What Is SI? Super Intelligence and the New U.S. AI Term

Suppose a federal customer sends your SaaS team a request for information about “SI.” Your product still uses the same model and retrieves the same documents. It sometimes makes the same mistakes too. The customer’s vocabulary has changed. As of October 1, 2026, SI means “Super Intelligence” in a new U.S. executive-branch naming policy. In research, superintelligence describes a much higher capability threshold: intelligence substantially beyond humans across almost every relevant domain. Explain the new term with both meanings in view. The September 29 executive order supplies the government meaning; Bostrom and Müller’s discussion of superintelligence supplies the research distinction.

For a software business, start by asking which meaning the buyer intends. A federal communications change can affect the wording of an RFP or a government-facing page. That wording change does not establish that the software has become more capable. Translate the terminology and retain accurate system descriptions. Before changing product claims, identify any actual change in the customer’s requirements.

Two meanings of SI lead to different conclusions#

The new federal use of SI covers technologies already within the statutory definition of artificial intelligence. The existing definition concerns machine-based systems making predictions, recommendations, or decisions for human-defined objectives. It sets no requirement to outperform humans across all fields. Read 15 U.S.C. § 9401(3) when an official reference to SI appears to imply a new class of technology.

The research meaning predates the naming policy. In his 1965 paper on an ultraintelligent machine, statistician I. J. Good considered a machine exceeding human intellectual activity, including the ability to design machines. He proposed that better machine designers could produce still better machines. That “intelligence explosion” was a conditional argument about what might follow from such a capability. It was not a finding that an existing computer had acquired it.

Nick Bostrom’s 2014 book, Superintelligence: Paths, Dangers, Strategies, developed questions about possible paths to machine intelligence, control, and consequences. His earlier explanation of superintelligence described superiority across practically every field, including scientific creativity and social skills. The breadth matters. A system that plays one game beyond the best human player’s ability can be extraordinary at that game without satisfying this broader concept.

AGI, or artificial general intelligence, usually concerns broad capability at a human-comparable level; ASI, or artificial superintelligence, concerns substantially superior capability. Definitions and thresholds vary. Google DeepMind researchers’ Levels of AGI framework separates performance from generality: how well a system performs and how wide a range of tasks it covers. Before accepting a general intelligence claim, ask both “better at what?” and “across how many kinds of work?”

Even the administration’s own earlier 2026 Economic Report of the President distinguished specialized current AI, hypothetical AGI, and artificial superintelligence. A later policy decision to use SI for the existing statutory category changes the administrative vocabulary. It does not, by itself, resolve the technical distinctions that the earlier report discussed.

Judge capability by the task and its conditions. Check the failure rate too. Stanford’s 2026 AI Index technical performance chapter documents rapid progress alongside uneven performance and limitations in evaluation. METR’s research on long software tasks examines how reliably systems complete work of different durations. Neither a good benchmark score nor success at a short coding problem automatically establishes reliable performance over a long, changing business process.

For example, generating a correct SQL query from a clear instruction is different from resolving an ambiguous customer problem. Resolving that problem includes obtaining authorization, querying the right records, handling an exception, and explaining the decision afterward. These capabilities have to work together under operational constraints. Forecasts about when more general systems might arrive also remain forecasts: a survey of thousands of AI researchers reports respondents’ expectations, rather than measuring a present superintelligent system. The naming order cannot replace that evaluation.

The order changes official language and requests a legislative proposal#

The directive applies to the executive branch. Section 2 directs agencies, where legally permitted, to use Super Intelligence and SI in correspondence, public communications, websites, reports, policy documents, and other nonstatutory documents. It does not require alteration of prior regulations, presidential actions, contracts, grants, or historical documents. The order’s implementation section is the controlling text for these boundaries.

Section 3 currently ties SI to the existing statutory AI definition. It also gives the Assistant to the President for Science and Technology 60 days to submit proposed legislative language, including possible changes to that definition and conforming amendments. Section 4 preserves agency and OMB authority, requires implementation consistent with law and available appropriations, and creates no enforceable private right. The definition and general provisions establish these limits. A requested proposal is a policy process; it is not an enacted amendment.

