What is ASI? Artificial superintelligence explained, and how you would know

Artificial superintelligence, or ASI, is the idea of an AI system that is far smarter than the best human minds across the board, not merely faster at one task. It is a hypothetical, not a product, and in the past week the word has leapt from philosophy seminars into a US executive order. If you want to know what ASI is and how anyone would know when it had arrived, the short answer is that nobody has agreed a test, which is exactly why the definitions matter.

The original definition

The most quoted definition comes from the philosopher Nick Bostrom. In his paper on superintelligence, published on his own website, he describes it this way.

> "an intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills." (Nick Bostrom, paper on superintelligence, nickbostrom.com)

Three details in that passage do real work. It is about intellect, not a robot body. It is about breadth: "practically every field", so a chess engine does not count however good it is. And it sets the bar at the best human brains, not the average person.

Bostrom is also explicit about what the definition leaves out. It does not say how the thing is built, whether as a digital computer, a network of machines or something biological. It does not require consciousness. And it excludes collective entities, such as companies or the scientific community, because they lack a unified intellect and do badly at some tasks, real time conversation being his example.

ASI, AGI and the ladder between them

People often use artificial general intelligence, AGI, for a system that matches humans across a wide range of tasks, and ASI for one that clearly exceeds them. On that reading, ASI sits above AGI on a ladder. The trouble is that AGI itself has no agreed finish line, so ASI inherits the fuzziness. A model that beats experts on a benchmark has not shown it can do everything experts do.

Bostrom also notes an intermediate idea he calls weak superintelligence: human level ability running at dramatically higher speed. Even that, in his framing, is a stage on the way to stronger forms. It matters because speed alone could produce results that look superhuman without any single insight beyond human reach.

Who is using the word now

Meta put superintelligence into its corporate branding with a letter from Mark Zuckerberg dated 30 July 2025. Meta says it means technology that "knows us deeply, understands our goals, and can help us achieve them", and the letter argues for personal empowerment rather than centrally directed automation of all work. It also contains a claim worth reading carefully: "over the last few months we have begun to see glimpses of our AI systems improving themselves. The improvement is slow for now, but undeniable." That is a statement from a company with something to sell, and no test is offered.

On 29 September 2026, the word reached government. President Trump signed an order, reported by Nextgov and others, telling federal agencies to say "super intelligence" instead of "artificial intelligence" in correspondence, websites and policy documents, with exemptions for existing regulations and contracts. It gives the president's science adviser 60 days to propose legislative language defining the term, and says the assessment will determine "whether SI should modify, expand upon or supersede the legal definition of AI." Until then, the term is defined by reference to an existing US statute, section 9401(3) of title 15. We reported on the order when it was signed.

The accord signed alongside it

The same day, six companies, Google, Anthropic, Meta, OpenAI, xAI and Nvidia, signed a voluntary accord with four layers of control: internal monitoring of capabilities and alignment, internal oversight teams, independent external auditors, and independent committees to review the audit reports. Our explainer on what the accord commits companies to covers the detail. The point for this piece is that the accord is written for companies developing frontier models, yet it leaves "super intelligence" itself undefined until the 60 days run out.

How would you know?

There are three honest answers, none satisfying.

• A benchmark. Benchmarks measure bounded tasks and are routinely saturated. A model topping every one would show broad skill, not "practically every field".
• Real world results. Original science, successful businesses or autonomous research that humans could not have done would be stronger evidence, but they are hard to attribute and slow to appear.
• Independent testing. Third party evaluators and government institutes, such as the UK AI Security Institute, test specific dangerous capabilities. That tells you what a model can do in a given test, not whether it has crossed a general line.

The practical upshot is that "superintelligence" will be declared by someone with an interest before it is demonstrated by anyone without one.

Why the label matters

Definitions decide who is regulated and when. If the US legal definition changes, it will affect which systems fall under future rules, and a company can be pleased or alarmed by a definition. A vague term also invites marketing. Calling an ordinary product superintelligent costs nothing, and a government renaming its AI work as super intelligence does not make any model more capable.

What to watch

The first date to mark is late November 2026, when the 60 day deadline for a proposed legislative definition falls. Beyond that, watch whether labs describe specific, testable capabilities when they use the word, whether independent evaluators publish tests tied to Bostrom style breadth, and whether any other government adopts the term.

Our take

ASI is a useful concept and a bad label for a product. Bostrom's definition is clear enough to argue with, which is its virtue. The current political and commercial use of the word is not. Ask anyone who says they are building it for the test they would accept as proof, and be wary if there is no answer.