What is the legal definition of super intelligence? (2026)

Clerion Research Team20 October 202613 min read

As of October 2026, the legal definition of super intelligence in the US is the existing AI definition in 15 USC 9401(3), adopted by Executive Order 14434. It covers almost any product with a model inside, even a spam filter, and a proposed new definition is due by 28 November 2026.

Near-black and blue blurred vortex gradient banner reading The legal SI definition

As of October 2026, the legal definition of super intelligence in the US is the old definition of AI with a new name. Executive Order 14434, signed 29 September 2026, says "Super Intelligence" and "SI" mean whatever counts as "artificial intelligence" under 15 U.S.C. 9401(3). A new federal definition is due by 28 November 2026.

That surprises most people. The order talks about a new era, so you'd expect a new bar to clear. There isn't one yet. Under the current text, the model behind your product, the recommendation widget on your pricing page and a decade-old spam filter all sit inside the same legal word.

Below: what the text says today, what could change on 28 November, how it stacks up against the EU AI Act and California's frontier AI law, and how to work out whether your own product is "SI" on paper. It's a plain reading of public documents, not legal advice.

Under US federal law today, super intelligence has no separate definition. Section 3(a) of EO 14434 says "Super Intelligence" and "SI" mean "the technologies and systems encompassed by the term 'artificial intelligence'" in 15 U.S.C. 9401(3). That holds until a later presidential action or an Act of Congress replaces it.

The order was published in the Federal Register on 2 October 2026 at 91 FR 63129. It tells executive departments and agencies to use "SI" in place of "AI" in correspondence, websites, reports and other non-statutory documents, "to the maximum extent permitted by law".

Three limits matter for you:

  1. It binds federal agencies, not companies. Arnold & Porter's 13 October analysis says it imposes no obligations on private parties and has no penalty or enforcement provision.
  2. It does not rewrite existing regulations, contracts, grants or past presidential actions.
  3. It creates no enforceable rights for anyone against the United States.

So the word changed. The legal category didn't.

What does 15 U.S.C. 9401(3) actually say?

15 U.S.C. 9401(3) defines artificial intelligence as "a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations or decisions influencing real or virtual environments." It comes from the National Artificial Intelligence Initiative Act of 2020, and it is now the working definition of SI.

The statute on Cornell's Legal Information Institute adds that such systems use machine and human inputs to do three things:

  1. perceive real and virtual environments;
  2. abstract those perceptions into models through automated analysis;
  3. use model inference to formulate options for information or action.

Read that as a builder and notice what's missing. No capability bar, no compute figure, no benchmark, nothing that compares the system to a person. The test only asks what it does: take goals a person set, and spit out predictions, recommendations or decisions that change something. Almost every product with a model inside passes.

Does that mean a spam filter is "super intelligence"?

On the current text, yes. A spam filter is a machine-based system that makes a prediction (spam or not) toward a human-defined objective and changes a virtual environment (your inbox). Under EO 14434 that puts it inside "SI", next to a frontier model.

This isn't a fringe reading. Memeburn's 6 October report ran under the headline that, under the law, a spam filter still counts as super intelligence, and noted the federal government "now calls a spam filter and a frontier model by the same name."

Call it the label gap: how far apart a word's promise and the law behind it sit. Right now it's about as wide as it gets. The preamble talks about systems that go beyond imitating human intelligence, while the operative text covers anything that predicts.

On paper, your autocomplete and the most capable model on Earth are the same thing until 28 November.

No. Legal SI, as of October 2026, means today's AI under a federal statute. Superintelligence and ASI are research terms for a hypothetical system that surpasses humans across nearly every field. No system meets that bar, and no US law uses it as a legal test.

Memeburn's report notes the term already had an established research meaning, a system that outperforms human capability across the board, and that it was unclear whether the White House meant to invoke it. The order's own text settles it for now. It points to the 2020 statute, not the research idea.

If you want the longer version of how SI, AGI and ASI differ as ideas, we covered it in SI vs AGI vs ASI: what's the difference? and in our guide to what SI means. For legal work, keep them apart. When a contract, a procurement form or a regulator says "SI" before 28 November, read it as "AI under 9401(3)".

When will the US get a new definition of super intelligence?