NIST provides an implementation example. NIST’s Super intelligence page says the agency is updating communications under the September 29 order. Meanwhile, the published NIST AI Risk Management Framework 1.0 retains its original name. NIST’s AI RMF resource page also identifies the Generative AI Profile as a companion resource. A communications label can coexist with a document’s established title. Cite the title and version of the framework you actually used.

OMB documents need the same care. M-25-21 addresses federal AI use, governance, and high-impact applications; M-26-04 addresses specified principles for federal acquisition of large language models. Read each document’s substantive instructions on their own terms. A seller should not interpret a customer’s new SI vocabulary as withdrawal of those instructions or as satisfaction of them.

As of October 1, 2026, the renaming order contains no new model performance test, procurement certification, or vendor licensing system. Its additional action is preparation of a possible definition and legislative changes. Any claim that it has already changed statutory obligations requires a separate legal instrument. The White House’s fact sheet explains the administration’s rationale and broader agenda, but the order itself determines what this particular directive requires.

The policy timeline helps identify which changes are substantive#

The September terminology decision is part of a series of federal actions. On January 23, 2025, Executive Order 14179 directed preparation of an AI Action Plan and review of policies adopted under the preceding administration’s AI order. That order established a planning and policy-review process. It did not make the later SI name a technical standard.

On July 23, 2025, the White House released America’s AI Action Plan. Its official announcement described more than 90 actions under three pillars: innovation, infrastructure, and international diplomacy and security. A SaaS team should trace a requirement in a later document to the relevant action or implementing guidance. The action plan’s breadth does not mean every recommendation applies to every software supplier.

Three orders issued that day addressed distinct subjects. The data-center permitting order concerned infrastructure approvals. The American AI technology-stack export order created an export-package initiative. The order titled Preventing Woke AI in the Federal Government addressed principles for certain federal AI acquisitions. Infrastructure approvals, export packages, and purchasing standards can affect a business differently even when they appear in the same press release.

The president’s objection to the old name also predates September 2026. In the official transcript of his July 23, 2025 remarks, Trump said he disliked the word “artificial” and wanted a different description. In September 2026, the White House’s release about his United Nations remarks recorded his preference for super intelligence. At the September 29 America.gov event, he said “the word super is the best word of all” and “it’s not artificial”; the White House recording of the event, preserves those remarks. They explain his stated naming preference, rather than supplying a scientific test of intelligence.

On November 24, 2025, the Genesis Mission order launched a Department of Energy effort to apply advanced computing and AI to scientific discovery. DOE’s mission page describes its research objectives and partnerships. This program concerns an application of the technology: assembling scientific data, computing resources, and tools to accelerate work on national research problems. Assess its operational substance through those resources and projects.

Keep the funding announcements separate. The September 29 White House fact sheet says the federal government has announced more than $5 billion for Genesis. Separately, DOE’s July announcement reported more than $800 million in partner commitments, and its March funding announcement described a $293 million research funding opportunity. Those are different announcements and categories. They should not be added into an invented new total, described as cash already spent, or treated as a single pot available to SaaS vendors. The larger figure is the administration’s reported announced amount. The White House fact sheet is the source for that attribution.

Other actions continued on separate tracks. A December 11, 2025 order called for a national AI policy framework and addressed potential conflicts with state laws. The March 2026 legislative recommendations were recommendations to Congress, rather than an enacted federal AI statute. On June 2, 2026, the advanced AI innovation and security order addressed innovation and cybersecurity. Each has its own scope; none should be replaced in a compliance analysis by the later naming announcement.

Immediately before the naming order, a September 25 White House fact sheet on the U.S.–China state visit reported agreement to use the SI term and establish an SI dialogue. On September 29, a separate America.gov order used super intelligence in a government-services initiative. The wording is entering diplomacy and service delivery. They do not establish that other jurisdictions, standards bodies, or private buyers share one definition.

