Superintelligence and today's tools: keeping the terms apart
Why the word 'superintelligence' in a brand name, a headline, or a policy document does not make a current product superintelligent, and how published definitions actually differ.
By Adam Green10 min read
This site's name contains the word 'superintelligent'. That is a brand, not a claim. We do not state or imply that any product we cover is superintelligent. This explainer is about why that distinction matters, and how to read the word when other people use it.
What do people mean by these terms?
The word is used for at least three different things, and arguments tend to go wrong when they are mixed.
| Referent | What it is | Can it be tested? | Example of how it appears |
|---|---|---|---|
| Current intelligent software | Products available today. | Yes. They can be tried, measured and compared. | A coding assistant, a research tool. |
| Research on machine intelligence beyond human performance | A research topic with differing definitions. | Partly. Narrow capabilities can be measured; the general concept lacks agreed tests. | Papers on levels of AGI or paths to superintelligence. |
| Speculation about future capabilities | Predictions, hopes and fears. | No. They cannot be verified today. | Forecasts, opinion pieces, slogans. |
Three different referents of the word. Editorial classification.
How do published definitions of superintelligence differ?
Researchers and institutions define terms such as general intelligence and superintelligence differently, and the definitions have changed over time. The sources below were opened for this article; each is quoted only briefly or paraphrased.
| Source | Framing | What it does not do |
|---|---|---|
| Bostrom, 1998 paper "How Long Before Superintelligence?" | An intellect much smarter than the best human brains in practically every field; not tied to a physical substrate or to consciousness. Groups such as companies are excluded because they are not single intellects. | Give a test for when the threshold is crossed. |
| Legg and Hutter, 2007 | Distil expert definitions of intelligence into one formal measure applicable to any machine. | Offer a practical benchmark for today's products. |
| Morris and others, Google DeepMind, 2023 (revised 2025) | Two axes, performance depth and generality breadth. Performance levels run from Emerging through Competent, Expert and Exceptional to Superhuman. The authors assessed leading language models of 2023 as Emerging general AI, and marked higher general levels as not yet achieved. | Speak to 2026 systems; the assessment is dated. |
| Genewein and others, Google DeepMind, 2026 | Describe artificial general superintelligence, intuitively, as more capable than large organisations of humans, and discuss four possible pathways: scaling AGI, paradigm shifts, recursive self-improvement and multi-agent collectives. | Settle which pathway, if any, is real. We read the abstract only. |
| Meta, 2025 memo | "Personal superintelligence": AI that helps each individual achieve their goals, contrasted with a centrally directed alternative. | Define the term technically or propose a test. |
| 15 U.S.C. 9401(3) | Defines artificial intelligence, not superintelligence: a machine-based system that makes predictions, recommendations or decisions for human-defined objectives. | Mention superintelligence. |
Selected definitions and framings. Paraphrased from the cited sources.
Notice what the differences are. Bostrom compares with the best individual humans; the 2026 DeepMind paper compares with large organisations. Morris and colleagues make it a matter of degree on two axes instead of a single line. Meta's usage is a product vision. A claim that something "is" or "will soon be" superintelligent is meaningless until you know which of these is meant.
Where do today's systems sit on published frameworks?
The most concrete framework in our sample is the Levels of AGI paper. In it, narrow systems such as AlphaFold, AlphaZero and Stockfish reach the Superhuman performance level, meaning they outperform all humans, but only on narrow tasks. For general systems, the authors rated the leading language models of September 2023 at Level 1, Emerging, and marked the higher general levels as not yet achieved.
Two cautions. First, that assessment is from 2023 and is the authors' judgment, not a measurement, so it should not be read as a statement about any 2026 product. Second, the framework's value is the vocabulary: it separates how well a system performs from how broadly. A tool can be superhuman at a narrow task without being general, and general without being expert. Marketing language usually blurs the two.
Do labels change what a product can do?
A change in terminology in a speech, a marketing page, or a policy document does not change what a product can do. Capability is established by testing, which is what our methodology describes.
