AGI vs superintelligence: what is the difference?

Artificial general intelligence (AGI) means roughly human-level ability across many tasks. Superintelligence (ASI) means performance far beyond the best humans across virtually all tasks. Narrow AI does one kind of task, sometimes better than any human. There is no agreed definition of AGI, and as of October 2026 no system has been shown to meet the standard definition of superintelligence.

What is the difference between narrow AI, AGI and superintelligence?

The three terms describe two separate things at once: how broad a system is, and how good it is.

• Narrow AI performs well on one task or a small set of tasks. A chess engine, a protein-structure predictor and a spam filter are narrow. Some narrow systems are already superhuman at their one task.
• Artificial general intelligence performs well across a wide range of tasks, at roughly the level of a capable human. Definitions differ on how wide and how capable.
• Artificial superintelligence (ASI) performs far beyond the best humans across virtually all tasks. This is the sense in Nick Bostrom’s 2014 definition, covered in the definition guide.

The common mistake is to treat these as three points on one line. They are not.

A system can be superhuman and still narrow. AlphaFold predicts protein structures better than any human. It cannot write a contract. Superintelligence requires both: superhuman depth and general breadth.

Side by side, the three terms differ on breadth, on level of performance and on whether anything meets them today.

Narrow AI vs AGI vs superintelligence (ASI)

Aspect	Narrow AI	AGI	Superintelligence (ASI)
What it means	Performs well on one task or a small set of tasks	Performs well across a wide range of tasks, at roughly the level of a capable human	Performs far beyond the best humans across virtually all tasks
Breadth	One task or a small set	Wide; definitions differ on how wide	Virtually all tasks
Level	Can be superhuman at its one task	Roughly a capable human; definitions differ on how capable	Far beyond the best humans
Example on this page	A chess engine, AlphaFold, a spam filter	The Levels of AGI paper classes chatbots such as ChatGPT and Gemini as “Emerging AGI”	None; the Levels of AGI paper lists it as not yet achieved
Exists today?	Yes; some narrow systems are already superhuman at their one task	Depends on whose definition is used	No system has been shown to meet the standard definition as of October 2026
Who defines it	The Levels of AGI paper applies its levels to narrow systems as well as general ones	No agreed definition: OpenAI’s Charter (economic work), Google DeepMind’s Levels of AGI (percentiles); Anthropic’s chief executive prefers “powerful AI”	Nick Bostrom (2014); the Levels of AGI paper labels Level 5 general performance “Artificial Superintelligence (ASI)”

Where did the term AGI come from?

The term dates to the late 1990s and caught on in the late 2000s.

Mark Gubrud used “artificial general intelligence” in a 1997 conference paper, “Nanotechnology and International Security.” According to MIT Technology Review, Shane Legg, later a co-founder of DeepMind, suggested the term to Ben Goertzel as the title of a 2007 book of essays, and Goertzel launched an annual AGI conference in 2008. The term spread from there.

The point of the word was contrast. Mainstream AI research at the time was narrow. “General” marked the thing it did not yet do.

How do the major definitions of AGI differ?

The major definitions of AGI measure different things. OpenAI’s measures economic work, Google DeepMind’s measures percentile performance across tasks, and Anthropic’s describes capability combined with autonomy and scale.

OpenAI: an economic test

OpenAI’s Charter, published in 2018, defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.”

This definition is about jobs, not about minds. It asks whether a system can do paid work better than people, and it requires autonomy. It does not mention reasoning, understanding or consciousness.

Google DeepMind: a ladder of levels

In November 2023, a team led by Meredith Ringel Morris at Google DeepMind published “Levels of AGI for Operationalizing Progress on the Path to AGI.” Instead of one finish line, the paper proposes levels of performance crossed with breadth.

1. Emerging: equal to or somewhat better than an unskilled human.
2. Competent: at least the 50th percentile of skilled adults.
3. Expert: at least the 90th percentile of skilled adults.
4. Exceptional: at least the 99th percentile of skilled adults. Earlier versions called this level “Virtuoso”; the September 2025 revision renamed it.
5. Superhuman: outperforms 100% of humans.

The paper applies each level to both narrow and general systems. It classes chatbots such as ChatGPT and Gemini as “Emerging AGI.” It labels Level 5 general performance “Artificial Superintelligence (ASI)” and lists it as not yet achieved.

