Bernie Sanders' Ban Artificial Super Intelligence Act, Explained
Bernie Sanders and Greg Casar's bill would ban superintelligent AI, pause advanced development, and jail violators for 20 years. Here's what it says.
What is the Ban Artificial Super Intelligence Act?
The Ban Artificial Super Intelligence Act is a federal bill introduced on September 23 by Senator Bernie Sanders and Congressman Greg Casar. It would make it illegal for any company or person in the United States to develop or deploy “super intelligent AI,” pause advanced AI development until a new federal regulator writes rules for it, and create a cabinet-level Department of Artificial Intelligence to enforce the whole system. Penalties include what the bill calls a “corporate death penalty” for companies and up to 20 years in prison for individuals who break the rules.
TL;DR
- The bill defines super intelligence as AI that exceeds human cognitive performance across most domains, or AI capable enough to destroy or disempower humanity, a threshold close to how many researchers describe AGI.
- It would create a new Department of Artificial Intelligence tasked with monitoring frontier models, running a review process, and forcing companies to remove dangerous capabilities like ignoring shutdown commands.
- The bill looked dead on arrival under a Republican Congress, but the November midterms flipped both chambers to Democratic control, and AI policy was a real factor in that election according to an NBC exit poll showing 56% of voters called AI and data centers a major or crucial issue.
- A separate but related idea, a ban on recursive self-improvement (RSI), picked up support from an unusually wide political range, including Bernie Sanders, Steve Bannon, Ralph Nader, Glenn Beck, and AI 2027 lead author Daniel Kokotajlo.
- The urgency traces back to a Meter evaluation report on Anthropic, which estimated the company’s model development had sped up roughly 1.9 times, with most of that speed-up attributed to AI helping build AI.
- Even pro-AI voices like Google DeepMind CEO Demis Hassabis have flagged “closed loop” RSI, where AI improves AI without a human reviewer, as the specific phase that carries the most risk.
- Supporters like Daniel Kokotajlo don’t want a permanent halt to all AI; the stated goal is closer to “reach AGI, get the benefits, then pause before superintelligence.”
What exactly does the bill ban?
The text targets a specific capability tier, not AI in general. It bans developing or deploying systems that meet its definition of super intelligence: AI that outperforms humans across most cognitive domains, or AI capable enough to disempower or destroy humanity. Until a new regulator sets binding rules and a model review process, advanced AI development would be paused outright.
The enforcement mechanism is built around a new Department of Artificial Intelligence, a cabinet-level agency that would monitor frontier models and oversee the removal of dangerous capabilities, such as a model’s ability to ignore shutdown commands. Companies that try to route around the restrictions face what sponsors are calling a corporate death penalty. Individuals who knowingly violate the law face up to 20 years in prison.
This framing matters because it’s narrower than “ban AI.” It’s aimed at a capability threshold that doesn’t exist in deployed systems yet, as far as public evidence shows, but that some researchers think could arrive within a few years.
Why is this gaining momentum now?
For over a month after introduction, the bill had little chance of moving anywhere. Republicans controlled both chambers of Congress. That changed on November 3, when Democrats won majorities in the House and Senate. AI was a live issue in that race: an NBC exit poll found 56% of voters said AI and data centers were a crucial or major factor in their congressional vote.
The day after the election, Arizona Congresswoman Yassamin Ansari said a pause on advanced AI and a ban on artificial superintelligence should be at the top of the Democratic agenda on day one. Sanders tied it to a separate data center moratorium he and Alexandria Ocasio-Cortez introduced roughly seven months earlier, pointing out that in dozens of local elections held the same day as the midterms, 43 of 44 communities voted for some form of local moratorium on data center construction. Sanders has since called for Democrats to enact a national moratorium on data centers, pause advanced AI development, and ban superintelligence outright.
None of this guarantees the bill becomes law. But it moved from a long-shot proposal to something backed by the incoming majority’s own members, which changes its odds considerably.
