Proposed bill from Sanders and Casar would license advanced AI models
The Ban Artificial Superintelligence Act is a proposal, not law. It would pause every AI system trained above 10^25 operations and require federal approval before any of them ships.
The proposed line sits at 10^25 operations of training compute. Above it, no model ships without federal sign-off.
Sen. Bernie Sanders and Rep. Greg Casar introduced the Ban Artificial Superintelligence Act on September 23, 2026, Casar's office announced. It is a proposal, not law. As written, it would ban artificial superintelligence, pause advanced AI development, and require every advanced model to get approval from a new federal Department of Artificial Intelligence before release.
"Advanced" has a number. The section-by-section summary defines an advanced AI system as one trained on 10^25 or more integer or floating-point operations. The bill text has the Secretary of Artificial Intelligence adjust that threshold every year to track gains in training efficiency. We found no bill number as of September 25.
How approval would work
The pause comes first. Under Section 8, no advanced system could be trained, modified, or fine-tuned, except to remove dangerous capabilities or shut it down, and no unreleased advanced system could be deployed. The pause would end when the Secretary decides the Department is fully staffed and has written its rules. The text sets no deadline for either.
Those rules would have to include four things: developers reporting pre-development plans, monitoring and evaluation during development and after deployment, audits of safety practices, and final approval before any advanced system reaches the public. Section 9 then says no person may "deploy, release, import, or transfer" an advanced system without that approval. The Secretary would inspect the system and withhold approval if dangerous characteristics are present.
Section 12 adds the license. Anyone who develops or distributes an advanced model would need a Department charter and would have to grant "full access" to systems, staff, and physical infrastructure.
How the ban is defined
The ban itself uses a capability test, not a compute number. The bill text defines artificial superintelligence as a system that exhibits, "or can easily be modified to exhibit," performance that "exceeds human cognitive performance and capabilities across most domains or tasks," or enough capability to plan and execute the destruction or disempowerment of humanity. The text names no benchmark for "most domains or tasks." The Secretary would make that call during inspection, with an AI Advisory Board of experts advising on the rules.
Six "superintelligence precursor characteristics" would be banned too. They include the capacity to "automate or greatly accelerate" AI research, to resist shutdown, and to deceive humans to avoid oversight. A system found with one would be cut off from the internet and rendered inoperative within 30 days unless the Secretary can verify the precursor is gone.
Violations would carry up to 20 years in prison for people in policymaking roles and rogue actors, a 10-year industry ban for other employees, and, for companies, charter revocation and forfeiture of assets. The one-page summary calls that the "corporate death penalty."
What the text does not say
The bill lists developing, deploying, distributing, and importing as separate acts, but it defines none of them. It does not mention APIs, hosted access, or companies that build products on someone else's model. It does not name open-weight models, although Section 9 bars releasing any system, "including any elements sufficient to reconstruct the system's capabilities," that "may be foreseeably modified" into superintelligence or a precursor. Approval would be required to import an advanced system. Whether calling a model hosted outside the US counts as importing, the proposal does not say.
The bill faces long odds in the Republican-controlled Congress, the AP reports via PBS. The same report quotes employees of OpenAI and Google DeepMind who back it as individuals, not on behalf of their companies.
Why a build studio cares
We build AI workflows and agents on frontier model APIs, so the 10^25 line is the part we would watch. It would turn every new model version from our providers into an approval event, and during the pause nobody could fine-tune those models. The undefined word "deploy" decides who carries the burden. If it reaches the app that calls a model, every team shipping an AI feature would need sign-off. If it stops at the lab, most builders file nothing. The EU's Article 50 rules regulate what the output discloses, not whether the model ships. Whatever happens to this proposal, a written list of the models your product depends on, and the smaller ones you could swap to, is cheap to keep.
Next step: read the 19-page bill text and the section-by-section summary. To map which models your product depends on, write to us at hello@gattyworks.com.