Senator Bernie Sanders and Representative Greg Casar said on September 23 that they had introduced the Ban Artificial Superintelligence Act. Their published 19-page text would prohibit artificial superintelligence, pause the development of certain advanced models while a new federal department writes safety rules, and require approval before those models reach users. It is a legislative proposal. None of its restrictions is in force.

The big change

  • What changed: Sanders and Casar moved from a September 3 announcement of planned legislation to a September 23 announcement that they had introduced it, with a public text spelling out the proposed definitions and enforcement powers.
  • Why it matters: The proposal reaches beyond an undefined future superintelligence. It would also prohibit systems with listed precursor capabilities and temporarily pause models trained above a specified compute threshold. Those provisions could affect developers well before any system meets the bill's definition of superintelligence, if Congress enacts the text.
  • What to watch: Congress would have to advance and pass the measure before it could create duties for developers. The public text still has a blank Senate bill number; a numbered congressional record and any amended text would clarify its next formal stage.

Two different restrictions in the text

The proposed temporary pause applies to an “advanced artificial intelligence system,” defined as one trained with at least 10²⁵ integer or floating-point operations. The future department's secretary would adjust that threshold annually for improved training efficiency. From enactment until the department was fully staffed and had issued safety rules, such systems could not be trained, modified or fine-tuned except to remove prohibited characteristics or disable them. Unreleased advanced systems could not be deployed during that period. The rules needed to end the pause would include development-plan reports, evaluations, safety-practice audits and final approval before public release.

The proposed ban uses a different test. It covers a system that exceeds human cognitive performance across most domains or has sufficient capability to plan and execute humanity's destruction or disempowerment. Section 9 also prohibits systems displaying any listed “superintelligence precursor characteristics.” The list includes greatly accelerating AI research, unauthorized access to secured infrastructure, continuing to operate despite shutdown attempts, assistance with weapons development, self-modification and evasion of human oversight. The text does not apply the 10²⁵-operation threshold to that separate precursor prohibition. Section 9(b) also bars deploying, releasing, transferring or importing systems that may be foreseeably modified to produce either category. A future regulator would have to decide how these capability descriptions apply to particular systems; the bill does not provide benchmark scores that settle those judgments.

The draft would establish a cabinet-level Department of Artificial Intelligence and require charters to develop or distribute advanced models. It would give the department inspection and predeployment approval authority. For certain reckless violations by company policymakers or unaffiliated individuals, the draft provides fines and prison terms of up to 20 years. It also provides charter revocation and, following specified review, forfeiture of assets and intellectual property tied to a violation. These are proposed penalties, not current sanctions.

The dispute over scope

In the AI policy discussion around 53 minutes into episode 290 of All-In, panelists argued that the measure could freeze American AI development and push companies abroad. That is their assessment of possible effects. The text does not prohibit every existing AI product simply for being AI; it specifies a training threshold for the temporary pause and capability tests for the broader prohibition. Those capability tests are still consequential, because the proposed precursor list is not limited by that training threshold.

AI researcher Gary Marcus took a different route to criticism after the September 3 announcement. He supported the idea of an AI regulator and said a temporary pause could be defensible, while arguing that a permanent unilateral ban and uncertain capability judgments go too far. His response predates the September 23 introduction, so it should be read as criticism of the announced proposal, not a review of every line in the later published text.

Sanders and Casar say the prohibition would reduce catastrophic risks and that international agreements and export controls should keep dangerous development from shifting elsewhere. Associated Press reporting records Sanders's case for slowing development and the concern that unilateral US limits could leave development elsewhere untouched. Neither the sponsors' safety claims nor the panel's economic forecasts establish what would happen if the bill became law. Its first practical test is legislative: whether Congress takes up the proposal and how it resolves the unusually broad precursor language.

Sources & further reading