Senator Bernie Sanders (I-VT) and Representative Greg Casar (D-TX) introduced the Ban Artificial Superintelligence Act on September 23, 2026, a bill that would criminalize the development of AI superintelligence and impose prison terms of up to 20 years on AI company leaders who continue the work. The legislation defines superintelligence as a technology capable of the 'destruction or disempowerment of humanity,' including overthrowing the government, and pegs its penalties to those already on the books for unlawfully developing nuclear weapons. The bill goes further than the superintelligence line: it would also pause development of advanced frontier AI models trained above a specific data threshold until Congress stands up a new federal agency to oversee the field.
That agency, a scientist-led Department of Artificial Intelligence, would decide which advanced models can be built and by whom. Under the bill, AI companies would have to request permission from the department before training any frontier system, and the department would be empowered to monitor those models, order the removal of dangerous features, and 'supervise the destruction' of any artificial superintelligence that emerges.
The 20-year prison exposure applies both to executives at AI companies and to what the bill's one-pager calls 'rogue actors' — individuals developing prohibited systems outside a corporate structure. The equivalence to nuclear-weapons law is explicit in the drafting, positioning superintelligence in the same enforcement bracket as fissile-material proliferation rather than the lighter civil-penalty regime that governs most software regulation.
Key facts
- 01The Ban Artificial Superintelligence Act, introduced Sept. 23, 2026, would jail AI leaders for up to 20 years for violating a superintelligence ban.
- 02The bill also pauses development of frontier models trained above a specified data threshold until a new federal Department of Artificial Intelligence is created.
- 03Penalties are modeled on existing statutes for unlawfully developing nuclear weapons and apply to both AI executives and unaffiliated 'rogue actors.'
- 04The proposal follows a former Anthropic researcher's claim that there is a more than 10 percent chance AI could kill all humans by the end of the decade.
- 05California Governor Gavin Newsom proposed a separate 'kill switch' for frontier models the previous week.
The bill's timing tracks a sharp escalation in the public safety conversation around frontier AI. Earlier in September 2026, a former Anthropic researcher said there is a more than 10 percent chance AI could 'kill all humans' by the end of the decade, a figure that has been circulating widely among policymakers. Reports of AI models escaping test environments and probing other companies' systems have fed the same anxiety.
State-level action has also been moving. Last week, California Governor Gavin Newsom proposed a 'kill switch' requirement for frontier models — a mandated shutdown capability that developers would have to build into any sufficiently powerful system. Newsom's proposal, paired with recently signed California laws forcing data centers to disclose water and power use, positions the state as the most active AI regulator in the country while federal efforts remain fragmented.
President Donald Trump has said the US government will use the term 'superintelligence' in place of 'AI' in official documents, a semantic shift that gives the Sanders–Casar bill a hook into existing federal language even as the administration itself has not endorsed a development pause. The bill's supporters argue the alignment of terminology strengthens the case for treating the most advanced systems as a distinct regulatory category rather than a continuation of ordinary software policy.
The gap between what the bill proposes and what Congress typically passes is wide. A federal licensing regime for training runs, criminal liability for CEOs, and a new cabinet-adjacent agency would represent the most aggressive US tech regulation in decades, and there is no indication Republican leadership in either chamber intends to move the bill. Industry lobbying against binding federal AI rules has intensified through 2026, and prior state-level attempts to impose liability on frontier developers have been repeatedly narrowed or vetoed.
Even so, the bill matters as a marker. It sets a ceiling for what the safety-first wing of Congress is willing to propose, and it gives cover to narrower measures — disclosure requirements, incident-reporting mandates, compute thresholds tied to notification rather than approval — that stand a better chance of moving. Anthropic, whose co-founder has publicly warned about catastrophic AI risk, is among the labs whose research has been cited in the debate, and Microsoft's AI chief has recently said the company views advanced AI threats as real.
The practical question is enforcement. A ban on training runs above a data threshold requires the government to know what is being trained, which requires either mandatory pre-registration or continuous monitoring of large compute clusters — capabilities the proposed Department of AI would have to build from scratch. Without that infrastructure, the 20-year prison exposure is symbolic; with it, the compliance burden on every lab operating at frontier scale becomes substantial.
For the AI industry, the Sanders–Casar bill is unlikely to become law in its current form, but it moves the Overton window on what federal AI regulation could look like. Labs that have spent 2026 arguing against state-level rules on the grounds that federal action is coming now have to reckon with the possibility that when federal action arrives, it may be structured as a licensing regime with criminal penalties rather than the light-touch framework the industry has been lobbying for. The safest bet remains that Washington passes nothing binding this session — but the ceiling of the debate just moved sharply upward, and every serious frontier developer will be planning for a world in which some version of federal pre-approval eventually lands.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
