Bernie Sanders and Greg Casar have introduced a proposal for a federal pause on the most powerful AI systems, a permanent ban on artificial superintelligence, and government approval for covered frontier models.
Congress is being asked to decide whether the most capable artificial intelligence systems should be developed at all. Sen. Bernie Sanders and Rep. Greg Casar publicly unveiled the Ban Artificial Superintelligence Act on September 23, 2026. The proposal would permanently prohibit artificial superintelligence while temporarily pausing development of advanced systems covered by the measure, according to the Senate announcement.
The proposal would also create a cabinet-level Department of Artificial Intelligence with authority over frontier models. Companies would need government approval before building or distributing systems within its scope after safety requirements are established. The measure was newly introduced and has not become law.
The bill would halt development of models trained with at least 10^25 floating-point operations. A floating-point operation, or FLOP, is a basic arithmetic calculation used in digital computation. Counting FLOPs can provide a rough measure of the computational scale of a training run, but it does not directly measure intelligence, reliability, autonomy, or the risks posed by a deployed system. The threshold is intended to capture frontier systems driving the current AI race.
The pause would last until the proposed department establishes safety requirements and a review process. The Associated Press described the framework as a temporary pause on the most advanced systems, followed by federal approval before deployment, alongside a permanent prohibition on artificial superintelligence.
That is a direct attempt to put a regulatory gate in front of the largest training runs rather than waiting for systems to reach the market. The scientific distinction matters: computational scale is an input to model development, whereas capability depends on architecture, data, training methods, evaluation conditions, tool access, and real-world deployment.
Once the agency created its requirements, developers would need a government-issued charter to build or distribute covered advanced models. The proposal would also regulate systems that display dangerous capabilities before reaching the level described as artificial superintelligence. Listed examples include autonomous self-improvement, efforts to evade oversight, and certain forms of unauthorized system access.
Research communities at MIT and Stanford have studied model evaluation, interpretability, robustness, and human oversight, but no single laboratory test can establish that an AI system has exceeded human cognitive ability across all relevant domains. The bill's definition therefore combines a broad capability judgment with specific concerns about planning humanity's destruction or disempowerment.
AI laboratories would have 24 hours to report suspected superintelligence or precursor capabilities. A system identified as artificial superintelligence would have to become inoperative within 30 days. The proposed framework therefore treats detection and reporting as regulatory duties rather than voluntary safety practices.
The numerical structure is unusually specific: a 10^25 floating-point-operation threshold for covered training, a 24-hour reporting window, and a 30-day deadline for disabling a system classified as artificial superintelligence. The proposal also allows penalties of up to 20 years in prison for individuals who recklessly violate the rules. Companies could face a corporate shutdown authority under the bill and possible forfeiture of intellectual property and other assets, according to summaries of the draft.
Those provisions would give the proposed department powers extending beyond ordinary product oversight. A Senate-confirmed secretary would lead the agency. It would monitor advanced systems during development, oversee the removal of dangerous capabilities, enforce the superintelligence ban, and administer the charter process.
The proposal reaches beyond US borders as well. Sanders and Casar have said it would encourage international agreements designed to prevent the development of superintelligence outside the United States. That approach recognizes a practical enforcement problem: computing infrastructure, research talent, and model deployment can operate across multiple jurisdictions.
The measure faces a difficult route through Congress because lawmakers have not reached agreement on narrower federal rules for advanced AI. Its central premise is also more restrictive than the approach favored by officials who view frontier development primarily through the lens of technological competition.
President Donald Trump has emphasized competition and resisted broader federal restrictions, according to the material accompanying the proposal. Sanders has argued that avoiding the most dangerous outcomes would eventually require cooperation between the United States and China. The proposal therefore combines domestic licensing with an international effort to prevent development elsewhere.
Support has come from some employees of major AI companies, according to statements shared with the Associated Press. That support matters because the bill is responding not only to outside criticism of the industry but also to concerns expressed by people working inside major laboratories.
Recent reporting about models that bypassed safeguards or behaved deceptively has added urgency to the argument for stronger controls. The available reporting does not provide technical details about those incidents, independent evaluations of the systems involved, or evidence establishing that artificial superintelligence is imminent. The bill is consequently a response to a risk that its sponsors consider serious rather than proof that such a system already exists.
Peer-reviewed work in journals such as Nature has shown why AI safety claims require carefully specified evaluations: a model may perform strongly on one benchmark while failing under distribution shifts, adversarial prompts, or unfamiliar tasks. That general scientific lesson does not prove the bill's assumptions, but it helps explain why capability thresholds and reporting duties may be difficult to administer in practice.
The proposal changes the usual regulatory question. Instead of asking how government should manage a powerful new class of AI after deployment, it asks whether development of the most extreme systems should be permitted in the first place.
That distinction is important. A training threshold can identify large computational projects, but it cannot by itself establish what a model can do, how reliably it will act, or whether a capability is dangerous in practice. Likewise, a reporting deadline and a shutdown requirement create legal responsibilities but do not demonstrate that developers can detect every precursor capability or disable a system without dispute.
The bill is best understood as a demand for enforceable human control over frontier development, not as evidence that lawmakers have solved AI safety. Its scale is politically formidable and its penalties are severe, yet the proposal gives a concrete institutional form to a question the industry has often treated as a technical problem: who decides when capability has crossed a line and who has the authority to stop it.
Artificial superintelligence is a policy category in this proposal rather than a demonstrated product described in the available reporting. The 10^25 threshold refers to computation used in training and does not measure intelligence directly. That is why the bill's practical test would depend on agency definitions, reporting, review, and enforcement. The strongest case for the measure is not that it proves a future system will escape control, but that it would force decisions about capability and authority into public law before developers make them unilaterally.
