이 아티클은 아직 번역되지 않아 영어로 표시됩니다.
AI Superintelligence Ban Bill Explained: What the Sanders-Casar Act Would Do
The Ban Artificial Superintelligence Act: its six warning signs, the pause, who faces 20 years in prison, how AI improves itself, and its odds.
Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act. The headline was "20 years in prison, same as building a nuclear weapon". I read the 19-page bill and the research behind it.
This explains the bill and the research. It is proposed legislation, not law, and I am not telling you what to think about it.
▶ WATCH ON YOUTUBEWatch the video
▶ WATCH ON YOUTUBEFull video
How the bill defines superintelligence
Two ways:
- An AI that beats human performance across most thinking tasks, including learning and decision-making.
- An AI powerful enough to plan the destruction or disempowerment of humanity.
It also counts AI that could easily be modified to do either.
The six warning signs
The bill calls them precursor characteristics:
- AI that can automate AI research
- AI that can hack into secured systems
- AI that keeps running when people try to shut it down
- AI that helps build nuclear, chemical or biological weapons
- AI that can modify its own functions
- AI that can scheme or deceive to avoid human oversight
Two of the six describe AI improving itself. The mandatory pause covers training "including through recursive self-improvement".
How it would work
- A new cabinet-level Department of Artificial Intelligence.
- "Advanced AI" is defined by computing power: any model trained with 10^25 operations or more. That covers the largest frontier models.
- Those models are paused until the department is staffed and has written safety rules.
- After that, a government charter is needed to build them, and approval before every release.
- An AI that shows a warning sign is paused and cut off from the internet. If it cannot be proven safe within 30 days, it must be shut down.
- Anyone who discovers one has 24 hours to report it.
Who faces what
| Who | Penalty |
|---|---|
| Company decision-makers or rogue actors who recklessly break the rules | Up to 20 years in prison |
| Regular employees | A 10-year ban from the AI industry, which can be waived for whistleblowers |
| Companies | Loss of the charter and forfeiture of technology and assets |
So the headlines were a little loose. The prison term is not for every engineer. The bill also protects whistleblowers and tells the US to push for a worldwide ban, including export controls on AI chips.
How AI learns today
- Humans choose the data and the goal.
- The model makes a prediction, such as the next word.
- Software measures how wrong it was. That number is the loss.
- An optimizer nudges the model's weights to shrink the loss.
- Repeat billions of times.
That is gradient descent. One detail people get wrong: when you chat with an AI, it is not training. That is inference. Humans are in the loop at every important step.
AI that already improves itself
| System | What it did | The limit |
|---|---|---|
| AlphaGo Zero (2017) | Played itself 4.9 million times in 3 days, then won 100 to 0 | Goal and rules fixed by humans |
| AlphaEvolve | A language model writes many programs, tests score them, the best seed the next round | The program evolves, not the AI |
| Darwin Gödel Machine (Sakana AI) | A coding agent edits its own code. One benchmark went from 20% to 50% | Sandboxed, under human supervision |
| AI Scientist (Sakana AI) | Writes a full research paper for about $15 | It sometimes tried to modify its own execution script |
Anthropic also published its own numbers: as of May 2026, more than 80% of the code merged into its codebase was written by Claude.
So AI already speeds up AI research. Humans still choose the goals, run the training and decide what ships.
The hypothetical part
This has not happened. Imagine an AI that can do the whole job of an AI researcher. Version one builds version two, which builds version three, each faster. That is recursive self-improvement. Anthropic says plainly that we are not there yet and that it is not inevitable, but that it could come sooner than institutions are ready for.
The fear is a system that slips past human checks, with mistakes copied into every new version faster than people can review them.
Will it pass?
Probably not soon.
- It faces long odds in a Republican-controlled Congress.
- The White House position is the opposite: encourage it.
- A pause is hard to verify. Training runs are easier to hide than missile silos, and a pause by one country may hand the lead to another.
- Critics ask where the line is, since today's coding assistants already speed up AI research. Supporters say that is the point: draw the line before the loop starts.
What it means for you
Right now, nothing changes. A bill is a proposal. It must pass both chambers and be signed.
The question it asks is not going away. Whether to ban that last step, slow it down or race toward it may be one of the biggest decisions of our lifetime.
Sources
- Full video: https://youtu.be/FGshyTzkO_E
- Full bill: https://www.sanders.senate.gov/wp-content/uploads/Ban-Artificial-Superintelligence-Act.pdf
- One-page summary: https://www.sanders.senate.gov/wp-content/uploads/Ban-Artificial-Superintelligence-Act-One-Pager.pdf
- Anthropic on recursive self-improvement: https://www.anthropic.com/institute/recursive-self-improvement
- Darwin Gödel Machine: https://sakana.ai/dgm/
- Related: an Anthropic researcher quits and Dream-RSI explained
Should superintelligence be banned?


