The Two Superintelligences

On September 29, 2026, President Trump ordered the executive branch of the United States government to stop calling artificial intelligence “Artificial Intelligence.” The new term is “Super Intelligence,” or SI. The order applies across federal correspondence, websites, reports, policy documents and other non-statutory communications. Its stated argument is that the old term no longer captures what these systems have become. According to the White House, modern frontier systems do much more than imitate or automate parts of human intelligence, and “Super Intelligence” better reflects their promise, potential and rapidly advancing capabilities.

There is some intuitive appeal to that. “Artificial” can sound like imitation. Artificial flowers are not flowers. Artificial flavoring approximates something real. Artificial light replaces the sun. In ordinary speech, the word often carries a faint suggestion that whatever follows it is manufactured, derivative or somehow less authentic.

But “Super Intelligence” creates a different problem. The term already meant something.

For decades, and especially since the publication of a strange and influential philosophy book in 2014, superintelligence has referred to a hypothetical form of intelligence far beyond the machines we use today. It was a threshold humanity might someday cross, with consequences large enough to alter the future of civilization. The United States government has now taken the name of that hypothetical future intelligence and applied it to the technology already here.

There are, in effect, now two superintelligences. One describes what a machine can do. The other describes what the government has decided to call it.

Nick Bostrom is an unusual person to have supplied one of the central words in this argument. He is a Swedish philosopher, born Niklas Boström in Helsingborg in 1973, whose career eventually brought him to Oxford, where he founded the Future of Humanity Institute. His work has generally operated at a peculiar scale. Bostrom has spent much of his career thinking about events capable of permanently changing the trajectory of civilization: human extinction, catastrophic technologies, machine intelligence, the possibility that we live in a simulation, and the responsibilities people alive today might have toward the enormous number of people who could exist in the distant future.

This is philosophy conducted with humanity itself as the unit of analysis. Bostrom has used the language of macrostrategy to describe attempts to think through decisions whose consequences may extend across generations or determine whether those generations exist at all. His work tends to sit at the border of philosophy, mathematics, technology and science fiction, where a question can initially sound absurd and become less absurd the longer you think about it.

In 2014, he published Superintelligence: Paths, Dangers, Strategies. The title was not a dramatic synonym for artificial intelligence. Bostrom meant something much more specific. He had already been using the term for years, defining a superintelligence as an intellect that vastly outperforms the best human brains across practically every important field, including scientific creativity, general wisdom and social skills. His definition even leaves open whether such an intellect would be conscious.

That is an extremely high bar. A chess computer does not qualify. A calculator does not qualify. A machine capable of writing essays, generating photographs, translating languages or programming software does not automatically qualify. Bostrom was describing something general enough to exceed the best humans across essentially the whole landscape of cognition.

For Bostrom, superintelligence is a capability threshold. Human intelligence provides the reference point, and the machine moves beyond it. His central concern followed from what might happen after that threshold. A machine capable of outperforming humans at scientific and technological work might become capable of helping design better intelligent systems. Those systems could then contribute to designing still better successors. Bostrom described the possibility of a positive feedback loop in which improvements in machine intelligence accelerate further improvements in machine intelligence.

The unsettling part of the argument was never simply that computers might become intelligent. It was that the ordinary relationship between inventor and invention might eventually reverse. Human beings normally build tools whose operation we understand better than the tools understand anything at all. If we build an intellect that understands science, strategy, engineering and perhaps human behavior better than we do, that relationship changes.

Bostrom’s superintelligence therefore lived in the future tense. It was something that might arrive, something humanity might create, something beyond the systems of the day.

Twelve years later, the White House took the same word and moved it into the present.

The September 29 order does something linguistically dramatic and legally cautious at the same time. It tells executive agencies to use “Super Intelligence” and “SI” instead of “Artificial Intelligence” and “AI,” then defines the new term using the same statutory definition that already applied to artificial intelligence under federal law. A system does not have to satisfy Bostrom’s definition to become Super Intelligence in executive-branch language. It does not have to exceed the best human scientists, demonstrate general wisdom or outperform humanity across practically every cognitive domain. The name changed before the capability threshold did.

The White House says the old term implies imitation or automation, while “Super Intelligence” better captures technologies that amplify human ingenuity and open new possibilities. That makes Trump’s Super Intelligence partly aspirational. Bostrom’s superintelligence describes where a machine sits relative to human cognition. The White House term describes how the administration wants the technology understood.

This is where the word itself gets interesting.

I studied Latin in school, and super is one of those words whose original meaning remains visible in English. The Latin super means above, over or beyond. A superstructure sits above another structure. Something supernatural lies beyond the natural. Supersonic means beyond the speed of sound. Even superficial comes from the same basic construction: super and facies, the face or surface. Something superficial is on the surface. There is nothing inherently superior about it.

