Advanced AI development would pause, and superintelligent systems would be banned

Advanced AI training, modification, and launches of unreleased systems would pause. Advanced AI means systems trained using at least 10^25 computing operations, with that threshold adjusted annually for technological changes. Work could continue to remove warning signs of superintelligence or disable a system. The Secretary would decide when the department was fully staffed and had issued rules ensuring safe development and deployment. The pause would begin upon enactment and end when the Secretary determines both requirements are met.

People would be barred from developing, deploying, acquiring, possessing, funding, importing, or transferring covered AI. Superintelligence means AI that exceeds human abilities across most tasks or can destroy or disempower humanity, including easily modified systems. Warning signs include greatly accelerating AI research or resisting shutdown. The restrictions would also cover components sufficient to reconstruct a system’s capabilities. Separate bans on deployment, release, transfer, and import would cover systems foreseeably modified to gain banned capabilities.

Officials would isolate or disable identified systems, and anyone discovering banned capabilities would have to report them. These duties would cover identified superintelligence and systems with warning signs, such as resisting shutdown. A mandatory pause would stop training and modification, except work to remove warning signs or disable the system. A system with warning signs could avoid disabling only if the Secretary verifies those signs are gone. Disabling would be required before 30 days after the Secretary identifies warning signs, unless their removal is verified. The reporting deadline would run from discovery.

Advanced AI developers and distributors would need a government charter, meaning permission to operate. These rules would cover advanced AI trained using at least 10^25 computing operations, with the threshold adjusted annually for technological changes. Charter holders would give officials full access to systems, processes, staff, and facilities as officials consider appropriate for enforcement. Officials would inspect systems and withhold approval if they find features that endanger the public.

Violations of the pause, bans, or charter rules would carry criminal, employment, and business penalties. Decision-makers at charter holders and unaffiliated people would face fines and up to 20 years in prison. People outside those decision-making roles would face a 10-year AI job ban; unaffiliated people could face both. Individual penalties would require reckless violations. Officials could waive the job ban for people who report violations if doing so serves the department’s purpose of protecting the public from AI dangers. A violation finding would require charter revocation. A court-upheld finding would expose intellectual property and assets to government seizure.