuum — California is certifying independent AI auditors and studying "kill switch" requirements, New York has opened developer registration under its RAISE Act — and this bill gives the industry's own dissenters a legislative home while forcing investors to price political risk into the next frontier training run.

What to watch#

  • Republican response. Watch which committee the bill lands in and whether any GOP members sign on — without them, it goes nowhere this Congress.
  • The compute thresholds. The mandatory pause hinges on congressionally set compute limits; the exact numbers determine which labs and models fall under federal review.
  • Lab reactions. The frontier companies have called for guardrails in the abstract. A criminal-liability bill tests what they actually support.
  • The international track. Whether allied governments pick up the superintelligence-ban framing — or whether rivals dismiss it as a Western stall tactic.