The first draft is due by 28 November 2026. Section 3(b) of EO 14434 gives the Assistant to the President for Science and Technology 60 days to propose legislative language for a federal definition. A proposal is not law: changing the statute still needs an Act of Congress.

Freshfields' 1 October note names the adviser as Michael Kratsios and puts the deadline at 28 November 2026. The proposal has to say whether the new definition should modify, expand or replace the existing AI definition, which other statutes need conforming amendments, and what further executive action is needed.

Two things are still open as of 20 October 2026:

  • Whether the proposal will be public. Arnold & Porter points out the order does not say.
  • Which direction it goes. A narrower definition, one with a capability or compute bar, would push most products out of "SI". A broader one would pull more in. Arnold & Porter's point is that either change could shift which technologies fall under laws that already rely on the current definition.

Kratsios used the term freely in his 8 October remarks ("frontier SI for science models", "SI-ready data") but did not define it. Separately, Memeburn reports, citing the Wall Street Journal, that the new Super Intelligence Force has 120 days to deliver a report on risks and opportunities. We explained that body in What is the Super Intelligence Force?

How does the SI definition compare with the EU AI Act and California's SB 53?

The US SI definition is the broadest of the four you are most likely to meet. The EU AI Act's "AI system" is close to it. The EU's "general-purpose AI model" and California's "frontier model" are narrow and aimed at whoever trains the model, not whoever builds on it.

Definition Where it lives The test Does a hosted-model app built by a 3-person startup fall in?
SI (= AI under 15 U.S.C. 9401(3)) EO 14434, US federal Machine-based system making predictions, recommendations or decisions for human-set objectives Yes, almost certainly
AI system EU AI Act, Article 3(1) Machine-based system with some autonomy that infers from inputs how to generate outputs Usually yes
General-purpose AI model EU AI Act, Article 3(63) A model showing significant generality, able to do a wide range of distinct tasks No, you integrate one, you do not provide it
Frontier model California SB 53 Trained with more than 10^26 integer or floating-point operations No, unless you trained it yourself at that scale
Superintelligence / ASI Research use only Far beyond human ability in almost every domain No, and no product does yet

The EU wording comes from Article 3 of the AI Act: an AI system is "a machine-based system that is designed to operate with varying levels of autonomy" that "infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions". The Commission's July 2025 guidelines, as summarised by LearnSignal, carve out basic data processing, classical heuristics and simple prediction systems, such as one that predicts an average. That carve-out is one thing the US definition does not have.

California's SB 53, signed 29 September 2025 and effective 1 January 2026 according to Orrick, covers models trained with more than 10^26 operations. Its heavier duties fall on "large frontier developers" with more than $500 million in prior-year revenue, with penalties of up to $1 million per violation. The Future of Privacy Forum notes its catastrophic-risk bar: a risk of contributing to 50 or more deaths or serious injuries, or at least $1 billion in damage.

There's a pattern here. Laws that put real duties on companies draw tight lines, by compute, revenue or harm. The broad ones have mostly been about wording and programmes. So if the 28 November draft adds a threshold, it will likely look more like the bottom rows of this table than the top one.

Is my AI product "super intelligence" under the law?

Probably yes under the US definition today, and probably no under any definition that carries duties. If your product uses a model to predict, recommend or decide something for a goal a person set, it fits 9401(3). Unless you trained a model past 10^26 operations, SB 53 does not reach you.

Here is a worked example. Say you run a hypothetical contract-review tool. Three people, built on a hosted model through an API, about 40 paying law firms. You never trained a model yourself.

  1. US, EO 14434. Your tool takes an objective set by a person (flag risky clauses) and produces recommendations that change a virtual environment (the marked-up document). Under 9401(3), it is AI, so on agency paperwork it is now "SI".
  2. EU AI Act, AI system. It infers from inputs how to generate outputs and runs with some autonomy. It very likely counts. Whether any duties follow depends on the risk category, which is a separate question from the definition.
  3. EU AI Act, general-purpose AI model. No. The model provider is the one in scope. You are the downstream integrator.
  4. California SB 53. No. You did not train a frontier model.
  5. Research superintelligence. No.