Industry uses superintelligence for ambitions with different scopes#

Companies also use the word in different ways. OpenAI’s Governance of Superintelligence discussed future systems with capabilities beyond AGI, including possible approaches to oversight. Its Introducing Superalignment article distinguished the challenge of controlling superintelligence from alignment work on existing systems. These articles use the word to discuss a prospective capability and the problems that might accompany it.

OpenAI’s 2026 principles statement also presents superintelligence within its account of future development. Reading an ambition in a company statement is different from seeing an independently assessed capability in a deployed product. For procurement or product messaging, ask for the named model and its available features, along with the evaluation conditions and known limitations. A corporate mission cannot answer those questions by itself.

Meta’s Muse Spark announcement names Meta Superintelligence Labs and describes a path toward personal superintelligence. Safe Superintelligence Inc.’s founding statement makes safe superintelligence the company’s sole mission. The shared word covers different kinds of claims: a laboratory name, a consumer-product ambition, and a research company’s mission. None should be quoted as a government certification of superhuman general capability.

Microsoft’s March 2026 Copilot leadership update describes a superintelligence mission alongside responsibilities for the Copilot product. Microsoft AI’s account of its lab and new MAI models uses a humanist framing focused on serving people. Buyers still need to evaluate the particular service they will receive, even when “superintelligence” names an investment program or a development philosophy.

Safety documents are more useful when they specify actions. OpenAI’s updated Preparedness Framework discusses monitoring capabilities and applying safeguards for frontier systems. Use the framework to ask which risks are measured, which thresholds trigger action, and who makes release decisions. It still needs to be assessed as that company’s published process. It is neither the September naming order nor proof that every product bearing the word SI meets a shared standard.

The September 29 accord concerns voluntary controls#

The White House meeting also produced an accord that needs to be read separately from the naming directive. In his official September 29 remarks, Speaker Mike Johnson identified the White House Accord on Super Intelligence as a voluntary statement of principles and standards. He summarized its expectations as robust internal controls and multiple layers of internal and external review.

The archived text page of the accord describes four layers. Companies should monitor model capabilities and alignment during training and deployment through internal controls. Those controls should cover cybersecurity, biosecurity, chemical threats, and unintended access to technical systems. An internal team should check that controls, monitoring, and detection work. It should also check that problems are addressed. An independent external auditor or evaluator should assess their operation. An independent board committee should receive reports and oversee remediation. These are commitments about how frontier-model companies manage development and review.

The document also calls for participating companies to meet regularly on standards and best practices, and says the steps might eventually be codified into laws or regulations. That possibility is distinct from an enacted requirement. The archived signature page names Trump and representatives of Google, Anthropic, Meta, OpenAI, xAI, and Nvidia. The two images reproduce the document posted by the president; Johnson’s official account confirms its voluntary character. Neither the accord’s wording nor a signature establishes that a participating company’s model has reached research-defined ASI.

If your SaaS team uses a frontier model through an API, ask the provider how its controls apply to the model and service you use. Request information about evaluations, incident handling, changes in safeguards, and the responsibilities allocated to your company. A voluntary public commitment may prompt those questions. It does not answer whether your own retrieval system, permissions, logging, or customer deployment works safely.

Federal sellers should translate the wording and check the requirement#

Suppose an RFP asks for an “SI assistant” to summarize benefits documents. I would answer using the solicitation’s wording, then describe the offer as a document-retrieval and language-model service. The response should identify its model version, supported sources, permissions, and review process. The buyer can then match the response to its preferred term and evaluate the service’s actual behavior. A claim that the product is broadly superhuman would require a very different demonstration.

The relevant acquisition reference remains OMB M-25-22, which addresses federal purchases of AI and emphasizes matters such as performance, risks, vendor documentation, competition, and portability. GSA’s Buy AI guidance directs agencies toward acquisition and security considerations. These are better starting points for a government sales team than a rush to replace every occurrence of AI in a proposal library.

I would update a government-facing terminology note first, then the response templates used for new opportunities. Preserve exact citations, defined terms, and quoted language from the applicable contract or framework. Ask the contracting officer for clarification: does SI refer to the established AI scope, or does the solicitation introduce a distinct capability requirement? That question can prevent a marketing interpretation from becoming an unintended contractual promise.

