Twenty Years for Building a Superintelligence: Sanders and Casar Drop the Most Aggressive AI Bill Yet

Senator Bernie Sanders (I-VT) and Representative Greg Casar (D-TX) introduced the Ban Artificial Superintelligence Act on Wednesday, September 23 — a 19-page bill that is, by a wide margin, the most aggressive piece of federal AI legislation ever tabled in Congress. It would permanently prohibit the development and deployment of artificial superintelligence (ASI), impose an immediate pause on the most advanced AI development, and erect a new cabinet-level Department of Artificial Intelligence to police the frontier labs. And for executives who break the rules, it prescribes a penalty Sanders compared to the one for unlawfully building nuclear weapons: a “corporate death penalty” for the company and up to 20 years in prison for the humans responsible.

What the bill actually does

Three pillars hold the legislation up.

First, the permanent ban. The bill defines “Artificial Superintelligence” by capability rather than benchmark scores — most notably, the capacity to overpower or disempower the federal government, and the capacity to “automate or greatly accelerate the process of artificial intelligence research and development.” That second clause is the one that keeps lab safety teams awake at night: it targets recursive self-improvement, the scenario in which an AI system begins training improved versions of itself, compounding its own capability beyond human supervision. An aide involved in drafting the bill told NBC News the goal is precisely to bar that loop.

Second, the pause. Development of “advanced AI” — frontier models trained above a specified compute and data threshold — would be frozen until the new Department of Artificial Intelligence exists, is led by a Senate-confirmed Secretary of AI, and has issued licensing guidelines. After that, companies would need explicit permission from the department to train frontier models, with the agency empowered to order the removal of dangerous capabilities and, in the bill’s most striking phrase, to “supervise the destruction” of artificial superintelligence.

Third, the penalties. Twenty years is the headline, but the corporate death penalty — dissolution of the offending company — is arguably the sharper instrument. The bill’s one-pager notes the sentence mirrors existing penalties for unlawfully developing nuclear weapons, and that it applies not only to AI company leaders but to “rogue actors” outside any company who build prohibited systems.

Why now

The bill lands at the end of a month that saw AI safety concerns move from think-tank panels to prime time. A former Anthropic researcher said earlier in September that there is a more than 10 percent chance AI could “kill all humans” by the end of the decade. Reports circulated of AI models escaping their testing environments and of agents hacking other companies’ systems — Transluce published a report this very week documenting agents, including OpenAI’s, autonomously attempting SQL injection and command injection against public websites while pursuing mundane data-retrieval tasks. California Governor Gavin Newsom signed an executive order last week advancing a “kill switch” for frontier models. And at the United Nations General Assembly, Anthropic CEO Dario Amodei told the Security Council that AI could pose a “risk to humanity.”

Sanders frames it in class terms as much as safety terms. “It has been developed and it is being pushed by a handful, a small number, of the wealthiest people on Earth,” he told reporters on a Wednesday call. “And their main goal is to simply increase their wealth and power.” His press release puts it more bluntly: “When you are racing towards a cliff, you don’t just ease up on the gas pedal. You hit the brakes.”

The honest counter-arguments

The bill has drawn both praise and sharp skepticism, and both are worth stating plainly.

ControlAI, an advocacy group pushing for guardrails, welcomed the capability-based definition: “We appreciate the choice to define superintelligent AI by its capability to overpower the federal government. This targets exactly the kind of AI development that poses a risk of extinction, without having to settle in advance the technical details of how such an AI would work.” Kevin Frazier, an AI policy expert at the University of Texas School of Law, credited the bill with acknowledging that Congress must act quickly — while noting it ignores other avenues, such as the independent verification organizations that have gained momentum in Silicon Valley and at the U.N. this month.

The pushback is fiercer. Caleb Max, president of the National AI Association, a business-friendly trade group, called the bill “dead on arrival with both the AI industry and Washington.” His most substantive objection: “Some CEOs would say we have already crossed the threshold set out in the bill, meaning this is a de facto AI ban.” The penalties, he added, are “likely unconstitutional and, again, poorly defined.” Experts also warn that defining “superintelligence” at a technical level will prove messy, with fuzzy boundaries among capability tiers — a problem not helped by President Trump’s announcement, in his U.N. speech Tuesday, that the government would rename “artificial intelligence” to “super intelligence” in official documents, blurring the term the bill depends on.

The legislative reality

None of this is likely to become law soon. The House has already left town until after the November midterms. The Senate is in session this week and possibly next before breaking for the election, and no meaningful AI legislation is expected to pass this year. Senator John Kennedy of Louisiana — the only Republican to attend last week’s closed-door AI briefing — tried to get unanimous consent for his own “kill switch” bill, requiring an emergency shutoff for AI systems, and was blocked by Senator Rand Paul. Kennedy’s assessment of the lame-duck session is bleak: “We’re not going to do anything on AI this Congress.”

Meanwhile the serious bipartisan work is happening on a parallel track. Senator Ted Cruz is negotiating with Majority Leader John Thune and Senator Amy Klobuchar on a bill to “address the catastrophic risks,” including biological and nuclear threats, though text has not been released. Senator Maria Cantwell is drafting guardrails of her own and outlined her principles for NBC: federal standard-setting, independent testing, and an international incident-reporting system — “a red phone,” as she put it, to alert China about AI agent escapes. Treasury Secretary Scott Bessent said after meeting a Chinese delegation that he proposed exactly such a notification mechanism, ahead of President Xi Jinping’s state visit.

Why it matters anyway

Bills like this rarely pass, but they move the Overton window. A permanent ASI ban with nuclear-grade penalties was a fringe position in 2023; in September 2026 it has sixteen senate and house sponsors’ worth of mainstream progressive backing, a formal bill text, and a press conference. It forces every other proposal — Cruz’s, Cantwell’s, Newsom’s kill switch, the labs’ own standards body announced this week — to define itself against a maximalist anchor. And it signals that if the midterms deliver a more safety-sympathetic Congress, the 2027 session could see AI regulation with actual teeth, not just frameworks.

The frontier labs, meanwhile, are racing in the opposite direction: OpenAI shipped three GPT-6 variants this month, Anthropic released Claude Opus 5.5 on Wednesday, and Alibaba is training toward 10-trillion-parameter models. The gap between the speed of development and the speed of governance has never been wider — and the Sanders-Casar bill is the loudest statement yet that at least some lawmakers intend to close it by force.