The Ban Artificial Superintelligence Act would create a Cabinet-level AI department empowered to approve advanced-model development and deployment, order certain systems offline or inoperative, revoke federal charters and impose penalties reaching 20 years in prison.

By yourNEWS Media Newsroom

Sen. Bernie Sanders, I-Vt., and Rep. Greg Casar, D-Texas, have introduced legislation that would place the development of the most computationally advanced artificial intelligence systems under a new federal regulatory structure, impose an immediate pause on qualifying models and permanently prohibit artificial superintelligence as defined by the proposal.

The Ban Artificial Superintelligence Act, introduced Sept. 23, would establish a Cabinet-level Department of Artificial Intelligence headed by a presidentially appointed secretary subject to Senate confirmation. The department would oversee advanced AI development, require federal authorization before qualifying systems could be released and possess authority to order certain AI systems disconnected from the internet or rendered inoperative. Senator Bernie Sanders

The complete 19-page bill text lays out a regulatory system built around computing thresholds, mandatory government review, developer charters, continuing federal monitoring and criminal and employment penalties for violations. The measure would apply its advanced-AI rules to systems trained using at least 10^25 integer or floating-point operations — 10 septillion operations — with the secretary required to adjust that threshold annually as training efficiency and technology change.

The bill distinguishes those “advanced artificial intelligence systems” from “artificial superintelligence,” which it defines through capabilities rather than computing power alone.

Under the proposal, artificial superintelligence would include a system that “exceeds human cognitive performance and capabilities across most domains or tasks,” including decision-making, learning and adaptive behavior. It would also include a system capable of planning and carrying out the destruction or disempowerment of humanity, including by overthrowing or undermining the federal government.

The legislation separately establishes a category of “superintelligence precursor characteristics.” Those include the ability to automate or greatly accelerate AI research and development; gain unauthorized access to secured digital or physical infrastructure; continue operating despite attempts to shut the system down; assist the design, production, modification or acquisition of nuclear, chemical or biological weapons; independently modify or improve its own functions; or deceive or otherwise evade effective human oversight.

That definition is significant because the bill’s restrictions would reach not only systems classified as artificial superintelligence but also AI exhibiting any of those precursor capabilities.

Upon enactment, every system meeting the bill’s advanced-AI computing threshold would enter a mandatory pause, and unreleased advanced systems could not be deployed. That pause would remain until the secretary determines the new department is fully staffed and has issued regulations governing safe and secure AI development and deployment.

Those regulations would have to require developers to submit plans before development begins. They also would establish monitoring and evaluation throughout development and deployment, including after deployment; audits of organizational safety practices; and final federal approval before an advanced AI system could be released publicly.

The legislation would go further than requiring disclosure or safety testing. No person could deploy, release, import or transfer an advanced AI system without pre-deployment approval from the Department of Artificial Intelligence.

Before issuing that approval, the secretary would inspect the model for characteristics determined to pose a danger to the public, including superintelligence or its precursor characteristics, and would be required to deny approval if those characteristics were present.

Development and distribution also would be subject to a federal charter.

The secretary could issue a charter permitting a person or company to develop or distribute an advanced AI model, but the bill states that no person could undertake either activity without such a charter. Applicants would have to provide whatever information the secretary requires.

As a condition of obtaining and retaining that authorization, developers would be required to give the secretary full access to advanced AI systems, internal processes, employees and physical infrastructure to the extent the secretary considers appropriate for enforcement.

The charter structure would therefore place covered advanced-model development under continuing federal supervision rather than relying solely on a one-time licensing decision. It also could have substantial implications for open-source and open-weight development when a model falls within the legislation’s definitions, because development or distribution of a covered advanced model would require a federal charter.

The bill grants the proposed department additional authority when the secretary identifies dangerous capabilities.

If an AI system is determined to possess a superintelligence precursor characteristic, the secretary would be required to place it under an immediate mandatory pause and have it sequestered from the internet. If the government cannot verify within 30 days that the system no longer exhibits the identified characteristic, the system would have to be rendered inoperative.

A system formally identified as artificial superintelligence would face an even stricter requirement: It would have to be rendered inoperative immediately.

The legislation also prohibits any person from developing, deploying, acquiring, possessing, funding, importing or transferring artificial superintelligence or AI displaying one or more precursor characteristics. The restriction extends to components sufficient to reconstruct the prohibited system’s capabilities.

It additionally prohibits releasing, transferring or importing an AI system that could foreseeably be modified to produce artificial superintelligence or the listed precursor characteristics.

Anyone who discovers artificial superintelligence — or an AI system possessing one of those precursor characteristics — would have 24 hours to report the discovery to the secretary. The reporting requirement does not apply merely because a system crosses the 10^25 training threshold; it is triggered by discovery of superintelligence or the specified precursor capabilities.

