US Democratic Lawmakers Propose Banning Superintelligence: Advanced AI Training Would Require a Federal License, Violators Could Face Up to 20 Years in Prison

In September 2026, US Senator Bernie Sanders (Independent, Vermont) and Representative Greg Casar (Democrat, Texas) formally introduced the Ban Artificial Superintelligence Act in Congress, with Representative Alexandria Ocasio-Cortez (AOC) also signing on to help advance it. The bill calls for a federal pause on advanced AI development, establishes a licensing regime led by a newly created cabinet-level Department of Artificial Intelligence, and imposes criminal penalties of up to 20 years in prison on violators. The bill gained significant political momentum after Democrats secured a congressional advantage in November, and some Democrats have listed it as a priority for “day one” of governing.

The bill’s core mechanism moves the point of government regulatory intervention from the “outputs” of AI systems to “the training process itself.” In the past, US AI-related regulations generally intervened only after AI products were deployed, checking whether they discriminated, produced harmful content, or were used for illegal purposes. The new bill requires AI labs to apply for and obtain a license from the Department of Artificial Intelligence before starting training of “advanced AI.” The Secretary would independently hold the power to define “advanced AI” and could adjust the threshold at any time as technology evolves; the bill calls this a “rolling threshold,” and Congress itself would not intervene in setting specific standards.

Even more disruptive are the infrastructure access provisions. The bill requires licensed labs to give the Secretary of Artificial Intelligence “complete access” to all their digital and physical infrastructure, explicitly including data centers. This means regulators would not only review results but also enter training clusters, inspect model weights, and review training data. For “rogue actors,” the bill designs three cumulative penalties: violating companies face judicial dissolution; responsible executives personally face up to 20 years in prison; and rank-and-file employees could be barred from the industry for 10 years.

The bill’s impact on the AI industry ecosystem would be highly differentiated by stakeholder. For leading labs such as OpenAI, Google DeepMind, and Anthropic, compliance costs are a secondary issue; the truly thorny part is “complete infrastructure access”—equivalent to exposing model architectures, training data, and even trade secrets to federal regulators, with no exemption for commercial sensitivity. For small and medium-sized AI startups, the compliance threshold of the licensing regime could constitute a substantive market-entry barrier, objectively creating protection for the largest incumbents. For the open-source community, the bill text does not currently clarify whether distributed training, individual researchers, or academic institutions would be subject to the same licensing obligations.

In terms of historical precedent, Casar and Sanders benchmark the 20-year criminal penalty against “illegal nuclear weapons development.” The US Atomic Energy Act of 1946 did establish an end-to-end federal licensing system for nuclear materials from production to disposal, with criminal liability attached. But the physical feasibility of nuclear control rests on scarce and traceable materials such as uranium ore and centrifuges; the compute, data, and code needed for AI training are highly digital, replicable, and able to move across borders, making enforcement boundaries far blurrier than for nuclear materials.

In terms of the international competitive landscape, if the United States makes frontier AI training subject to criminal control while China, the United Kingdom, the European Union, and other jurisdictions have no equivalent restrictions, AI training workloads would in theory have an incentive to migrate to jurisdictions with lower regulatory costs. In a statement accompanying the bill, Sanders explicitly called for promoting an “international agreement” to stop other countries from developing superintelligence, but international negotiations typically take years; if unilateral legislation moves first, it could cause competitive losses for US AI labs before any agreement is reached.

In terms of legislative prospects, the bill has a very low probability of passing under the current political configuration. The bill’s sponsors are all progressive Democrats, but even if Democrats controlled both chambers, many moderate lawmakers have clear reservations about defining AI development as a criminal offense. The bill’s more practical function lies in the “Overton window effect”: by proposing an extreme version of regulation, it makes AI accountability legislation previously seen as harsh appear moderate by comparison, thereby securing more political space for other bills.

Two key signals will test this judgment: first, how the lobbying power of large technology companies responds, and whether cracks appear in their relationship with Democratic donations; second, whether the committee review stage amends the most controversial provision, “complete infrastructure access.”