Government contractors should distinguish agency communications policy from the requirements that apply to their own work. Check the solicitation, its amendments and incorporated clauses, the subcontract, and the task order. FAR Part 27 illustrates why existing provisions on patents, data, and copyrights still need attention when software and model services are procured. A new label does not settle who may use data, what deliverables are owed, or which rights attach to them.

In compliance documents, keep the connection between the customer’s SI terminology and the older framework traceable. A statement can describe the service as SI in the customer’s terminology and name the NIST AI RMF used for its assessment. Record the version and scope of that assessment. The cost is a small amount of editorial maintenance; the benefit is avoiding a global rename that breaks document references or implies capabilities your team has not established.

I would reconsider that limited approach if an agency issued binding follow-up guidance, a contract amendment changed the defined term or requirements, or customers consistently asked for a new naming convention. A general announcement justifies monitoring vocabulary and preparing an explanation. Changing the product name, legal descriptions, model documentation, and sales claims simultaneously needs a stronger basis.

Search content needs the full phrase before the abbreviation#

“SI” already has several meanings. Microsoft’s partner terminology uses SI for a system integrator. The International Bureau of Weights and Measures uses SI for the International System of Units. Sports Illustrated’s SI TV page uses the same letters within its brand. An increase in searches containing SI therefore cannot be assumed to represent interest in super intelligence.

For a B2B SaaS site, I would keep established AI product and solution pages and add a concise definition of the new federal term where it serves the reader. Use the full phrase “super intelligence” beside SI on first mention. A page answering “what is SI” should immediately state its domain and date. It should also distinguish the official naming policy from the research concept so a search result does not carry an exaggerated capability claim.

Measure the vocabulary shift through relevant queries and outcomes. Google’s Search Console guidance describes reports on queries, pages, clicks, and impressions. Compare queries containing the full phrase, agency names, federal procurement language, and your product category. Then examine the landing pages and qualified inquiries. This is a proposed measurement approach, not a claim that demand has already shifted as of October 1, 2026.

Google explains that some queries are anonymized, so the visible query table does not account for every search. Google Trends describes normalized, sampled data, which cannot be read as an absolute count of prospective buyers. A sudden SI spike could reflect interest in a news event, a measurement unit, sports, or an integrator. Read the full query in its category context before changing the content plan.

Use a clear title that answers the question, such as “What Is SI? Super Intelligence Explained.” Google’s title-link guidance favors descriptive, accurate titles and notes that its systems can generate a different displayed title. Explain the ambiguity on the page instead of repeating every possible acronym. Google’s helpful-content guidance similarly directs attention to usefulness for readers. Publish a focused explanation, connect it to relevant existing pages, and revise it when the underlying definition or customer language changes.

State and international terminology still need separate treatment#

California has taken a different terminology position. California’s September 30 announcement describes an order retaining artificial intelligence terminology. The signed Executive Order N-10-26 directs agencies under the governor’s authority to continue using Artificial Intelligence and AI, subject to applicable law. A company serving federal and California customers may therefore need different communication conventions for the same technical system.

Outside the United States, the European Commission’s AI Act guidance continues to describe the Artificial Intelligence Act and AI systems. The OECD’s explanatory memorandum on its updated AI-system definition provides another established definitional reference. Neither source should be silently rewritten as an SI framework in a cross-border compliance document. Explain any translation and preserve the official name of the instrument being discussed.

As of October 1, 2026, the next federal definition, the extent of additional agency implementation guidance, and any eventual congressional amendments remain matters to follow. Johnson’s official remarks say Congress will continue deliberating; they do not establish a legislative outcome. The California order establishes California’s direction, not a conclusion about all states. The White House’s report of a bilateral terminology agreement likewise does not establish international adoption.

A published implementing document or enacted amendment should trigger a review. A forecast about political reaction should not. Until then, keep the jurisdiction, document date, and meaning of SI visible. The business can then follow a customer’s vocabulary and describe what its software does in consistent terms.

Frequently asked questions#