So what changes in practice? Today the SI label mostly matters if you sell to the US government. Freshfields expects agency solicitations, forms and correspondence to switch to the new terms, so your proposals should be ready to use "SI" where the agency does, while still citing "AI" where a statute, regulation or existing contract uses it. Arnold & Porter makes the same point.

For everyone else, nothing you must do this quarter. Whatever lands after 28 November might change that.

What should founders watch before 28 November?

Watch three things: whether the APST proposal is published, whether it adds a capability or compute threshold, and whether it is written to replace or sit beside 9401(3). Together they decide whether "SI" stays a synonym for AI or turns into a narrower legal class with duties attached.

A short watch list:

Signal Where it would show up Why it matters to you
Proposal text published whitehouse.gov or the Federal Register Until it is public, every summary is a guess
A threshold (compute, capability, revenue) The proposal's definition section Decides whether a small product is in or out
"Replace" vs "expand" 9401(3) The conforming-amendments section Replacing could change scope of existing federal AI programmes
A bill in Congress congress.gov Only Congress can change the statute
Procurement language Agency solicitations The first place a small team will meet the term

Keep one thing in mind while you watch. We argued in Should I rebrand my AI startup to SI? that you should keep "AI" in titles and product names and add "SI" in copy where it helps. Nothing here changes that advice. For now SI is still a new name pointing at an old statute.

Do this tonight

  1. Read Section 3 of EO 14434. It is two short paragraphs, and it is the only text that currently defines SI.
  2. Run your product through the five rows of the comparison table above. Write down where you fall in each one. That is your baseline before 28 November.
  3. If you sell to US agencies, add a line to your proposal template: "SI as defined in EO 14434 Sec. 3(a), equivalent to AI under 15 U.S.C. 9401(3)."
  4. Put 28 November 2026 in your calendar with a link to whitehouse.gov presidential actions, and check for the proposal that week.
  5. Ask ChatGPT, Claude and Perplexity "what is the legal definition of super intelligence?" and note which pages they cite. That tells you which sources the assistants trust on this topic.

When people start searching for the new definition, some of them will ask an assistant first. If you want to see which of your pages ChatGPT, Perplexity, Claude and Gemini send people to, Clerion reports AI assistants as their own traffic channel. Here is how to track ChatGPT traffic to your website.

Frequently asked questions

Does the super intelligence executive order apply to private companies?

No. EO 14434 directs federal executive departments and agencies to use "SI" instead of "AI". It imposes no obligations on private companies, has no penalty provision and creates no enforceable rights. Federal contractors will still see the new term in solicitations and forms.

Who is writing the new legal definition of super intelligence?

The Assistant to the President for Science and Technology, Michael Kratsios, consulting other agencies. Section 3(b) of EO 14434 gives him 60 days from 29 September 2026, so the proposed legislative language is due by 28 November 2026.

Can the White House change the legal definition of AI on its own?

No. The definition in 15 U.S.C. 9401(3) is a statute, so only an Act of Congress can change it. The executive order only asks for proposed legislative language and says SI means AI under that statute until Congress or a later presidential action says otherwise.

Is a chatbot built on the OpenAI or Anthropic API super intelligence under US law?

On the current text, yes. It is a machine-based system that makes predictions or recommendations for human-set objectives, which is the 15 U.S.C. 9401(3) test. It is not a frontier model under California SB 53 unless you trained a model past 10^26 operations yourself.

Is legal super intelligence the same as artificial superintelligence (ASI)?

No. As of October 2026, legal SI means today's AI under a 2020 statute. ASI is a research term for a hypothetical system that outperforms humans across almost every field. No system meets that bar and no US law uses it as a test.

Will the new super intelligence definition be made public?

Unknown. EO 14434 does not say whether the proposal due by 28 November 2026 will be published. Watch whitehouse.gov presidential actions and the Federal Register, and treat any summary as a guess until the text is out.

Sources

legal definition of super intelligencesuper intelligencesiexecutive order 14434ai policyai regulation
20 October 2026 · 13 min read

Clerion is website intelligence for AI founders.

Web analytics, AI traffic, conversions, errors and speed from one script tag.

Start a 14-day trialSee pricing

Keep reading

All writing →

Written from real site data by the team building Clerion. Analytics with no cookie banner and nothing to consent to. Start cookie-free. Nothing charged for 14 days.