A sitting U.S. senator just proposed treating AI research like manufacturing fentanyl.
The Summary
- Sen. Bernie Sanders and Rep. Greg Casar introduced legislation that would ban the development of artificial superintelligence and impose up to 20 years in prison for violations
- The Ban Artificial Superintelligence Act defines the banned technology as AI capable of the "destruction or disempowerment of humanity," including overthrowing the government
- The bill would also pause development of "frontier models" trained above a specific data threshold, creating a dual-tier regulatory framework
- Sanders framed the legislation around a simple question: "Do we really want to develop a superintelligence that could act independently of human control?"
The Signal
This isn't a policy proposal. It's a line in the sand drawn with no understanding of where the shore ends. The bill attempts to criminalize a category of technology that doesn't yet exist and has no agreed-upon definition in the research community. Artificial superintelligence, or ASI, remains theoretical. The world's leading AI labs are still trying to reach artificial general intelligence, the predecessor state where AI matches human cognitive abilities across domains.
What Sanders and Casar have created is legislation with two fatal problems. First, it criminalizes intent without defining capability. How do you prove someone was developing superintelligence versus very advanced narrow AI? Second, it establishes a "frontier model" threshold based on training data volume, a metric that tells you almost nothing about an AI system's actual capabilities or risks.
"The bill doesn't just ban the more powerful artificial superintelligence; it would also pause the development of advanced AI systems."
The frontier model pause is where this gets interesting for anyone building in the agent economy. If Congress decides that models trained on X petabytes of data need government approval before deployment, every AI lab operating at scale becomes subject to regulatory capture. The bill's language around "specific threshold of data" suggests a hard line, but data volume is a terrible proxy for danger.
GPT-4 was trained on less data than GPT-3 but is demonstrably more capable. Constitutional AI techniques can make smaller models safer than larger ones. Multimodal models combine text, image, and code in ways that make simple data metrics meaningless. A threshold based on training compute or dataset size would either be so low it catches everything useful, or so high it misses the actual risks.
The 20-year prison sentence is the tell. This isn't regulation designed to guide development. It's prohibition masquerading as safety policy. When you threaten researchers with two decades in federal prison, you don't get careful AI development. You get all the serious work moving to Beijing, London, and Tel Aviv. You get American labs that stop publishing, stop collaborating, and start optimizing for legal defensibility instead of capability or safety.
The Implication
If you're building AI agents, this bill won't pass, but the sentiment behind it will shape what does. Expect more proposals that try to draw capability lines in the sand, each one creating new compliance burdens for frontier labs while doing nothing to address actual AI risks like synthetic media, automated hacking, or algorithmic discrimination.
The real policy fight isn't about banning superintelligence. It's about whether the U.S. maintains an environment where AI research can happen openly, or whether we push it underground and offshore. Watch what actually moves forward in committee, not what gets headlines.