Sanders–Casar AI bill would pause training at a 10²⁵-operation threshold

Sen. Bernie Sanders and Rep. Greg Casar introduced the Ban Artificial Superintelligence Act on September 23, 2026. It proposes a temporary pause on developing advanced AI systems, a lasting prohibition on systems the bill defines as superintelligent or showing specified precursor abilities, and a new federal Department of Artificial Intelligence. It is a bill, not a rule in force today. The sponsors' announcement establishes the introduction date; the published bill text sets out the proposed mechanism. The Verge's source report was published the same day.

The immediate reader question is therefore about scope, not compliance. A developer using an existing model would not suddenly have to stop because the proposal was introduced. If enacted as written, it would regulate the development and deployment of certain systems through a training-compute threshold, capability prohibitions and approvals from a department that does not yet exist.

The pause has a compute threshold and an end condition

The text defines an advanced AI system as one trained using at least 10²⁵ integer or floating-point operations. It directs the proposed department's secretary to adjust that threshold each year for changes in training efficiency. This is a measure of training work, not a parameter count, consumer-device specification or blanket definition of every AI application. The bill does not establish a public list of models that would meet it.

From enactment, covered advanced systems would face a mandatory pause on training, modification and fine-tuning, subject to narrow safety-related exceptions; unreleased covered systems could not be deployed during that period. The pause would end only after the proposed department was fully staffed and its secretary had issued safety rules. Those rules would require plans before development, monitoring and audits during the model lifecycle, and final approval before public release. Roll Call's September 23 account independently describes the proposal as a pause pending a new agency and rules, rather than a ban on all AI use.

The permanent prohibition is broader than one model label

The bill separately defines artificial superintelligence by capabilities: exceeding human cognitive performance across most domains, or being capable of planning and executing humanity's destruction or disempowerment. It would also prohibit systems with listed precursor characteristics, including unauthorized access to secured infrastructure, resistance to shutdown, certain weapons-related assistance, or independent self-modification. These are proposed legal tests, not findings that any named public model has them.

The planned department would inspect advanced systems before deployment and could withhold approval when it finds dangerous characteristics. The text also calls for a charter to develop or distribute an advanced model. Its headline prison provision applies to covered policy-making individuals or unaffiliated actors who recklessly violate specified sections or rules, with a maximum of 20 years. That is a conditional proposed penalty; it does not mean ordinary users or every AI engineer face prison for using current tools.

What developers can conclude today

There is no enacted pause, no new department and no approval process under this bill. Its congressional path and final wording are unresolved. Even if it advanced, applying capability terms such as “across most domains” and measuring training work across different development methods would require choices by lawmakers and regulators. The bill proposes annual threshold adjustment but does not answer every case in advance.

For teams planning frontier-model work, the useful next step is to read the bill's definitions and operative sections, track its actual legislative status, and separate vendor controls from statutory requirements. Our earlier report on Claude Opus 5.5's cybersecurity routing describes a provider's product safeguard; this proposal would create a different, federal approval regime if it became law. Neither the sponsors' risk claims nor opponents' objections by themselves establish what Congress will enact.