ShrikeShrikeRuntime Governance
Reference · Updated 2026-10-01

What Is Super Intelligence (SI)?

Super Intelligence, abbreviated SI, is the term United States executive agencies now use in place of Artificial Intelligence. An executive order signed September 29, 2026 defines SI as the technologies and systems already covered by the statutory definition of artificial intelligence at 15 U.S.C. 9401(3). It renames the field without changing its scope.

Two different things are now called super intelligence, and a reader cannot tell them apart from the phrase alone. One is a federal naming decision. The other is a research idea that predates it by decades. This page separates them, quotes the order rather than summarizing it, and is explicit about which claims come from the order and which come from reporting.

What does the executive order actually say?

Three passages carry the whole of it. On scope, section 3:

"For purposes of this order, and except where otherwise provided by law, the terms 'Super Intelligence' and 'SI' mean the technologies and systems encompassed by the term 'artificial intelligence' as defined in section 9401(3) of title 15, United States Code."

That sentence is the most consequential one in the document, and it is the one most often left out of coverage. The new term is pointed at an existing statutory definition. Nothing is added to coverage and nothing is removed from it. A document reading SI today describes exactly the set of systems it would have described reading AI yesterday.

On where the term applies, section 2:

"executive departments and agencies (agencies) shall use 'Super Intelligence' and 'SI' in place of 'Artificial Intelligence' and 'AI' in official correspondence, public communications, websites, reports, policy documents, and other non-statutory documents."

Non-statutory is the operative word, and section 3's "except where otherwise provided by law" says the same thing from the other direction. Statutes keep their existing language. So the two vocabularies will run in parallel: an agency web page saying SI, citing a statute saying artificial intelligence, describing one set of systems.

On what happens next, section 3(b) gives a deadline:

"the Assistant to the President for Science and Technology (APST)...shall submit to the President proposed legislative language to establish a Federal definition of 'Super Intelligence' and 'SI' that reflects the capabilities described in section 1 of this order."

Sixty days from a September 29 signature places that submission in late November 2026. It is proposed legislative language, not a definition in force. Until Congress acts on something like it, the statutory artificial intelligence definition remains the operative one.

What Super Intelligence does not mean

The research literature has used superintelligence for decades to name a hypothetical system whose general capability exceeds human capability. That idea has its own arguments, its own critics, and no settled definition.

The order does not adopt that idea, define it, or gesture at a capability threshold a system must cross to qualify. Under the order, a document classifier that sorts invoices is SI, for the same reason it was AI. Treating a federal use of SI as a statement about frontier capability misreads it, and so does treating it as a new regulatory category.

This matters practically for anyone reading a federal solicitation, a grant program description or an agency policy document over the next year. SI in those documents is a vocabulary change. Whether a particular program targets frontier systems is a question about that program, and the word tells you nothing either way.

What is the White House Accord on Super Intelligence?

A separate document, signed the same day by the leaders of several companies that train frontier models, committing them voluntarily to four layers of control.

A caveat that belongs in the open rather than a footnote: the accord is not published alongside the executive order, and the fact sheet accompanying the order does not reproduce it. What follows is therefore drawn from reporting by the outlets cited at the end of this page, not quoted from the document. Where the order is quoted above, the accord is summarized here, and the difference is deliberate.

As reported, each signatory commits to four layers:

  1. Internal controls that monitor model capability and alignment during training and deployment, in areas including cybersecurity.
  2. An internal team responsible for verifying that those controls work.
  3. A partnership with independent external auditors or evaluators.
  4. An independent committee of the board that reviews what the internal team and the external auditors report.

Reporting also notes that signatories agreed to meet regularly on shared standards, that the commitment was characterized as morally binding rather than legally enforceable, and that the text contemplates these steps being codified into law or regulation over time.

Who does the accord apply to?

Its signatories, who train frontier models. It is voluntary, and it is not regulation.

An organization that deploys agents built on somebody else's model takes on no obligation from it. Anyone telling you otherwise is selling something. The honest reading is narrower and still useful: the accord is the first widely endorsed structure for what oversight of a capable system is supposed to look like, and structures of that kind travel. A buyer, an auditor or a board can reasonably ask a deployment to describe itself against four layers that six frontier labs have publicly accepted for themselves.

What would demonstrating those four layers require?

Read as an engineering question rather than a policy one, the four layers are a sequence, and each one constrains the next.

Layer one includes deployment, not only training. Training-time evaluation produces a result about a model; deployment-time control produces a decision about an action. Those are different artifacts and only the second one accumulates.

Layer two asks whether the controls work, which cannot be answered by the absence of incidents. A control that stopped running produces silence that looks exactly like a control with nothing to do. Answering it needs a record of decisions the control actually made, including the ones where it allowed something, because a refusal-only log cannot distinguish a quiet control from an absent one.

Layer three is the binding constraint on the other three. An independent external auditor cannot be handed model internals: they are proprietary, and interpreting them is not what an auditor does. Whatever evidence satisfies layer three has to be legible from outside the model, which pushes the useful record toward the boundary where actions are adjudicated rather than toward the weights.

Layer four needs that same material summarized for people who are accountable without being practitioners. A board committee reviewing reports is not reading decision logs.

None of that is a conformance standard, and nothing here is a checklist to be certified against. It is the shape the four layers imply, written down so a reader can judge a claim against it.

Sources

Shrike builds the deployment-time half of this picture: an action governance layer that adjudicates what an agent does before it does it, and records every decision, including the allows, so the record reads from outside the model. See action governance for what the layer does, or the quickstart to run it against your own agent.