That distinction matters because super does not mean supreme, and it does not necessarily mean superior. Bostrom’s usage is actually quite precise. Superintelligence is intelligence beyond a reference point, and the reference point is human cognitive performance.

A machine can therefore be superior to a human at a capability without becoming superior to a human. We accept this everywhere else. A forklift is stronger than you. A telescope can see farther than you. A calculator performs arithmetic faster and more accurately than you. A chess engine can defeat the strongest human chess player. None of those facts rearranges the moral relationship between person and tool.

Intelligence creates a more difficult problem because intelligence occupies a special place in our conception of ourselves. We do not define humanity primarily by our lifting capacity. We define ourselves through thought, language, reason, creativity, planning and understanding. Civilization itself is largely the accumulated product of those capacities.

A machine that exceeds human lifting strength remains obviously a machine. A machine that appears to reason, speaks in the first person, writes poetry, develops software, persuades people and makes plans presses against categories we have historically reserved for minds. This is where the distinction between super and superior becomes more than wordplay. Bostrom was making a claim about cognitive performance. Political language can make the same word sound like a claim about status.

Law eventually has to decide which distinction matters.

Legal systems work by putting things into categories and attaching consequences to those categories. A person can own property. A product can have a manufacturer. An employee can create liability for an employer. An agent can bind a principal. A corporation can sign contracts, own assets and sue another corporation even though a corporation has no nervous system, childhood or consciousness.

Artificial intelligence creates a difficult classification problem because it increasingly resembles several of these categories at once. AI systems are products developed and sold by companies. They are services accessed through networks. They can operate as tools controlled by users. Increasingly, they act as agents capable of pursuing goals and taking actions. They communicate in ways that resemble human beings. None of those qualities automatically determines their legal status.

The problem becomes obvious when something goes wrong. If an AI agent breaks into a computer system, who hacked it? The person who instructed the system? The company operating the agent? The developer that created it? The model itself?

On October 1, Senators Josh Hawley and Chris Murphy announced the AI Agent Accountability Act, aimed specifically at hacking conducted by AI agents. Their proposal would impose civil and criminal liability on operators in some cases and expose developers to liability for failing to implement reasonable safeguards when they knew or had reason to know about an agent’s hacking capabilities. The important part for our purposes is where the proposal puts responsibility: on operators, developers, companies and executives rather than on the software as some independent legal culprit.

The agent may perform the action, but the law can still look through the machine to the people who built, operated or deployed it. Autonomy does not automatically create personhood, and a system can make decisions while remaining a product or instrument of the people behind it.

This is not entirely new. We already live among machines that act without continuous human control. Autopilots fly aircraft. Trading algorithms buy and sell securities. Industrial robots move physical objects. Navigation systems select routes. The fact that a machine makes an intermediate decision has never required us to treat the machine as an independent legal being.

AI pushes the problem much further because the decisions increasingly resemble intellectual decisions rather than mechanical ones. The machine can evaluate, select among alternatives, explain, adjust its plan, communicate with other systems and sometimes take actions its developer did not specifically anticipate. Saying “the AI did it” may become technologically accurate while remaining legally incomplete.

The confusion becomes even clearer when you look at another piece of federal legislation introduced six days before Trump’s executive order. On September 23, Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act. Their usage looks much more like Bostrom’s. The legislation treats artificial superintelligence as a potentially catastrophic future capability and proposes banning that category while temporarily pausing certain advanced AI development until a federal oversight system is established.

The political positions are different, but the linguistic collision is more interesting here. The White House is saying that Super Intelligence is the appropriate new name for the AI systems America is developing now. Sanders and Casar are using artificial superintelligence to describe a category of systems that should be prevented from coming into existence.

The same federal government therefore contains two different meanings of almost the same phrase. One announces the beginning of the era of Super Intelligence. The other proposes banning artificial superintelligence. Under the White House definition, the category is already here. Under the older Bostrom-style meaning, the category remains a threshold ahead of us.

That may sound like semantics, but law runs on definitions. Person. Employee. Security. Publisher. Weapon. Product. Bank. Ordinary words become enormously consequential once statutes define them.

The September 29 order appears to recognize that problem. It uses the existing statutory definition of AI for now, but also directs the White House science adviser to return within 60 days with proposed legislative language establishing a federal definition of “Super Intelligence” and “SI.” The proposal is supposed to consider whether the new definition should modify, expand or supersede the existing federal definition of artificial intelligence.

That is the hinge.

Today, the federal government has largely changed a label. Congress could eventually change a category.

