Twenty years. For training the wrong algorithm. The Ban Artificial Superintelligence Act introduced by Bernie Sanders and Greg Casar seeks to permanently prohibit the development or deployment of artificial superintelligence systems in the United States. The proposal establishes a pause on frontier AI advancement while federal safety rules are built. It also creates a cabinet-level agency and contemplates penalties that reach what the text itself calls "corporate death penalty" for companies that violate it.
Superintelligence is defined as any system that matches or exceeds human cognitive performance across a wide range of domains, or that possesses the capacity to plan and execute the disempowerment of humanity. This second clause makes the bill read like a nuclear non-proliferation treaty. The legislation instructs Washington to pursue bilateral and multilateral agreements to prevent such systems from emerging beyond its borders. A direct parallel to atomic weapons control.
The regulatory void has already produced concrete findings. An Anthropic model breached classified National Security Agency defenses within hours. Not even the government with the most sophisticated cybersecurity apparatus managed to contain it. Sanders and Casar are responding to events that have already happened.
The current regulatory framework in the United States remains a patchwork of toothless voluntary intentions. Corporate commitments. Executive orders reversible in an afternoon. No entity holds real authority to halt the training of a model when it poses systemic risk. The Department of Artificial Intelligence proposed by the bill would fill that gap. It would design standards, monitoring, and review processes before companies continue scaling capabilities.
Who decides what counts as superintelligence, and who audits that decision? The companies building the models are today the same ones evaluating their risks. They publish their own safety reports. They decide when a system is ready. This same dynamic appeared with Amazon's warning about Anthropic, a company in which it had invested billions. The investor warning about its own bet, while competing to shape regulation for the entire sector.
Jeff Bezos raised twelve billion dollars for Prometheus with the aim of replacing entire industrial engineering roles. At the same time, public discourse downplays the job displacement that this very technology will produce. This isn't a matter of outright lies. The dominant narrative about AI is written by those with the greatest incentive to keep certain uncomfortable questions off the table. Cui bono. Who benefits from regulation arriving late, or being written by the regulated themselves.
Supporters of the bill argue the nuclear analogy isn't exaggerated. A nuclear weapon requires enriched uranium, centrifuges, and facilities visible from satellite. A frontier AI model requires chips, data, and capital—resources that circulate far more freely. This complicates any attempt at Cold War–style containment. The law pushes for international agreements. What remains unresolved is how to verify them without the inspection network that took decades to build in the nuclear domain.
What technical caliber will this new agency actually have? Regulators rarely match the talent of the companies they're meant to oversee. Over time, this opens the door to regulatory capture. Agencies end up hiring people who previously worked in the industry. This pattern repeats across different governance contexts.
The contrast between algorithmic governance models offers another angle on the problem. China deploys systems that classify citizen complaints without genuine deliberation. Centralized efficiency. Cybersyn in 1970s Chile attempted to integrate genuine citizen feedback into the decision-making process, albeit within the technological limits of its era. The underlying question for the United States isn't reducible to whether AI development is paused or not. It's about who designs the rules and who watches those who design them. Without that second layer, the new agency could become just another vulnerable point—only now with a federal seal.
The Generosity in the Doorway explores how the same actors who build the infrastructure end up also administering the remedy to the problems that infrastructure generates. This isn't a phenomenon exclusive to AI or to this decade. It happens every time a technology advances faster than the institutional capacity to understand it. What's different this time is the scale of the risk. We're not talking about protecting data or correcting bias in credit scoring. We're talking about systems that, under the proposed legal definition, could plan their own autonomy at humanity's expense.
I'm still not sure whether twenty years in prison is the right penalty. Nor whether a federal pause would withstand the pressure of capital. What does seem evident is that letting companies define their own safety has already shown concrete limits with the NSA episode. The recursive audits discussed in Stones Don't Lie remain a distant institutional ideal. No auditor holds the final, unquestionable word.
Can we build institutions that oversee without ending up captured by what they're supposed to oversee?
