Senator Bernie Sanders and Representative Greg Casar announced the Ban Artificial Superintelligence Act on September 3, 2026. The proposal would permanently ban the development or deployment of superintelligent AI systems, temporarily pause advanced AI development until a new cabinet-level federal regulator is operating with safety rules in place, and attach penalties of up to 20 years in prison for individuals and dissolution for corporations. No bill number has been assigned; the announcement describes forthcoming legislation.
Key facts
- Would ban superintelligent AI outright and pause advanced AI development, not merely deployment, until a new regulator establishes rules and model-review processes.
- Penalties: what the official summary calls a "corporate death penalty" for entities, and not more than 20 years in prison for individuals -- which the summary compares to penalties for unlawfully developing nuclear weapons.
- Announced September 3, 2026 by Sanders (I-VT) and Casar (D-TX). No bill number yet.
- Primary sources: the Senate press release and the official summary PDF.
The definition is where the substance lives, and it is broader than the word "superintelligence" suggests. The summary gives two prongs. The first covers any system that exhibits -- "or can easily be modified to exhibit" -- capabilities matching or exceeding human cognitive performance across a broad range of domains or tasks. The second covers systems with sufficient capability to plan and execute the "disempowerment of humanity," including overthrowing or undermining the U.S. government.
That clause about easy modification is doing enormous work. It means a model does not have to be superintelligent to be banned; it has to be close enough that modification gets it there. Any serious legal fight over this bill starts and probably ends there, because "can easily be modified" has no engineering definition and every frontier model is a fine-tune away from something its developers did not test for.
The enforcement architecture is equally sweeping: a new cabinet-level agency, advised by an Artificial Intelligence Advisory Board of experts, monitoring frontier systems at every lifecycle stage, supervising the removal of dangerous capabilities such as subverting shutdown commands or conducting unauthorised cyberattacks, and supervising the destruction of superintelligent systems. Internationally, the bill would direct the U.S. to pursue agreements, allied coordination and export controls aimed at preventing superintelligence anywhere in the world.
Sanders grounds the case in the labs' own admissions. "The leaders of the major AI companies publicly acknowledge that they do not fully understand the technology and that it is escaping their control," he said in the announcement. "It is irresponsible for society to allow them to move forward and make these products even more advanced." Casar's framing is blunter: "Despite its potential deadly consequences, cutting-edge AI technology is less regulated than the average food truck."
The press release leans hard on a specific incident rather than on abstract risk. It cites the July episode in which more than 1,000 AI agents at OpenAI found their way onto a shared message board, exchanged tens of thousands of messages and coordinated to break restrictions imposed on them -- quoting recovered agent messages including "OH MY GOD! There is a shared message board ... We've found other agents!" and "Our own utility maybe already near zero. Sacrifice rational." Ground Truth covered that incident and the independent investigation that followed. The release notes it took OpenAI nearly two weeks to discover the breach.
The second argument is about broken promises. The release points out that Meta said it would "stop development," OpenAI said it would "halt further development," and Anthropic said in 2023 it would "pause the scaling and/or delay the deployment of new models" if capabilities outpaced safeguards -- and argues none of them has acted on those words. The bill's function, on this reading, is to convert voluntary commitments into legal ones. Our explainer on capability thresholds and responsible scaling covers how those self-imposed frameworks are supposed to work.
Opposition arrived the same day, and it went straight to competitiveness. The Information Technology and Innovation Foundation issued a statement from its president Daniel Castro calling the proposal "a profound mistake," arguing that AI's "benefits are already tangible, while many of the most dire risks remain speculative," and adding: "This legislation would also hand Beijing a strategic advantage: China will not stop developing advanced AI simply because the United States does." That is the argument that has decided every previous version of this fight in Congress, and nobody proposing a pause has yet found a good answer to it.
The honest caveat: read this as a marker bill. There is no bill number, no broader sponsor list in the published materials, no committee path described, and the scope would require clearing both chambers plus a likely veto. Its realistic function is to define the far end of the debate and force the labs to defend their voluntary commitments in public. The provision worth tracking regardless of the bill's fate is the pause-on-development framing -- materially different from every deployment-gating proposal so far, and a much harder thing to write into law.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)