The Short Version
AGI, artificial general intelligence, usually means a system that matches people across most or all cognitive work. Superintelligence means a system that goes well beyond that. ASI, artificial superintelligence, is the same idea as superintelligence in most usage, and in one Google DeepMind paper it is a specific level on a scale. None of the three has an agreed definition. Each company uses its own, and since September 2026 the US government has used "Super Intelligence" as its official name for AI in general, which is a fourth meaning.
The Terms Side by Side
| Term | Common meaning | A widely cited wording | Source |
|---|---|---|---|
| AGI | Matches human ability across most or all cognitive tasks | An AI that can match or exceed the cognitive versatility and proficiency of a well-educated adult | Hendrycks and 32 co-authors, "A Definition of AGI", 2025 |
| Superintelligence | Far exceeds human ability across nearly all domains | "An intellect that is much smarter than the best human brains in practically every field" | Nick Bostrom, "How Long Before Superintelligence?", 1998 |
| ASI | Used interchangeably with superintelligence | General capability at a level that outperforms 100% of humans | Google DeepMind authors, "Levels of AGI", 2023 |
| Super Intelligence (SI) | US government term for AI as a whole | The technologies covered by the statutory definition of artificial intelligence in 15 U.S.C. 9401(3) | Executive Order 14434, September 29, 2026 |
How Each Organisation Uses the Words
| Organisation | Preferred term | Its wording | Source |
|---|---|---|---|
| OpenAI | AGI, and since 2025 superintelligence | AGI: "highly autonomous systems that outperform humans at most economically valuable work". Superintelligence is described, not defined: its chief executive wrote of aiming at "superintelligence in the true sense of the word" and of not knowing how far beyond human level intelligence can go | OpenAI Charter. Sam Altman, "Reflections" and "The Gentle Singularity" |
| Google DeepMind | AGI | A scale of five performance levels from Emerging to Superhuman, crossed with narrow or general scope. ASI is the cell that is both general and Superhuman, marked as not yet achieved. Demis Hassabis describes AGI as a system with all the cognitive capabilities the brain has | "Levels of AGI" paper. Hassabis essay, July 14, 2026 |
| Anthropic | "Powerful AI" | A model that outperforms Nobel-level experts across most fields, uses the interfaces a remote worker uses, works autonomously for hours to weeks, and runs as millions of copies at 10 to 100 times human speed. Summarised as a "country of geniuses in a datacenter". Its chief executive writes that he avoids "AGI" because of the baggage the term carries | Dario Amodei, "Machines of Loving Grace", October 2024 |
| Meta | "Personal superintelligence" | An AI that knows a person deeply, understands their goals and helps them reach them, delivered through personal devices. A description of a product direction, not a capability threshold | Mark Zuckerberg's letter, July 30, 2025 |
| Microsoft AI | "Humanist superintelligence" | AGI matches human performance across tasks. Superintelligence goes far beyond it. The humanist version is problem-oriented, domain-specific and limited in autonomy | Mustafa Suleyman's essay, November 6, 2025 |
| Safe Superintelligence | Superintelligence | Not defined on the company's site. In interview, its chief executive called AGI a loaded term, said a human being is not an AGI in the usual sense, and agreed with the description of superintelligence as a mind that can learn to do any job | Ilya Sutskever, interview of November 25, 2025 |
Where the Definitions Disagree
- Economic or cognitive. OpenAI's charter defines AGI by economically valuable work. The Hendrycks paper defines it by cognitive abilities, splits those into ten domains and scores systems against human tests. On that scale the paper scored GPT-4 at 27% and GPT-5 at 57%.
- A threshold or a scale. Most definitions describe a line to be crossed. The DeepMind paper describes levels: Competent is at least the 50th percentile of skilled adults, Expert the 90th, Exceptional the 99th, and Superhuman outperforms everyone.
- Knowing or learning. Most definitions list what the system can do. Sutskever's describes a system by its ability to learn, which he compared to a very capable 15-year-old who still has to learn the job.
- General or scoped. Bostrom's wording covers practically every field. Microsoft's humanist version is deliberately domain-specific, and its essay talks of "medical superintelligence".
- Capability or label. The US government's "Super Intelligence" is a naming decision. The executive order says the term covers the existing legal definition of AI and asks for a legislative definition within 60 days.
Which Term to Use
| If you are writing about | Use | Why |
|---|---|---|
| Systems that match skilled people at most cognitive work | AGI, with the definition you mean | The word alone does not say whether you mean economic work or cognitive breadth |
| Systems far beyond the best people in nearly every field | Superintelligence, or ASI | The two are used interchangeably outside the DeepMind scale |
| A specific company's claim | That company's own term | "Powerful AI", "personal superintelligence" and "humanist superintelligence" are not synonyms |
| US federal policy after September 29, 2026 | "Super Intelligence (SI)", with a note that it means AI | Otherwise readers will take it to mean superhuman systems |
Why It Matters in November 2026
The naming change has practical effects. The Justice Department's Super Intelligence Task Force and the U.S.-China Super Intelligence Dialogue both cover AI broadly, despite the name. A complaint filed on November 12, 2026 uses the research meaning and asks a court to stop the development of "AGI or superintelligence". And predictions about arrival dates, collected in our timeline tracker, cannot be compared without checking which definition each speaker has in mind. The organisations that use the word for their own goal are in the superintelligence lab tracker.
Sources
Definitions are from the OpenAI Charter, Altman's Reflections and The Gentle Singularity, the Levels of AGI paper, Hassabis's July 2026 essay, Amodei's Machines of Loving Grace, Zuckerberg's letter, Suleyman's essay, Sutskever's interview, A Definition of AGI, Bostrom's 1998 paper and Executive Order 14434. The OpenAI Charter page blocked our reader, so its wording is given as quoted in a November 2026 court filing that cites it. Read on November 13, 2026.