The paper also sets six principles for any AGI definition. One of them is “Focus on Potential, not Deployment”: a system counts by what it can do, not by whether anyone has put it to work yet.

Google DeepMind again: a short working definition

In an April 2025 post introducing its paper on AGI safety, Google DeepMind used a simpler line: AGI is AI “at least as capable as humans at most cognitive tasks.” The post said such systems could arrive “within the coming years.”

Anthropic: “powerful AI” instead

Anthropic’s chief executive, Dario Amodei, wrote in his October 2024 essay “Machines of Loving Grace” that he dislikes “AGI” as an imprecise term with science-fiction baggage. He uses “powerful AI” instead: a model smarter than a Nobel Prize winner across most relevant fields, able to work autonomously on tasks lasting hours or weeks, and able to run as many copies at once. His shorthand is a “country of geniuses in a datacenter.”

That description compares the system with Nobel Prize winners rather than with workers, so it sets a higher bar than the economic AGI of OpenAI’s Charter. Amodei uses neither “AGI” nor “superintelligence” for it.

Researchers’ surveys: high-level machine intelligence

The AI Impacts surveys of AI researchers avoid “AGI” and ask about high-level machine intelligence (HLMI). The 2023 survey defined it as the point when “unaided machines can accomplish every task better and more cheaply than human workers.”

“Every task” is a very high bar. By that wording, HLMI is close to what many people would call superintelligence, although doing every task better than human workers is not the same as greatly exceeding the best humans.

Philanthropy and policy: transformative AI

Open Philanthropy (now Coefficient Giving), a major funder of AI safety research, uses transformative AI, a term it introduced in 2016 posts by Holden Karnofsky. It defines AI by its effect: a transition comparable to, or more significant than, the agricultural or industrial revolution. A system could be transformative without being general, or general without being transformative.

Why are the lines so fuzzy?

The lines are fuzzy because performance is uneven, because the definitions measure different things, and because some definitions carry money. Each changes how to read a claim that AGI or superintelligence has arrived.

Performance is uneven

The Levels of AGI paper shows the problem in its own table. It rates state-of-the-art language models as Competent on a subset of tasks, yet classes them only as Emerging when judged as general systems. The paper asks for performance across a broad set of tasks, not peak performance on a few. A single impressive benchmark says little about generality.

Definitions measure different things

A lab that grades paid work and a lab that grades test percentiles are not measuring the same thing. The same system could count as AGI under one of the definitions above and fall short under another.

Definitions carry money

Some AGI definitions are written into contracts. Fortune reported in November 2025 that Microsoft’s earlier agreement with OpenAI had barred Microsoft from pursuing its own AGI research, and that Microsoft formed its superintelligence team after the deal was restructured. When a company declares or denies that AGI has arrived, the declaration can have commercial consequences. That does not make the claim false. It does mean the claim needs its definition attached.

Is superintelligence just “more AGI”?

Not exactly. Superintelligence is usually treated as a step beyond AGI, and many researchers expect it to follow. How quickly is disputed, and some reject the step altogether.

The philosopher David Chalmers framed the step in a 2010 paper, “The Singularity: A Philosophical Analysis.” He used AI for human-level systems, AI+ for greater than human level, and AI++ for far greater than human level. His argument: if AI arrives, AI+ will soon follow, and AI++ soon after that, absent defeaters such as a global catastrophe.

Surveys of AI researchers show the expectation and the disagreement together.

• In a 2012–2013 survey by Vincent Müller and Nick Bostrom, the median respondent expected systems to move from high-level machine intelligence to superintelligence in less than 30 years.
• In the 2023 AI Impacts survey, the median respondent gave a 10% chance that superintelligence would follow within two years of HLMI, and a 60% chance within thirty years.

Others reject the step entirely. Arvind Narayanan and Sayash Kapoor, computer scientists at Princeton, argue in “AI as Normal Technology” (April 2025) that superintelligence is incoherent as usually conceptualized. The arguments are in the intelligence explosion guide and the debate guide.

Which term should you use?

Use the one that matches what you can actually show.

• Use narrow AI for a system built for one kind of task, however good it is.
• Use AGI only with a definition attached, because readers will otherwise fill in their own.
• Use superintelligence for the Bostrom sense: far beyond the best humans in virtually all domains.
• When quoting a company, use the company’s own term and say whose definition it is.

The company definitions guide quotes each lab’s own wording.