A current example is policy language. Executive Order 14434 directs US executive agencies to say Super Intelligence in place of AI in non-statutory documents where law allows, yet it points to the existing statutory definition of artificial intelligence. That is a naming decision for federal documents. We set it out section by section in what Executive Order 14434 changes for AI buyers and examined the surrounding claims in Trump wants to rename AI Super Intelligence. A related lesson on vendor releases is in the month AI got a new name.
What should I ask when I see the word?
- Who is saying it?A researcher, a vendor, a regulator, a journalist? Each has different incentives and different standards of proof.
- What do they mean by it?Beyond which comparison class: the best individual, an expert, an organisation? Is it a single system or a collective?
- Is it about a product, a research direction or a forecast?Only the first can be tested today.
- What is the evidence?Look for defined tasks, published conditions, repeatable results and independent checks. See reading benchmark claims critically.
- What would change your mind?If no result could count against the claim, it is a slogan, not a finding.
How does this site handle the terms?
- Product pages describe measured behaviour only, once measured.
- Research explainers are labelled as editorial interpretation unless they cite sources, and sources are labelled by type.
- Policy entries record what an official source says, with a link, and not what commentators say it means. See the policy tracker.
- We use no ranking or score language for "intelligence". Dimensions such as capability and reliability are scored separately, and only with evidence. See how to read an evaluation.
- Definitions of terms we use are collected in the glossary.
Frequently asked questions
What does superintelligence mean?+
Generally, hypothetical machine intelligence that greatly exceeds human performance across most areas. Nick Bostrom's 1998 definition describes an intellect much smarter than the best human brains in practically every field, but researchers and institutions define it differently.
Is any product on this site superintelligent?+
We do not state or imply that any product is. The word is part of our brand name, not a claim.
What is the difference between AGI and superintelligence?+
AGI usually means artificial intelligence with human-level generality; superintelligence means going well beyond human performance. Google DeepMind's Levels of AGI framework places superhuman performance at its top level and marks it as not yet achieved for general systems.
Does a policy that uses the term change what today's tools can do?+
No. Terminology in a policy or marketing page does not change a product's capabilities. For example, Executive Order 14434 directs US agencies to say Super Intelligence in official documents but defines it by reference to the existing statutory definition of artificial intelligence.
Is personal superintelligence a technical term?+
Not as far as we can verify. Meta's July 2025 memo uses the phrase to describe a vision of AI that helps each individual pursue their goals. The page gives a vision, not a technical definition or test.
How would anyone know a system was superintelligent?+
Only by evidence: defined tasks, published conditions, repeatable measurements and independent checks. No such evidence base exists for the term today, which is why we treat it as a research topic and a label.
Sources
- Levels of AGI for Operationalizing Progress on the Path to AGI — Morris and others, Google DeepMind, arXiv 2311.02462 (v5, September 2025) (accessed October 14, 2026). Third-party research. Its assessment of then-current systems dates from 2023.
- How Long Before Superintelligence? — Nick Bostrom, International Journal of Futures Studies, 1998 (accessed October 14, 2026). Source of the definition quoted. We did not verify the wording in his 2014 book.
- Universal Intelligence: A Definition of Machine Intelligence — Legg and Hutter, arXiv 0712.3329, 2007 (accessed October 14, 2026). Third-party research; abstract read.
- From AGI to ASI — Genewein and others, Google DeepMind, arXiv 2606.12683, June 2026 (accessed October 14, 2026). Third-party research; abstract only.
- Personal Superintelligence — Meta, 30 July 2025 (accessed October 14, 2026). Company statement of vision; vendor-reported, not a technical definition.
- 15 U.S.C. 9401, Definitions — Legal Information Institute, Cornell Law School (accessed October 14, 2026). Statutory definition of artificial intelligence, item (3).
Related reading
- Trump Wants to Rename AI “Super Intelligence.” What Actually Changes?
- Executive Order 14434: what it changes for AI buyers, and what it doesn't
- Reading benchmark claims critically
- How to read an intelligent-software evaluation
- Three Years Watching AI Become Infrastructure
- The Month AI Got a New Name, a New Interface, and Little New Evidence
This piece was first published on BestSuperintelligent.com. Read the original →