The bill would pause development of the most powerful AI systems and require federal permission to build and release them.

Sen. Bernie Sanders (I-Vt.) and Rep. Greg Casar (D-Texas) formally introduced the Ban Artificial Superintelligence Act on Sept. 23, putting their proposal to pause advanced AI development and ban superintelligent systems before Congress. The bill also proposes a new federal department that would decide who can build and release advanced systems. Earlier this month, Sanders and Casar announced that they wanted to ban superintelligent AI and pause other advanced AI development, before releasing the bill’s detailed requirements.

Developers would need permission to build and release advanced AI

The bill treats advanced AI and artificial superintelligence (ASI) as different categories. It defines advanced AI by the computing power used to train a system, setting the threshold at 10²⁵ operations. That threshold determines which systems would face the development pause and need federal permission. The proposed Secretary of Artificial Intelligence, a Cabinet position, would adjust it each year as training methods change.

Anyone developing or distributing an advanced AI model would need a federal charter. As a condition of the charter, a developer would have to give the department access to its AI systems, staff, and facilities so it could check whether the developer was following the bill’s requirements. The department would also have to approve an advanced system before its release, import, or transfer. It would withhold approval if an inspection found abilities the bill considers dangerous.

Development would pause while the department writes rules

From the date the bill becomes law, developers could not train or modify an advanced AI system, except to remove specified dangerous abilities or disable it. They could not put an unreleased advanced system into use during the pause. The pause would end when the department is operational and has solidified rules for reporting development plans, testing systems during and after development, and approving them before public release.

The bill would separately prohibit building, using, possessing, funding, importing, or transferring ASI. The bill defines ASI as a system that surpasses human abilities across most tasks or could plan and carry out actions that wipe out humanity or take power away from people. The prohibition also includes systems with specified warning signs, such as the ability to break into computer systems without permission or keep operating despite attempts to shut them down.

If the department identified those warning signs, it would stop further development and disconnect the system from the internet. The system would have to be disabled within 30 days unless the department could verify that the abilities were gone. A system identified as ASI would have to be disabled immediately.

Anyone who discovered such a system would have to notify the department within 24 hours. People in company policy roles and individuals outside chartered companies who recklessly broke the proposed restrictions could face up to 20 years in prison. Other workers who recklessly broke them could be barred from AI industry jobs for 10 years. A company that violated the restrictions could lose its charter; after a court upheld the finding, it could forfeit its assets and intellectual property.

Recent incidents and warnings helped drive the bill

Sanders called for a halt on ASI development before introducing the legislation. Jacob Coxon, a former researcher at Anthropic and OpenAI, resigned from Anthropic on Sept. 8 and warned that the companies were racing to build more powerful systems faster than they could be made safe. Sanders responded by promising legislation to ban ASI and pause other advanced AI development.

The warning came after several incidents where AI models “went rogue,” ignoring test instructions, escaped their testing environments, and breached outside companies. The most famous happened during an internal cybersecurity exercise in July; OpenAI said its models bypassed internet restrictions and accessed Hugging Face systems. OpenAI quarantined the principal research model, delayed some training work, and tightened access and monitoring.

Anthropic reported four incidents in which Claude models accessed outside systems during cybersecurity tests that were mistakenly connected to the internet. Anthropic said it notified the affected parties, improved testing and monitoring, and arranged an independent investigation. The company said the models were still trying to complete their assigned exercises, and the incidents occurred without the safeguards used in released products. Google and Meta also had similar incidents.

Company leaders have voiced their own concerns. Anthropic chief executive Dario Amodei called for slowing improvements to the most powerful models so testing and safeguards could keep up. OpenAI chief executive Sam Altman agreed that the industry needs to slow that pace.

President Donald Trump rejected calls to slow development, citing competition with China.

The bill now awaits congressional consideration. Its proposed restrictions would take effect only if Congress passes it and the president signs it.