AI or SI? The Techno-Autocracy Will Not Be Elected

The short version

• On 29 September 2026 an executive order told US federal agencies to call artificial intelligence "Super Intelligence", and the president signed a one-page accord on superintelligence with executives of Google, Anthropic, Meta, OpenAI, xAI and Nvidia.
• The accord commits each company to internal controls, an internal team, an external auditor it partners with and a board committee, and it sets no penalties and gives the government no enforcement role.
• My argument is that this is the form a techno-autocracy takes: rules for the most powerful technology of the age, written as a compact between an executive and a few firms, with no legislature, no public and no vote.
• What is not established is whether the accord will ever become law, what the FTC investigation confirmed the next day will find, and whether anyone set out to avoid a vote, which no document shows.

Two documents left the White House on Tuesday 29 September, and they belong together.

The first is an executive order, "Inaugurating the Era of Super Intelligence", which tells federal agencies to say "Super Intelligence" instead of "artificial intelligence" in their own documents. The second is a single page of text followed by seven signatures: Donald J. Trump's, and those of Sundar Pichai for Google, Dario Amodei for Anthropic, Mark Zuckerberg for Meta, Greg Brockman for OpenAI, Elon Musk for xAI and Jensen Huang for Nvidia. It commits their companies to watch over themselves. Asked whether it was binding, ABC News reports, the president said it was "morally" binding.

As this publication's founder and editor-in-chief, with a governance and public policy background, I keep returning to one question: does the AI age lead to a techno-autocracy? My answer is yes, and the evidence is already signed. The danger is not that a machine will rule. It is that the institutional form of rule by a few is being written now, in documents, before any superintelligent system exists, and, at the federal level, outside the institutions that vote. The strongest case against this opinion closes the page.

Why did Trump rename artificial intelligence “Super Intelligence”?

Because, in his order's words, the capabilities of today's frontier systems "increasingly represent not merely artificial intelligence, but a new era of Super Intelligence." For now the order changes the word, not the legal object, and it commissions proposed legislation to give the new word a federal definition of its own.

Section 1 makes it policy that the executive branch "will not acknowledge the usage of 'Artificial Intelligence' and 'AI' in any applicable setting", and Section 2 applies the new term to "non-statutory documents within the executive branch." Section 3(a) then defines Super Intelligence, for the order's purposes, as "the technologies and systems encompassed by the term 'artificial intelligence' as defined in section 9401(3) of title 15, United States Code."

That definition is Congress's. Since 1 January 2021 (Public Law 116-283) it has described "a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations or decisions influencing real or virtual environments." The order keeps it for now, "except where otherwise provided by law". But Section 3(b) gives the Assistant to the President for Science and Technology 60 days to "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", with an assessment of whether the new definition "should modify, expand upon, or otherwise supersede the existing statutory definition of 'artificial intelligence'".

So the legal object has not changed yet, but the order points to where it could go: from a machine serving "human-defined objectives" toward a definition that "reflects the capabilities described in section 1", the section that announces an era. The name comes first, and it is doing work.

Does it matter whether we call it AI or SI?

Yes, because the name shapes which institutions answer for the thing. Artificial intelligence, in the statute's sense, is a product: a machine pursuing objectives a person set, answerable to the legislatures that define it, the agencies that police its sale and the courts that hear the people it harms. Superintelligence is framed as an era or a threat, and eras and threats answer to whoever claims to understand them.

The shift is written into the titles, and among the documents I read, the companies used the new vocabulary before the government did.

The 2023 title lists what you demand of a product, that it be safe, secure and trustworthy, and the two stages of its life, development and use. The 2026 titles name stretches of time: an age, a new age, an exponential, an era. Nobody recalls an era, and nobody sues one.

Anthropic supplies the threat. It plans to caution investors in its IPO that advanced AI could pose "catastrophic or existential risks to humanity", Reuters reports. Cohere's co-founder and chief executive, Aidan Gomez, a competitor with his own interest, named the political use of that frame: "Convince a government that AI is an existential threat and you can convince it to outlaw your competition."

The political scientist Robert Dahl had a name for one alternative to democracy: guardianship. In my words, not his, it is the claim that a few, qualified by special knowledge, should govern for everyone because citizens and their representatives cannot grasp the stakes. He set it against the gravest technology of his own time in a 1985 book whose title is the contrast itself, Controlling Nuclear Weapons: Democracy versus Guardianship.

