20 Years in Prison for Building Superintelligence: Sanders' Bill Won't Pass, but It's Already Redrawing AI Politics

THE TEXT, NOT THE TWEET. On September 23, Senator Bernie Sanders and Representative Greg Casar publicly unveiled the Ban Artificial Superintelligence Act, a 19-page bill according to NBC News. It permanently prohibits developing or deploying 'artificial superintelligence' in the United States. The penalties are the harshest ever proposed for a civilian technology: up to 20 years in prison for individuals and a 'corporate death penalty' — forced dissolution of the company — for firms. Sanders equates the punishment to that for illegally building nuclear weapons. The bill also pauses advanced AI development until a new cabinet-level Department of Artificial Intelligence writes safety rules. And it declares it official US policy to pursue international agreements and export controls to stop superintelligence worldwide.

THE DEFINITION IS THE BATTLEFIELD. The text defines superintelligence through two prongs. One: a system that 'displays, or could easily be modified to display' capabilities matching or exceeding human cognitive performance across a broad range of domains. Two: a system with sufficient capabilities to plan and execute 'the disempowerment of humanity,' including overthrowing the US government. It also bans systems showing any of six 'precursor characteristics': automating AI research itself, resisting shutdown, self-modification, unauthorized access to infrastructure, enhancing biochemical weapons, and strategic deception. Here is the problem. According to a technical analysis by Pebblous, the word 'benchmark' (a standardized capability test) appears zero times in the text; 'standard' also appears zero times. The only concrete number is a threshold of 10^25 training operations — a measure of the compute used to train a model which, as distillation (copying a large model into a small one) has shown, correlates ever more poorly with real capability. Researcher Heidy Khlaaf sums it up in Science: the definition is 'hypothetical and unfalsifiable.' The toughest clause makes it worse: if the government cannot verify within 30 days that a system no longer exhibits precursor characteristics, the system must be destroyed. AI evaluations only demonstrate what a model can do; they never prove what it cannot do. The law demands proof of absence, and measurement science cannot deliver that today.

AI'S NEW POLITICAL MAP. The week of the bill was a portrait of realignment. Donald Trump told the UN General Assembly that the word 'artificial' 'makes intelligence sound fake,' and a State Department bureau promptly ordered, via internal email, that 'artificial intelligence' be replaced with 'superintelligence' in all its documents and positions, according to The Hill. That same September 23, the UN Security Council held its first session dedicated to loss of control over AI, with Sam Altman (OpenAI), Dario Amodei (Anthropic), scientist Yoshua Bengio and Clément Delangue (Hugging Face) as briefers. Meanwhile, alliances now cross the spectrum: Sanders shared a stage with Steve Bannon at the 'Pro-Human Assembly,' and the Future of Life Institute's Statement on Superintelligence has gathered more than 850 signatures, from Geoffrey Hinton and Steve Wozniak to Bannon and Glenn Beck. Public opinion is along for the ride: 68% of likely voters support the bill versus 25% opposed (Data for Progress, 1,300+ respondents), and 74% believe the government is not regulating AI enough (Quinnipiac). The money has arrived too: roughly $58 million from outside groups in the 2026 midterms — $33 million pro-safety versus $25 million pro-innovation, according to Transformer.

THE REAL ODDS. Let's separate the gesture from the law. As of this analysis, no bill number appears on congress.gov, there is no Republican cosponsor, and no hearing is scheduled. The CEO of the National AI Association declared it 'dead on arrival' in comments to NBC News, and Senator Rand Paul blocked a related AI bill days earlier. Criticism comes from both sides: the R Street think tank calls it 'AI authoritarianism,' and Gary Marcus — who favors creating a federal AI agency — considers a permanent, unilateral ban a mistake that would cede ground to other countries, calling it 'naive' about benchmarking. Even Democrats are split: Transformer describes four internal factions that cannot agree on what the main risk even is.

OUR READING. Three takeaways. First: this is not a law, it is dormant legislation. Texts like this sit in a drawer until a serious incident opens the political window — and then they pass in weeks, flaws intact. That is why its definition must be criticized now, not when it is too late. Second: the real battle is over vocabulary. Trump adopts 'superintelligence' as a brand of national greatness; Sanders, as an existential threat to eradicate. The same word sustains two opposite projects, and whoever fixes its legal meaning — with or without an activation metric — wins the debate. We have written it before: without a measurable threshold, the rule is symbolic and power stays with whoever controls the dictionary. Third: the brake as a moat. A pause freezes the current ranking. Today it would benefit established leaders over those coming up behind, and a unilateral US ban does not stop Chinese open-weight models (models whose weights are published for anyone to run), which already compete with the closed frontier.

REGULATE CAPABILITY, NOT PANIC. The short-term risks are real — jobs, agent-automated fraud, concentration of power — and they justify serious governance. But regulating with unfalsifiable definitions produces only two possible outcomes: dead letter or arbitrariness. The most useful thing about Sanders' bill is that it forces us to build what is missing: a public metrology of AI capabilities, with independent evaluations, measurable thresholds and verification protocols — exactly what Bengio and Altman asked for before the Security Council. What to watch in the coming months: whether the bill gets a number and a hearing, whether a Republican cosponsor appears, whether states copy individual pieces, and whether evaluation science matures. Our horizon does not change: well-governed AI points toward eradicating disease, extending life and generating abundance. The bridge there is neither panic nor euphoria. It is learning to measure before we ban.