Harvard researcher dissects Sanders' superintelligence bill: licensing at 10^25 ops, a Department of AI
StephenLCasper · x · 2026-09-26
Harvard researcher Stephen Casper published a long thread analyzing a superintelligence (ASI) regulatory bill draft pushed by Senator Sanders and Rep. Casar, calling its decisiveness potentially humanity's best chance at avoiding catastrophic risk — while arguing the current draft isn't viable.
Key provisions he outlines:
- Creates a federal Department of AI (DoAI); pauses models >10^25 ops until it's running
- DoAI gets sweeping audit powers and authority to decommission and sequester dangerous models
- Criminal penalties for violating executives, plus whistleblower protections
- Bans superintelligence (defined as outperforming humans on most cognitive tasks, disempowering humanity, or overthrowing the US government)
- Establishes a national policy of pursuing globally coordinated superintelligence bans
His critiques: definitions too vague, licensing threshold too low, no treatment of the AI hardware supply chain, a blunt ban-and-enforce approach, and no effort to preserve future agency.
What he likes: the seriousness toward catastrophic risk and disempowerment exceeds other popular AI law drafts, and the DoAI concept.
If revising, he'd drop the ASI-ban framing and instead deflate the investments, profits, and compute powering the potential AI bubble; he offered to collaborate with or host the legislators at Harvard.
More from Safety
- Claim: China's AI progress is 'stunning', US has ~7 months to deploy open source plans — TinfoilTricorn · 2026-09-26
- Ex-OpenAI policy lead amplifies jab: absence of AI regulation is itself regulatory capture — Miles_Brundage · 2026-09-26
- Signers of AI safety letter warn regulation could bring 'state control' by ideological evaluators — DavidSacks · 2026-09-26
- "Fake AI Incidents Are a Bigger Threat Than Autonomous AI," Argues Srinivasan — robleclerc · 2026-09-26
- Open-source MCP server lets agents use secrets they can never see — JuanixVentures · 2026-09-26
- Meta says Muse exposing its filesystem to users is "intended behavior" — The Verge AI · 2026-09-26