Amodei sees the tension, to his credit. Legislation "moves very slowly," he wrote in June, adding: "Often this is for good reasons: governments have grave powers, and it's usually for the best that they aren't used too hastily." But "in the several years that it can take Congress to act, AI can go from an amusing toy to the full country of geniuses." His answer, regulation modelled on the Federal Aviation Administration, is delegation by the test this piece applies below, and I credit it. Yet the premise he starts from, that a legislature is too slow for the thing, is the one guardianship needs, and an era makes it easier to believe than a product does.

What is the White House Accord on Super Intelligence?

It is a one-page pledge, signed on 29 September, in which the signatories say each company "should implement the following four layers of controls and audits" on its own frontier models. Its subtitle is "Joint Commitment on Frontier Responsibilities". Everything it asks of a company is below, beside the party each layer reports to.

The right-hand column is my reading, and it is the point: every layer reports back into the company. The text adds: "The participating companies will meet regularly to establish standards and best practices to improve the safety of their systems." And: "Over time, it may make sense to codify these steps into laws or regulations."

Is the White House AI accord legally binding?

No. Al Jazeera reports that "The pact, however, is voluntary and not legally binding", that "It does not establish penalties for companies that fail to comply" and that it does not "give the government an enforcement role". The president told reporters it is "morally binding", Al Jazeera adds.

The text bears that out. It names no penalty, no agency, no legislature and no date. The public appears only as an audience, to be given "confidence that the technology is operating as intended." And the commitment is made on the companies' own terms: "Regardless of whether this is required of companies, we believe that implementing these controls and audits is critical to ensuring a safe future for everyone, and each of our companies are committed to doing this."

None of these absences proves that anyone set out to avoid the law. They are absences, and an absence falls on someone: here, on everyone who did not sign.

Who audits AI companies under the White House accord?

The company does the choosing. The accord tells each company to "Partner with an independent external auditor or evaluator", says nothing about anyone else selecting, accrediting or paying that auditor, and, Al Jazeera notes, "does not require companies to publicly disclose audit results".

Before the accord existed, a competitor named the problem. Writing on 13 September about Amodei's roadmap, under which a handful of labs "would agree on shared standards", not about this accord, Gomez warned against "auditors who not only have financial or ideological conflicts of interest with those they audit, but who are handpicked by them", and asked that third parties "never be paid by the party they're reviewing". The accord's text is silent on both.

Who regulates superintelligence in the US?

At the federal level, the rules written for frontier AI come from the president and the companies, through executive orders and a voluntary accord, with the FTC investigating the companies over potential harms to consumers. At the state level, California has enacted a frontier AI law with penalties. In the record below, three entries came out of a legislature: Congress's 2021 definition, a Senate vote that struck an AI provision rather than enacting one, and California's statute.

The June order is explicit that "Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement" for new models. Under the December order, the Justice Department task force's job is to challenge state AI laws that conflict with the order's policy, which section 2 states as "a minimally burdensome national policy framework for AI". We documented that litigation, and what the 99 to 1 vote did and did not decide, in Who Limits AI in America. None of the three companies has confirmed the reported standards body, as our news desk reported, and BankInfoSecurity adds that "if it wants to wield any legal authority, it may need to register with a government agency."

The strongest fact against my argument is in that table, and it is not federal. California's legislature passed SB 53, and its governor approved it on 29 September 2025, a year to the day before the accord. The Transparency in Frontier Artificial Intelligence Act requires a "large frontier developer" to "write, implement, comply with, and clearly and conspicuously publish on its internet website a frontier AI framework", and a frontier developer to report critical safety incidents to the state's Office of Emergency Services, with civil penalties, for a large frontier developer, that do "not exceed one million dollars ($1,000,000) per violation", recovered "only by the Attorney General". The Legislature's own site shows its chapter in the code, "Effective January 1, 2026". That is a legislature doing what this piece asks for, and I say so plainly.

It does not settle the federal question, for three reasons. It is state law, and the December order, besides creating the task force, directs an evaluation to identify state laws "that may compel AI developers or deployers to disclose or report information in a manner that would violate the First Amendment", an evaluation we could not find published when we wrote Who Limits AI in America; the order does not name California, but SB 53 is a disclosure and reporting statute. Its duties are a published framework, which must describe how the developer approaches "Using third parties to assess the potential for catastrophic risks" and "Instituting internal governance practices", plus transparency and incident reports, and whistleblower protections, including an anonymous internal channel whose disclosures go to officers and directors each quarter; nothing in it requires an independent audit or a board committee. And it is not an Act of Congress, which is the first thing that would change my mind.

Is the FTC investigating OpenAI and Anthropic?

Yes. "The Federal Trade Commission has opened a broad probe into the safety of AI systems, including Anthropic and OpenAI, a senior FTC official confirmed to ABC News on Wednesday", 30 September, the day after the accord was signed. ABC reports that it covers "allegations of unfair or deceptive acts by AI companies and potential harms to consumers", and that the agency "could issue formal demands for AI companies to turn over information". Our news report has the detail.

