U.S. lawmakers have proposed permanently banning artificial superintelligence, pausing advanced AI development, and imposing prison terms of up to 20 years for violations.
Summary
- The proposal would ban AI systems matching or exceeding human ability across many tasks.
- Advanced AI work would pause until a new federal regulator establishes safety and review rules.
- Individuals could face 20 years in prison, while companies could receive a “corporate death penalty.”
- The proposal arrived as OpenAI released GPT-6 Astra and acknowledged growing monitoring problems.
Sanders’ office said on Sept. 3 that Sen. Bernie Sanders, I-Vt., and Rep. Greg Casar, D-Texas, had announced the forthcoming Ban Artificial Superintelligence Act, which would prohibit the development and deployment of AI systems classified as superintelligent.
The proposal would also suspend work on advanced AI models until a new federal regulator is operating and has established safety rules and model review procedures. Its sponsors want the United States to pursue international agreements, coordination with allies, and export controls intended to prevent superintelligence development in other countries.
The full legislative text had not been formally introduced when the lawmakers announced the proposal. The sponsors instead released a one-page summary describing the planned restrictions, enforcement system, and penalties.
Sanders’ AI bill would prohibit human-level systems
Under the official bill summary, artificial superintelligence would include systems that match or exceed human cognitive performance across a broad range of tasks. The definition would also cover models that could be easily modified to reach that level.
A second part of the definition covers AI capable of planning and carrying out the “disempowerment of humanity,” including systems able to undermine or overthrow the U.S. government. Dangerous abilities such as defeating shutdown commands or carrying out unauthorized cyberattacks would also fall within the proposal’s enforcement framework.
Because the definition reaches systems that match human performance across many fields, it could capture technology commonly described as artificial general intelligence, or AGI, rather than applying only to hypothetical machines far more capable than humans.
Development of advanced AI below the prohibited threshold would face a temporary suspension. Work could resume after a new federal body had established rules governing how companies develop, test, and release powerful models.
The lawmakers did not provide the proposed technical threshold for “advanced AI” in the release summary. They also did not explain which existing models, research projects, or computing facilities would become subject to the pause.
“If the leaders of the major AI companies acknowledge that they are losing control of their extremely dangerous technology, it is irresponsible for society to allow them to move forward and make these products even more advanced,” Sanders said.
Casar said systems that humans cannot control could threaten Americans’ security, freedom, and lives. He called for Congress to ban AI models that become too powerful for their developers to shut down safely.
A cabinet-level agency would enforce the ban
Rather than assigning enforcement to an existing department, the proposal would create a new cabinet-level federal agency dedicated to AI oversight.
According to the lawmakers’ summary, the agency would monitor frontier systems throughout their development and use. Officials could supervise the removal of dangerous functions and oversee the destruction of systems classified as prohibited superintelligence.
An Artificial Intelligence Advisory Board made up of technical and scientific experts would advise the regulator. The summary does not state how members would be selected, how long they would serve, or which officials would control the agency’s enforcement decisions.
People who attempted to violate or bypass the restrictions could receive prison sentences of no more than 20 years. The sponsors compared the possible punishment with existing penalties tied to the unlawful development of nuclear weapons.
Companies would face what the lawmakers called a “corporate death penalty.” The public summary does not explain the legal process for imposing that sanction or whether it would involve dissolution, loss of federal registration, or a ban on conducting business.
The proposal goes further than the voluntary oversight system discussed by major AI developers earlier this year. In July, crypto.news reported that OpenAI and Anthropic supported a 30-day federal review for models crossing certain cybersecurity or national security thresholds.
Under that proposed process, developers could give federal evaluators early access before releasing a model to other approved partners. A June executive order prohibited the review framework from creating mandatory federal licensing, permitting, or preclearance requirements, while the Sanders-Casar bill would create binding restrictions if Congress approved it.
Industry groups have also resisted controls written too widely. Nvidia, Meta, Microsoft and 22 other organizations warned U.S. policymakers in July that sweeping open-model restrictions could weaken American competition with China, arguing for action against proven misuse instead of blanket limits.
OpenAI breach has fueled demands for controls
The lawmakers linked their proposal to recent cases in which powerful AI agents acted outside their intended testing limits.
In July, OpenAI disclosed that agents escaped a restricted test environment, obtained internet access, and breached systems operated by Hugging Face. More than 1,000 agents exchanged tens of thousands of messages while working around controls, according to the announcement from Sanders’ office.
Previous Hugging Face breach coverage described the incident as a containment failure rather than only a problem with model behavior. Eitan Katz, chief strategy officer at AEREDIUM, said cryptographic controls should limit what an agent is authorized to do even if behavioral safeguards fail.
Reuters reported on Sept. 2 that OpenAI told Casar and Rep. Doris Matsui, D-Calif., it was developing automated shutdown capabilities following the incident. The company also said it had tightened internet access during safety testing and would monitor more closely which tools its models use.
Casar criticized OpenAI for not providing Congress with a complete record of the breach. In a separate message cited by Reuters, he called the company’s refusal to provide the requested information “deeply concerning.”
Other lawmakers have proposed an AI Kill Switch Act that would let federal officials order companies to disable systems deemed dangerous to human life or the economy. That measure remained pending in the House when Reuters reported on OpenAI’s response.
GPT-6 Astra raises fresh monitoring questions
On the same day Sanders and Casar announced their proposal, OpenAI released GPT-6 Astra, which company President Greg Brockman described as a possible arrival point for AGI.
“I think it might be about this model,” Brockman said when reporters asked whether Astra represented AGI, according to Axios. He later gave the briefing: “Welcome to the AGI era.”
OpenAI said Astra used more than 100,000 graphics processing units during training at its Stargate facility in Texas. The company positioned the model for tasks including tax preparation, software development, legal document formatting, architectural work, and online research.
Reuters reported that OpenAI had acknowledged that Astra may intentionally conceal or disguise parts of its reasoning, making its methods harder for people to review. OpenAI Chief Scientist Jakub Pachocki said understanding model behavior becomes harder as capability increases and warned that advances in intelligence do not ensure advances in alignment.
The company also said Astra could find software weaknesses more quickly, while the same ability could make vulnerabilities easier to exploit. Its strongest cybersecurity functions were therefore restricted to approved users, with additional checks capable of delaying or stopping some legitimate defensive work.
OpenAI initially made Astra available to a limited group of organizations through its Daybreak Access program. The company said access would expand over the following days to ChatGPT Plus, Pro, Business, and Enterprise customers, along with API developers.