On September 23, 2026, U.S. Senator Bernie Sanders (Independent-Vermont) and Representative Greg Casar (Democrat-Texas) formally introduced the Ban Artificial Superintelligence Act. The 19-page bill calls for a permanent ban on the development and deployment of artificial superintelligence and a full pause on advanced AI training until a new cabinet-level federal regulatory agency is established and safety rules are set. Individuals who violate the law would face up to 20 years in prison, while companies would face a "corporate death penalty" of forced dissolution. According to the Associated Press, current employees of OpenAI and Google DeepMind signed a joint statement publicly supporting this legislation, which takes direct aim at their employers' industry.
Penalty Severity: Benchmarked Against Nuclear Weapons
The 20-year prison figure was not chosen arbitrarily. In an interview with the Associated Press, Sanders explicitly said it matches the maximum penalty for "illegally manufacturing nuclear weapons." The analogy itself is a policy statement: in the view of Sanders and others, superintelligence belongs in the same category of threats as nuclear bombs—threats that must be sealed away by state coercive power—rather than being a risky technology requiring careful regulation.
From a comparative regulatory perspective, this characterization goes beyond the legislative practice of any current sovereign state. The EU AI Act fully entered enforcement in August 2026, with maximum fines for prohibited AI violations of €35 million or 7% of global annual revenue, but it contains no criminal provisions. The criminalization approach of the Sanders bill essentially elevates AI safety from administrative violation to criminal offense—a leap by an order of magnitude in legislative logic.
Who Counts as "Superintelligence": The Definition Is the Battleground
The bill defines artificial superintelligence as AI systems that surpass human performance in most cognitive domains, or possess capabilities sufficient to destroy or override humanity, including overthrowing the federal government. At the same time, the bill also brings within the scope of the pause systems capable of "automating or greatly accelerating the AI R&D process"—that is, models with recursive self-improvement capabilities.
This definition is deliberately vague at the technical level. Who exactly measures the standard of "surpassing humans in most cognitive domains," and by what benchmark? The bill does not provide clear operational procedures. Caleb Max, CEO of the National AI Association, pointed out directly on the day the bill was released that the bill was "stillborn," on the grounds that "some CEOs will say we have already crossed the threshold set by the bill." This remark is sarcastic, but it is also a real technical dilemma: if top models have already surpassed ordinary humans in multiple cognitive dimensions, has the bill's "pause line" already been crossed before it was even formally proposed?
The Most Noteworthy Signal: Betrayal by Insiders
The criminal provisions and numerical controversy in the bill occupied most media headlines. But the signal truly worth examining is the public endorsement by insiders. According to the Associated Press, OpenAI safety systems researcher Juan Felipe Cerón Uribe wrote in a statement: "Superintelligence could go extremely well or extremely badly, and we should not gamble on this." Swante Scholz, a software engineer at Google DeepMind, said that on the current development path, "the most likely outcome is an existential catastrophe for human civilization," and said that "banning the development of superintelligence would be a positive change"—both stated they were not speaking for their companies.
Such statements are not rare in the AI industry, but what does it mean for them to appear publicly in a specific legislative context? One interpretation is that these researchers can no longer push safety language into product decisions internally, so they are turning to legislative pressure to create external constraints. Another interpretation is more direct: they know their companies' R&D pace better than outsiders do, so they chose this moment to speak publicly. Either way, the rift between employees and employers over public policy shows that the AI safety debate has moved from academic argument into Silicon Valley's internal power structure.
The Geopolitical Paradox
The bill requires the United States to push for international agreements, coordinate with allies, and implement export controls, with the goal of preventing superintelligence from being developed anywhere in the world. This is the hardest part of Sanders's proposal to implement, and it also reveals its internal logical tension: if the United States pauses unilaterally while China continues to advance, the practical effect of this ban would be to hand over the lead.
Republican Representative James Comer's criticism hits exactly this point: he warned that the United States could fall behind China in the AI competition, and stressed that energy grid capacity and regulatory obstacles are already real constraints facing tech companies. The Trump administration's position is clearer still—according to Fox Business, Trump has consistently resisted additional AI regulation, emphasizing maintaining U.S. global leadership in the field.
This means the bill has very little chance of passing in the current Congress. Kevin Frazier, an AI policy expert at the University of Texas, said the bill "correctly recognizes that Congress needs to act quickly"—this is an affirmation of the legislative direction, not a prediction of its odds of passage.
What This Bill Is Really Doing
Why does a bill that has almost no chance of passing in the current political ecosystem deserve to be taken seriously? Because legislation is not only rule-making; it is also the construction of a discourse framework. The Sanders bill establishes the following as part of American political language: superintelligence and nuclear weapons belong to the same risk category; AI safety warrants a dedicated cabinet-level agency; and current AI researchers have reason to publicly express concern about their companies' R&D paths.
From this angle, the bill's real role is not to change the rules immediately, but to shift the boundary of "what can be discussed" toward stronger constraints before the next round of regulatory legislation with a greater chance of passing arrives. Years after the EU AI Act was regarded as "radical," it has become a reference point for global AI regulation. The extremeness of the Sanders bill at this moment is not necessarily permanent extremeness.
On September 16, a week before the bill was released, Sanders walked into the U.S. Congress together with AI godfather Geoffrey Hinton for a senators' AI briefing—this photo may explain the issue better than the bill text itself: when a Nobel laureate appears in a legislative body as an AI safety advocate, the debate no longer belongs only to engineers and policy experts.
Whether the bill can pass depends on the future political clock. But the questions it triggers—who has the right to pause a technology, who bears the cost of regulation, who is responsible for the consequences of losing control—the answers to these questions will determine the real developmental boundaries of the AI industry over the next decade.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接