After California's statute, this is the strongest fact against my argument: law pointing the other way, confirmed the day after the pledge. ABC's wording is the language of section 5 of the FTC Act, which declares "unfair or deceptive acts or practices in or affecting commerce" unlawful; that link is my reading, not the agency's statement. An agency acting under that law can demand what a pledge can only promise.

But notice which law bites: the old law of products and consumers, applied after the fact to what companies did. It has not decided what frontier models may be built, under which controls, or who checks them. The accord answers those questions, without a vote. The investigation proves that democratic institutions can still reach these companies, not that they are writing the rules.

How does the EU AI Act differ from the White House accord?

In who wrote it and what stands behind it. The AI Act is "Regulation (EU) 2024/1689 of the European Parliament and of the Council", adopted under the ordinary legislative procedure, and the Commission's enforcement powers over providers of the most advanced models, fines included, have applied since 2 August 2026.

Europe also starts with dialogue and a code of practice, which most providers of the most advanced models have signed, the Commission says. The difference is that in Europe the dialogue happens in the shadow of a law two co-legislators adopted, and in the United States it stands in for one. The Act has gaps of its own, set out in our GPT-6 Astra analysis, but they sit inside a law somebody voted for.

Is voluntary AI self-regulation a first step toward law?

Sometimes, but the American precedent its defenders cite ended in a revocation, not a statute. The July 2023 commitments came with a promise that the administration "will pursue bipartisan legislation". What followed was Executive Order 14110, which Executive Order 14148 revoked on 20 January 2025, as item "(ggg)" on a list. What one signature makes, another unmakes, and the accord is not even an order.

The serious version of the objection is that democracies hand technical questions to experts. They do, and how they do it is the answer. "The Federal Reserve Act of 1913 established the Federal Reserve System," in the Fed's words. The Nuclear Regulatory Commission exists because a statute says "There is established an independent regulatory commission", whose members "shall be appointed by the President, by and with the advice and consent of the Senate." Even the self-regulator Hassabis offers as his model answers to a public agency: "FINRA is registered with the Securities and Exchange Commission, which approves any changes to its rules," BankInfoSecurity notes. A legislature wrote each charter and can rewrite it. That is delegation. When the experts write their own charter, it is guardianship, and on its text the accord is that.

Can AI lead to a techno-autocracy?

Yes, and not in the way the films imagine. By techno-autocracy I mean rule over a technology everyone lives with, by the few who own it, on authority drawn from technical claims rather than consent. On that definition, the institutional form is visible in what was signed and reported in September.

Start with who signed. Besides the model developers, the signatories include Nvidia, which says it "pioneered accelerated computing". Anthropic's draft IPO prospectus is not public and I have not seen it; these figures are as reported by Reuters.

Four of the accord's seven signatures belong to companies that appear in that one filing as Reuters reported it: Anthropic's own, Google's, xAI's and Nvidia's, since what the xAI agreements buy is, in Reuters' words, "Nvidia-based computing capacity". Meta and OpenAI are not among the counterparties Reuters names. That is my reading of Reuters' figures, and it describes a market, not a conspiracy: companies at this scale buy from each other. But it means several of the parties to the accord are each other's counterparties, with the state as the seventh signature.

Then compare what each side asks of the other. OpenAI's paper proposes "new public-private partnership models to finance and accelerate the expansion of energy infrastructure required to power AI", for the grid, through "targeted investment credits, direct and indirect flexible subsidies, or equity stakes", structured "to minimize taxpayer exposure to commercial losses". It also asks that companies be able to share safety information "without running afoul of antitrust or competition constraints, using clear safe harbors". Amodei says the US government needs "to issue a narrow waiver for certain kinds of safety conversations." In fairness, the same OpenAI paper says AI data centers "should pay their own way on energy so that households aren't subsidizing them", and names the risk I am describing: "There is also a risk that the economic gains concentrate within a small number of firms like OpenAI, even as the technology itself becomes more powerful and widely used." Still, OpenAI asks the state for public financing tools, and OpenAI and Anthropic ask it for antitrust room. In September the state signed a pledge with them.

The alignment crosses borders. After the US-China summit the White House said the two leaders "agreed to use the term 'super intelligence' rather than 'artificial intelligence'", and that the two countries agreed "to establish a bilateral communication channel for SI incidents", which Beijing had not confirmed in anything our news desk read. That week, Fortune reports, two American AI chief executives asked "world governments to write shared rules" for AI. So far the shared rules are a term and a channel.

