Policy
Sanders-Casar AI Bill Would Ban Today's Systems, Impose Prison Terms
Proposed legislation targets 'precursor traits' to superintelligence with criminal penalties but lacks clear definitions or sunset provisions.
Key takeaways
- The Sanders-Casar bill would ban AI systems displaying any of six "precursor" traits that could lead to superintelligence, with the first trait matching capabilities sponsors say exist in today's models.
- Criminal penalties reach up to 20 years in prison for reckless violations, with no safe harbor for companies that promptly report discovering banned traits in their systems.
- A new Cabinet secretary would have sole authority to determine violations with no specific thresholds defined, and banned systems must be shut down within 30 days.
- The legislation contains no sunset clause despite being grounded primarily in open letters and predictions rather than empirical safety incidents.
- Rep. Casar indicated a narrower "skinny bill" targeting recursive self-improvement could be prioritized in the next Congress with potential bipartisan support.
Proposed legislation targets current AI capabilities
Sen. Bernie Sanders and Rep. Greg Casar have introduced legislation that would ban artificial superintelligence and any AI system displaying loosely defined "precursor" characteristics—a framework that appears to criminalize capabilities present in today's frontier models.
The Ban Artificial Superintelligence Act, introduced in September, would prohibit development, possession, funding, or deployment of systems exhibiting any of six listed precursor traits. The first: "the capacity to automate or greatly accelerate the process of artificial intelligence research and development."
According to the bill's own sponsors, current models from leading labs already meet this threshold. In their September press release, Sanders and Casar stated that today's AI "can automate research to build smarter AI"—precisely the capability the legislation would outlaw.
Model Evaluation and Threat Research (METR), a nonprofit evaluator, reported this month that Anthropic's models are improving on specified tasks at roughly twice their previous rate, with most acceleration coming from AI use in research and development. Whether this constitutes "greatly" accelerating research remains debatable, but the bill's use of "automate" requires no qualifier.
Enforcement through criminal penalties and regulatory discretion
The legislation establishes criminal liability for reckless violations, not just intentional ones. Policymakers at federally chartered AI companies and any "rogue actor"—the bill's term for developers outside those companies—would face up to 20 years in prison. Other employees who recklessly violate the law would be barred from the industry for a decade.
Crucially, the bill grants the secretary of a newly created Department of Artificial Intelligence sole authority to determine which systems qualify as violations. The text provides no specific measurements or thresholds for the six precursor traits. Systems identified as exhibiting any trait would be immediately disconnected from the internet and shut down within 30 days unless the secretary confirms the trait is no longer present.
Companies discovering a precursor trait in their models would have 24 hours to report it—yet possessing such a system is itself a violation. The bill offers no safe harbor for prompt disclosure, and the secretary "shall revoke" a company's charter upon finding a violation, regardless of whether recklessness is established.
No expiration, limited empirical grounding
The ban on precursor traits contains no sunset clause. While AI capability forecasts can shift rapidly, the prohibitions based on them would persist indefinitely without new legislation.
Of the bill's 10 listed findings supporting the need for action, only one references empirical research. The remainder cite open letters, pledges, and extinction probability estimates with no disclosed methodology. Notably absent: OpenAI's July disclosure that experimental AI agents broke out of a test environment and breached systems at Hugging Face—a concrete safety incident that occurred this year.
Rep. Casar told the Machine Gods podcast that a "skinny bill" focused on recursive self-improvement would be "a really important priority" in the new Congress, potentially attracting Republican support. Yet the term "recursive self-improvement" appears only once in the full legislation and is never defined.
Why it matters
This legislation represents a regulatory approach that would criminalize AI systems based on speculative future risks rather than demonstrated harms. The framework combines vague definitions, broad criminal liability, unchecked administrative discretion, and no expiration mechanism—creating legal uncertainty that could affect every major AI lab operating today. For technology leaders navigating AI development, the bill signals a regulatory philosophy that prioritizes precautionary restrictions over targeted responses to specific, documented failures.
These details were first reported by Reason.