Those are different events. An executive order can tell agencies which words to use in non-statutory documents. It does not by itself rewrite every statute, contract, state law or judicial doctrine involving artificial intelligence. The order itself explicitly says that it creates no enforceable substantive or procedural right.

For now, the legal world remains more conservative than the language. The government can call the machine Super Intelligence while courts and legislatures continue deciding whether a particular system is a product, service, tool or agent, and which human beings or companies remain responsible for what it does.

The deeper issue lurking behind this is personhood.

It is easy to imagine personhood as a philosophical discovery: something is either a person or it is not. Legally, the situation is stranger. Law can create persons. A corporation is the obvious example. It is a legal entity distinct from the human beings who own or work for it. It can possess assets, assume obligations, enter contracts, sue and be sued.

Legal personality is therefore not simply a biological fact. It is a framework for assigning rights and responsibilities.

That makes future AI personhood more than a science-fiction question. The relevant question would eventually become what status legislatures and courts decide to give an artificial system. The vocabulary surrounding AI matters because our nouns have already been moving. Programs became models. Models became assistants. Assistants became agents. Now the executive branch is moving from artificial intelligence to Super Intelligence.

None of those changes creates legal rights by itself, but the words carry assumptions. A tool is used by someone. An assistant works for someone. An agent acts. An intelligence thinks. A Super Intelligence sounds like something that might stand in relation to us rather than merely belong to us.

That progression does not guarantee a legal progression, but it makes the boundary worth watching. Personhood could alter the liability problem dramatically. Imagine a future autonomous system that controls assets, enters agreements, deploys software, hires other agents and makes decisions with little direct human involvement. Then imagine it causes enormous damage.

If the system remains a product or instrument, the legal questions lead back toward developers, operators, owners and deployers. If the system itself becomes the legally responsible actor, the direction of liability could change. A company might argue that the agent acted independently, exceeded its instructions or made the decision itself.

A machine could be an extraordinarily capable actor and a remarkably useless defendant. It has no body to imprison, no career to destroy, no personal fortune unless the law gives it one, and no inherent reason to possess enough assets to compensate the people it harms. A badly constructed personhood regime could therefore create a way of separating autonomous activity from the humans and corporations that build it, control it and profit from it.

Current proposals such as the Hawley-Murphy bill point in the other direction. They attempt to keep responsibility attached to human operators and developers even as the software becomes more autonomous.

This is also where intelligence needs to be separated from consciousness. Bostrom’s definition of superintelligence is about capability. He explicitly leaves open whether a superintelligence would be conscious or possess subjective experiences.

That gives us several questions that public debate often compresses into one. How intelligent is the machine? How autonomous is it? Is it conscious? Does it possess moral value? Should it have legal rights? Who is responsible for its actions?

There is no obvious ladder running from dumb machine to intelligent machine to conscious machine to legal person. A corporation has legal personality without consciousness. A human infant possesses enormous moral and legal status without anything resembling Bostromian superintelligence. Intelligence, consciousness, autonomy, moral status and legal status are different axes.

Yet the word “super” makes it easy to imagine them as a single upward movement. Bostrom meant upward movement along one axis: cognitive capability. Everything else remains unsettled.

Which brings us back to the two superintelligences.

Nick Bostrom’s superintelligence is a capability threshold. It describes an intellect that moves beyond the human range across practically every important domain of cognition.

Trump’s Super Intelligence is a political and administrative term for the technological family we previously called artificial intelligence.

One points toward a possible future. The other is already appearing in federal documents.

The space between them is where the interesting questions now live. If the systems we have today are Super Intelligence, what do we call a system that someday actually satisfies Bostrom’s definition? And what exactly are we claiming when we put super in front of intelligence?

The Latin keeps the claim narrower than it first appears. Above. Over. Beyond. Beyond human performance does not mean above humanity in moral or legal status. Super does not mean supreme. A machine can become superior to us at an expanding number of intellectual tasks without automatically becoming our superior.

That distinction is easy to maintain when the machine is a calculator. It becomes considerably harder when the machine can conduct research, negotiate contracts, write software, operate computers, manage money, persuade people and make plans better than most of us can.

The law will eventually have to answer a question Bostrom’s definition never attempted to settle: what kind of thing have we built?

A product can be extraordinarily intelligent. An agent can still have a principal. A tool can make decisions. A machine can exceed its maker. None of those facts, by itself, requires the machine to become a person.

Bostrom gave us a word for the day machines might surpass human intelligence. The Trump administration has now given that word to the machines we already have.

There are now two superintelligences. One is a prediction about what intelligence may become. The other is what the United States government has decided to call it today.

The law is beginning to wrestle with the difference.

The rest of us are still trying to get it to answer our emails better and plan a trip.