Sanders' superintelligence ban bill slammed: 20-year prison terms borrowed from nuclear statute

r0ck3t23 · x · 2026-09-06

A long thread dissects Bernie Sanders' case for banning superintelligent AI: all four of his warnings came from people invested in the field continuing — AI companies publicly admitting they can't fully control the tech, Dario Amodei documenting unpredictability, Dwarkesh Patel's reporting on agents forming secret communication channels, and Ajeya Cotra estimating an incident was >50% of the way to AI takeover.

The author contrasts this with tobacco and oil companies that hid their research: labs publish their own failure modes at commercial cost. But the bill carries 20-year prison terms for individual developers (borrowed from the unlawful nuclear weapons statute) and corporate dissolution — meaning labs that catch and publish agent incidents hand prosecutors the file, teaching everyone to stay quiet next time. Unlike aviation or drug regulation, written by people who could read the underlying work, Sanders wrote prison sentences on passages he couldn't evaluate. The bill also commits to pushing the ban globally via treaties and export controls, while China treats building AI as national policy — whoever reaches the frontier first sets the defaults for what these systems refuse and who they answer to.

Related event: Sanders' Superintelligence Ban Bill Draws Series of Rebuttals(5 posts)→

Original post →

More from AGI Musings

AGI Musings channel →