This is why I call it guardianship and not mere concentration. Dahl's guardians claimed to rule from knowledge; these firms hold knowledge and ownership at once, and under the accord they choose who checks. Standards set by a market's incumbents, which is what the accord's regular meetings "to establish standards and best practices" would produce, would become rules for everyone who enters after them if formalised. Hassabis says where his own proposal could lead: "formalisation could quickly follow, meaning that Frontier Models would be required to pass it to be deployed in the US market."

Is AI self-regulation ethical?

Not in the form this accord takes, and the reasons are about legitimacy rather than safety: morality replaces law, the governed never consent, and the auditor is chosen by the audited.

"Morally binding" is a precise phrase. A legal obligation has someone at the other end who can enforce it. A moral one is owed to a conscience. Calling the arrangement moral tells the public who can call it in: the signatories themselves.

Consent is the second problem. Gomez wrote about Amodei's roadmap, not the accord, but his sentences fit it: "There is no public comment period here. There is no consultation, and there is no vote. The public will be forced to live with the outcome regardless." A signatory's own paper states what is missing. Of the transition to superintelligence, OpenAI writes that "we should navigate it through a democratic process that gives people real power to shape the AI future they want", and it proposes "structured ways for public input so that alignment isn't defined only by engineers or executives behind closed doors." The accord creates no such way.

The third is the auditor. An audit is a promise made to someone other than the audited, and when the audited chooses the auditor and need not publish what it finds, the promise is made to itself.

None of this says the signatories will implement the controls badly. They may do it well. The ethical problem is that their conduct, good or bad, answers to nobody who did not sign.

What would democratic oversight of AI look like?

Four changes would move these rules back into institutions that answer to voters, and a fifth needs no one's permission. Each is specific enough for a company to adopt or a regulator to point at.

Row 5 costs a web page. If the signatories believe what they signed, it is the cheapest way to show it.

Can democratic institutions govern superintelligence?

Not at the federal level in the United States, on the current record, and not for want of institutions: a statute defines the technology, a consumer-protection agency confirmed an investigation the day after the accord, California has legislated for frontier developers, and Europe has penalties in law. The capacity to govern is the capacity to decide, and at the federal level the decisions about the frontier are being made somewhere else.

My standing question about AI, and this publication's, has never been whether the systems are capable. It is whether democratic institutions have the capacity to govern what is deployed. At the federal level, September 2026 answered it without them.

Congress defined artificial intelligence as a machine that serves "human-defined objectives". An executive order renamed it an era. A page of text, signed by a president and six companies, committed those companies to watching themselves, and the president called it morally binding. Nothing in that sequence was unlawful, and nothing after Congress's definition was put to a vote. That is how this kind of rule arrives: not by a coup, not by a machine, but by signatures.

Nobody will vote for this techno-autocracy, and nobody will need to.

What documents is this opinion based on?

Primary documents, read in the original on 1 and 2 October 2026, and named outlets' reporting where a sentence says so.

• The accord: the Wikisource transcription, a US government work, checked word for word against the signed scan.
• Executive orders: "Inaugurating the Era of Super Intelligence", 14409, 14365, 14148 and 14110, the last read for its title and its citation, 88 FR 75191.
• Federal statutes and Congress: 15 U.S.C. 9401, 15 U.S.C. 45 with the FTC Act's short title at 15 U.S.C. 58, 42 U.S.C. 5841, the Daily Digest, page D678, and the Fed's page on the Federal Reserve Act.
• California: SB 53's status and chaptered text, and the code chapter it added, including sections 22757.12 and 22757.15.
• White House fact sheets: July 2023 and 25 September 2026.
• The companies' own words: OpenAI, Hassabis, Amodei in June and September, Gomez, and Nvidia's about page.
• The EU AI Act: Regulation (EU) 2024/1689 on EUR-Lex, Article 101(1) on two independent renderings that agree, and the AI Act Service Desk.
• Reporting: Al Jazeera, ABC News, Fortune, Reuters via Yahoo Finance and KSL, and BankInfoSecurity.
• Dahl: Controlling Nuclear Weapons: Democracy versus Guardianship (Syracuse University Press, 1985). I read its catalogue record and its table of contents on Google Books, not its text, so this piece describes his concept and quotes none of his sentences.

What I could not open: Anthropic's prospectus, which is not public; The Information's paywalled report on the standards body; and the Forbes and Axios reports on the accord and the FTC. Nothing above rests on them.

---

This is an opinion piece. It argues a position and it is signed. I do not test AI products, and nothing here rests on testing; it is a reading of public documents, with the reporting I relied on named in the sentence where I relied on it. We use Anthropic's models in our work, and Anthropic is one of the six companies that signed the accord.