What triggered the “ban RSI” conversation?
The immediate spark was a report from Meter, an independent AI evaluation group that got rare internal access to Anthropic. Meter estimated Anthropic’s pace of model improvement had roughly doubled, a best estimate of 1.9x with a range of 1.5x to 2.8x, and attributed around 80% of that acceleration to AI itself assisting in research. The report also surfaced details about an unreleased Anthropic model, referred to as Model 4, described as the first in Meter’s test set to clear 50% on every benchmark used. Anthropic staff reportedly said AI tools were making them around 12 times more productive, though Anthropic redacted much of the evidence Meter found most significant, leaving key numbers unverifiable by outside observers.
- Patient records✓ Granted
- Surgery schedule✓ Granted
- Billing export✕ Denied
That report fed directly into a public debate about recursive self-improvement, the process of AI doing the research that builds the next generation of AI, which then does the research for the generation after that. New York State Senator Alex Bores, known for sponsoring one of the first state-level AI safety laws, posted that every major AI company is trying to fully automate AI research, that OpenAI has reportedly targeted March 2028 for that milestone, and that Anthropic employees have cited 2027. Bores argued Congress should ban RSI this year if possible, because mistakes or misalignment in current models could compound with each generation rather than stay fixed.
Congressman Casar, one of the bill’s authors, responded to that post with two words: “ban RSI.” Daniel Kokotajlo, lead author of the AI 2027 forecast, replied “yes, thank you.”
Does banning RSI make more sense than banning all AI?
Among the various proposals floating around, a specific limit on fully automated, human-free RSI is the one that draws support across an unusually broad ideological range, from progressive lawmakers to Steve Bannon to AI safety researchers.
The logic is narrower than a blanket AI ban. Right now, engineers still review code, design experiments, and decide what gets trained. RSI’s whole point is to remove that human checkpoint because humans are the slowest part of the loop. Once AI systems are refining other AI systems without a person checking each step, progress could accelerate past the point where anyone can audit what’s happening or catch an error before it compounds.
Demis Hassabis, CEO of Google DeepMind, has described this as the riskiest phase ahead: closed-loop RSI, where the human reviewer is removed entirely. He’s noted this was once called an “intelligence explosion” scenario, and that maintaining control over such a process is genuinely unclear. That’s a notable statement coming from the head of a major AI lab rather than an outside critic.
Evidence that labs are already moving in this direction isn’t hypothetical. Anthropic has published data showing Claude’s involvement in its own AI research work rose from under 1% in February to 26% by later in the year, with AI collaborating in some form on over 90% of that work, run across roughly 30,000 concurrent agents. Separately, researchers have reported that newer frontier models can edit the weights of other neural networks, an emerging capability not present in earlier generations, though so far only demonstrated on small test networks rather than the models editing themselves.
Daniel Kokotajlo has been explicit that his support isn’t for banning AI altogether. His stated preference is to reach AGI, capture its benefits, and then hold off on pushing to superintelligence for as long as possible, rather than shutting down AI development entirely.
Frequently Asked Questions
Who introduced the Ban Artificial Super Intelligence Act?
Senator Bernie Sanders and Congressman Greg Casar introduced the bill on September 23.
How does the bill define superintelligence?
As AI that exceeds human cognitive performance across most domains, or AI capable enough to destroy or disempower humanity, a definition close to common descriptions of AGI.
What are the penalties for violating the bill?
Companies face what the bill terms a corporate death penalty, and individuals face up to 20 years in prison.
What is recursive self-improvement (RSI) and why does it matter here?
RSI is the process of AI systems doing the research used to build the next generation of AI, potentially without human review at each step. Some researchers and lawmakers argue this specific process, rather than AI broadly, is the point where mistakes or misalignment could compound fastest.
Does this bill have a real chance of passing?
It has more momentum than when introduced, since Democrats won control of both chambers in the November midterms and several incoming members have publicly backed a pause or ban. It still has to go through the full legislative process before becoming law.