Violations could carry consequences for companies, executives, independent developers and employees.

A policymaking employee of a chartered entity, or a “rogue actor” unaffiliated with one, who recklessly violates specified provisions or regulations could be fined under federal law and imprisoned for up to 20 years.

Non-policymaking workers found to have recklessly violated the same provisions would face a 10-year prohibition on employment in the artificial intelligence industry. They could ask the secretary to review that finding, and the secretary could overturn it.

Sanders and Casar’s offices have described the company-level sanction as a “corporate death penalty” and compared the potential 20-year prison term with existing penalties related to unlawfully developing nuclear weapons. Senator Bernie Sanders

A company found to have violated the covered provisions would lose its charter. The company could seek review from the U.S. Court of Appeals for the Federal Circuit, which could uphold or overturn the finding.

While an appeal is pending, the court could place the company’s intellectual property into receivership at the company’s request until judicial review is completed and the secretary has ensured that prohibited systems have been rendered inoperative.

If the Federal Circuit upheld the violation, intellectual property and assets would become subject to federal forfeiture for the stated purpose of ensuring the complete destruction of systems and hardware associated with the violation. Charter applicants would agree to that potential forfeiture as a condition of receiving authorization to develop or distribute covered models.

The proposal has drawn opposition from some AI industry advocates.

Caleb Max, president and CEO of the National AI Association, said some current technology may already satisfy language used by the legislation.

“Some CEOs would say we have already crossed the threshold set out in the bill, meaning this is a de facto AI ban,” Max said. He separately argued that the proposed penalties were likely unconstitutional and poorly defined.

Sanders and Casar have framed the proposal differently, saying federal intervention is necessary because increasingly powerful AI systems could develop capabilities that humans cannot reliably control. Their legislation’s findings cite warnings from researchers and executives at major technology companies about loss of control, biological risks, autonomous AI research and other potential consequences of increasingly capable systems.

The proposed department would also have an Artificial Intelligence Advisory Board whose members would be selected by the secretary from people the secretary considers experts in AI or other fields relevant to the board’s work. The panel would provide scientific and technical advice concerning the department’s regulations.

Department employees would face conflict-of-interest restrictions. They could not participate in matters in which they held a financial interest or maintain a controlling financial interest in an entity regulated by the department. Former department officials and employees also would be prohibited from representing people before the agency.

The measure contains whistleblower protections covering current and former employees as well as current and former independent contractors.

Employers could not fire, demote, suspend, threaten, blacklist, harass or otherwise discriminate against covered individuals for lawfully reporting suspected violations to the AI department, law enforcement or Congress; participating in investigations or proceedings; or providing information internally to someone reasonably believed to have authority to address misconduct.

The secretary also could waive the 10-year industry employment ban for a non-policymaking worker who provided information about a violation when the secretary determines a waiver is consistent with the department’s public-safety mission.

The proposal extends beyond domestic regulation.

The secretary of artificial intelligence, working with the secretary of state, would be directed to pursue international agreements and coordination with U.S. allies aimed at preventing artificial superintelligence from being developed anywhere in the world.

The bill would establish prevention of global superintelligence development as official U.S. policy and specifically permits the use of export controls targeting AI computing infrastructure as one tool toward that objective.

Federal money also could not be used for activities violating the legislation’s pause, prohibition or charter requirements. The bill provides a limited exception allowing the Department of Artificial Intelligence to finance research involving certain AI capabilities when used for defensive cybersecurity measures.

Sanders has previously advocated federal legal action in another technology-and-information dispute involving climate claims. In 2015, he asked the Justice Department to investigate possible corporate fraud by Exxon Mobil over discrepancies between what the company allegedly knew about climate change and what it told shareholders and the public. In 2023, Sanders and other senators again urged the Justice Department to bring civil cases against fossil-fuel companies over what they described as a coordinated misinformation campaign, alleging potential violations of racketeering, consumer-protection and truth-in-advertising laws. Those were allegations advanced by Sanders and the other lawmakers, not judicial findings presented in those requests.

The new AI legislation represents a substantially broader regulatory approach in a different field. Rather than addressing allegedly unlawful conduct after it occurs, the bill would require advance government permission for development and distribution of covered advanced models, federal approval before deployment, continuing inspections and monitoring, and mandatory intervention when systems exhibit capabilities the legislation identifies as precursors to superintelligence.

Whether those provisions become law will depend on Congress and the normal legislative process. As introduced, however, the Sanders-Casar proposal sets out one of the most expansive federal frameworks yet proposed for controlling advanced AI: a permanent prohibition on artificial superintelligence combined with a federal charter system governing development of the most computationally intensive models.

Original article