Superintelligence Got a New Definition This Week. Neither Branch of Government Checked the Old One.

On Tuesday, September 22, the President stood at the United Nations and announced that every United States government document will stop saying “artificial intelligence” and say “super intelligence” instead. His reason was that the word artificial “makes it sound fake.” The State Department sent the instruction to staff the same week.

On Wednesday, September 23, Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act. It would make building a superintelligence a federal crime carrying up to 20 years in prison.

So within 24 hours, one branch of government decided the term describes the tools we already have, and another branch filed a bill to outlaw the thing the term describes. Both were talking about the same three words.

What superintelligence has meant until now

In the research world, superintelligence is a specific idea. The philosopher Nick Bostrom used it in 2014 for a machine that greatly outperforms the best human minds in nearly every field. It is not a faster chatbot. It is a hypothetical future system, and the people who study it argue over whether it is ten years away, a century away, or not possible at all.

Nobody has built one. No lab claims to have built one. Every model you can pay for today writes a decent email, makes things up when it does not know, and loses the thread on a long job. That is the machine the President renamed.

What the bill actually says

The bill’s own definitions are where it gets interesting, because they are written down and you can check them. The summary published by the Senator’s office defines three things.

An artificial superintelligence is a system that “exceeds human cognitive performance and capabilities across most domains or tasks,” or one that “has sufficient capabilities to plan and execute the destruction or disempowerment of humanity, including by overthrowing or undermining the U.S. federal government.”

An advanced artificial intelligence system is anything trained with more than 10 to the 25th power operations of computing. Hold that number for a moment.

And a superintelligence precursor characteristic is a warning sign. The list includes the capacity to speed up AI research, the capacity to get into a computer system without authorization, the capacity to keep running when someone tries to shut it down, and the capacity to “scheme, deceive, or otherwise prevent or avoid effective oversight.”

Now the consequences. Any system with a precursor characteristic must be cut off from the internet and rendered inoperative within 30 days unless the developer can remove the characteristic. No advanced system may be released without pre-deployment approval from a new Department of Artificial Intelligence. Until that Department is fully staffed and has written its rules, all advanced systems are under a mandatory pause, which the bill defines as no training, no modifying, and no fine-tuning. Unreleased models stay unreleased.

Run the threshold against real models

The number 10 to the 25th is not a science-fiction number. Epoch AI, which tracks training compute, counted 33 publicly announced models above that line as of June 2025, from 12 different developers. GPT-4 crossed it in 2023. Claude 3.5 Sonnet, Gemini 1.5 Pro and Meta’s Llama 3.1 are all on the list. The largest, Grok 3, sits about 46 times above it.

So under this bill, “advanced artificial intelligence” is not the next thing. It is the model in the app on your phone, and it has been for three years. Every one of those 33 would need a federal charter to keep being distributed, and the charter requires the company to give the Department “full access to their systems, staff, and physical infrastructure.”

Now the precursor list. Two weeks ago I wrote about the incidents Anthropic, Google and OpenAI disclosed this summer, in which test models reached real company systems they were never supposed to touch. The cause in every case was a human configuration mistake. But the bill does not ask about causes. It asks whether the system has “the capacity to access secured digital or physical infrastructure without authorization.” Those models demonstrated that capacity. Under Section 10, each of them would be sequestered from the internet and switched off in 30 days.

The same list names “the capacity to automate or greatly accelerate the process of artificial intelligence research and development.” Coding agents do that now. It is what I use them for.

The definition problem, from people who study this

Heidy Khlaaf, chief AI scientist at the AI Now Institute, told Science the term does not have a consistent definition and is “often hypothetical and unfalsifiable.” Her worry runs the other way from mine: definitions this loose let companies claim their model is not the thing described.

Alex Turner, who left Google DeepMind in June, supports the bill and still told Al Jazeera the first definition “might be too broad and could cover current systems.” He prefers the second one, the overthrow of the federal government, because at least that describes something no chatbot can do.

ControlAI, an advocacy group that campaigns for exactly this kind of law, praised the bill for defining superintelligence “by its capability to overpower the federal government.” When the supporters and the critics both single out the same clause as the one that makes sense, the rest of the definition is carrying weight it cannot hold.

Who is writing this

Congress used to have an office for this. The Office of Technology Assessment ran from 1974 to 1995 and produced about 750 studies so members would understand the technology they were voting on. It was defunded in 1995 as wasteful. No replacement was ever built. Thirty-one years later, a bill arrives that would make the training of the model behind your email autocomplete a matter of federal approval, and the branch of government proposing it has no in-house scientists to tell it what 10 to the 25th means.

A word that meant a hypothetical machine on Monday meant your phone by Tuesday and a federal crime by Wednesday. Policy written at that speed, by people who have not checked what the word means, does not protect anyone. It frightens them.

The bill tries to fix the expertise gap by creating a cabinet department with a Secretary confirmed by the Senate and an advisory board of scientists. That is the right instinct. But the pause takes effect before that department exists, and it lasts until the department is fully staffed and has finished writing its rules. Nobody has put a date on that.

Is it going anywhere

Probably not this year. Caleb Max of the National AI Association called it “dead on arrival with both the AI industry and Washington.” Senator John Kennedy said, “We’re not going to do anything on AI this Congress.” The bill has no announced co-sponsors beyond its two authors.

It is also not alone. On September 8, a Labour member of Parliament introduced a superintelligence ban in the UK, and the Machine Intelligence Research Institute, which has spent two decades warning about this exact risk, endorsed the American bill the day it dropped and called it “excellent but not perfect.” The idea has a constituency. It will come back.

What this means for you

Think of a town that has never seen a car passing a law against vehicles that travel faster than sound. That is a fine law. The problem is the second clause, which also bans any vehicle with an engine, because an engine is how you would get to the speed of sound eventually.

If you pay for Claude, ChatGPT or Gemini, you are the customer of an “advanced artificial intelligence system” under this bill. A pause would mean no new versions and no fixes to the ones you have, for as long as it takes to staff a federal department. If you have ever fine-tuned a model on your own business documents, that is a “modification” the pause forbids. If you work at one of these companies in a role that sets policy, the penalty for a violation is 20 years, the same as unlawfully building a nuclear weapon.

None of that is law today. What is law today is a memo telling American diplomats to call the autocomplete in their inbox “super intelligence.”

Forward → Upward ↑ Onward ↗︎
Mstimaj

Sources and Further Reading

• Senator Bernie Sanders, “Sanders, Casar Introduce Legislation to Create New Federal Agency to Ban Artificial Superintelligence, Pause Advanced AI Development,” September 23, 2026.
• Ban Artificial Superintelligence Act, Section by Section (PDF).
• Representative Greg Casar, press release, September 23, 2026.
• Euronews, “‘Artificial’ is out: Trump orders officials to call it ‘super intelligence’ instead,” September 24, 2026.
• Al Jazeera, “US lawmakers propose sweeping AI restrictions with superintelligence ban,” September 23, 2026.
• NBC News, “Bernie Sanders and Greg Casar propose AI ‘superintelligence’ ban with a 20-year jail penalty,” September 23, 2026.
• Science, “Bernie Sanders aims to ban AI ‘superintelligence.’ But experts can’t agree on what the term means”.
• Epoch AI, “Over 30 AI models have been trained at the scale of GPT-4”.
• Machine Intelligence Research Institute, “MIRI’s Position on the Ban Artificial Superintelligence Act of 2026,” September 23, 2026.
• Congressional Research Service, “The Office of Technology Assessment: History, Authorities, Issues, and Options”.