What is AGI? AGI vs ASI vs SI explained

What is AGI? AGI stands for artificial general intelligence: a hypothetical SI (AI) system that could do most of the thinking work people do, across many fields, at least as well as a skilled human. It does not exist yet, at least not under any definition researchers agree on, and that is the first thing to know about it. The second is that the word now sits next to two other acronyms that look alike: SI, the new official US name for today's AI, and ASI, artificial superintelligence. This guide explains where the term AGI comes from, the main ways it is defined, how it differs from ASI and SI, whether anyone has built it, and what researchers expect.

AGI meaning in plain words

Most SI tools you use are built for one kind of job or are judged on one kind of job: translating text, recognising faces, recommending videos, playing chess. AGI is the idea of a single system that would be general: able to handle a wide range of tasks it was not specifically built for, learn new ones, and do them reliably well.

Two dimensions matter in almost every definition:

• Breadth: how many different kinds of tasks the system can handle.
• Depth: how well it does them, compared with people.

A calculator is deep but narrow. A chatbot that can talk about anything but often gets facts wrong is broad but shallow on many tasks. AGI would need both breadth and depth at once. Most definitions also focus on cognitive work (reasoning, writing, planning, learning) and leave physical tasks aside, because robots lag far behind software.

Where the term comes from

The question behind AGI is old. In 1950 Alan Turing proposed his "imitation game", now called the Turing test, to ask whether machines can think. But the phrase "artificial general intelligence" is much younger.

According to a review of AGI definitions published by Google DeepMind researchers, the first known use of the term was in a 1997 article about military technologies by Mark Gubrud, who described "AI systems that rival or surpass the human brain in complexity and speed, that can acquire, manipulate and reason with general knowledge, and that are usable in essentially any phase of industrial or military operations where a human intelligence would otherwise be needed". The same review says the computer scientists Shane Legg and Ben Goertzel popularised the term among their peers in 2001, describing AGI as a machine able to do the cognitive tasks that people can typically do.

The word was needed because "AI" had become a name for any clever software, including narrow systems that solve one problem. AGI was a way to say: not that, but the original ambition of a general thinking machine.

Five ways people define AGI

There is no official definition, and the differences are not cosmetic. Here are the main families, as summarised and compared in the DeepMind paper Levels of AGI.

1. Human-level performance on cognitive tasks

The Legg and Goertzel framing: a machine that can do the non-physical tasks a typical person can do. It is intuitive, but it leaves open which tasks count and which people are the benchmark.

2. The ability to learn new tasks

In his 2015 book The Technological Singularity, Murray Shanahan defined AGI as "artificial intelligence that is not specialized to carry out specific tasks, but can learn to perform as broad a range of tasks as a human". The key word is learn: a system that only repeats what it was trained on would not qualify.

3. Economically valuable work

OpenAI's charter, written in 2018, defines AGI as "highly autonomous systems that outperform humans at most economically valuable work". It gives a measurable yardstick, economic value, but the DeepMind authors note that some parts of intelligence, such as artistic creativity or emotional intelligence, have no clear price, and that the definition implies a system must be deployed before it counts.

4. Flexible and reliable

The cognitive scientist Gary Marcus describes AGI as "shorthand for any intelligence (there might be many) that is flexible and general, with resourcefulness and reliability comparable to (or beyond) human intelligence". He proposed concrete tasks, such as understanding a novel or a film, or working as a cook in an unfamiliar kitchen, a version of the "coffee test" credited to Steve Wozniak. That last task requires a robot body, which most other definitions leave out.

5. Learning efficiency

The ARC Prize Foundation, built around the benchmarks created by the researcher François Chollet in 2019, rejects the economic definition as "an incorrect measure of intelligence", because skill can be bought with enough training data. Its own definition: "AGI is a system that can match the learning efficiency of humans." Its ARC-AGI puzzles are designed to be easy for people and hard for machines, so that a high score reflects learning rather than memorised knowledge.

Levels of AGI: a ladder instead of a finish line

To get out of the yes-or-no argument, the Google DeepMind team proposed in November 2023 a scale inspired by the levels of driving automation used for self-driving cars. It crosses performance (how good) with generality (narrow or general). The general column looks like this:

The authors placed the frontier chatbots of September 2023 at level 1, "Emerging AGI": competent at some tasks such as short essays and simple coding, but still at the emerging level on most tasks, including maths and factual accuracy. They wrote that "Competent AGI" is probably the best match for what most people mean by AGI. In the narrow column, by contrast, some systems already sit at the top: AlphaFold, AlphaZero and the chess engine Stockfish are classed as superhuman, but only at one job each. (Level 4 was first called "Virtuoso"; the latest version of the paper renames it "Exceptional".)

The useful lesson of this ladder is that progress is uneven. A model can write a decent cover letter in seconds and still invent a citation, which is what SI hallucination looks like in practice.

AGI vs ASI vs SI

Since 22 September 2026, the US government uses "super intelligence" (SI) as its name for artificial intelligence. That makes three terms easy to mix up. They do not mean the same thing:

The trap is the word "super". The SI label describes the tools we already have; it does not claim they are smarter than people. ASI, in the sense popularised by the Oxford philosopher Nick Bostrom, is a hypothetical intellect that greatly exceeds human performance in virtually all domains. AGI sits between the two: more than today's tools, less than superintelligence. In DeepMind's ladder, ASI is simply the top rung of the general column. Our guide What is SI? covers the renaming and the difference with superintelligence in more detail, and AI is now SI explains what the new name changes and what it does not.

Has AGI been achieved?

Not by any shared standard. But the claims have started.

In March 2023, a team at Microsoft Research published "Sparks of Artificial General Intelligence", a study of an early version of GPT-4. The authors argued that it "could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system", and also listed its limitations, including the possible need for "a new paradigm that moves beyond next-word prediction". The DeepMind paper mentions researchers who went further and argued that large language models already are AGIs because they are so general; the DeepMind authors disagreed, saying generality must be paired with reliable performance.

The ARC Prize Foundation takes the opposite view of its own benchmarks: "Today, AI can automate many valuable tasks, but only humans are capable of open-ended innovation and discovery."

AGI is also a contract term. When Microsoft and OpenAI signed a new agreement on 28 October 2025, Microsoft wrote that "once AGI is declared by OpenAI, that declaration will now be verified by an independent expert panel", and several rights and revenue terms between the two companies depend on that moment. The announcement does not say who sits on the panel or which test it will use. In other words, one of the most consequential definitions of AGI in business will be decided by a committee, not by a benchmark.

When could AGI arrive? What researchers say

Predictions vary from "within a couple of years" to "never", and the loudest come from company leaders with a stake in the answer. A more measured source is the large survey run by AI Impacts with researchers from several universities. In October 2023 it questioned 2,778 researchers who had published at top AI conferences. Two results stand out:

• Asked when "unaided machines can accomplish every task better and more cheaply than human workers", the aggregate forecast gave a 10% chance by 2027 and a 50% chance by 2047. The 50% date was 13 years earlier than in the same survey a year before (2060).
• Asked instead when all human occupations would be fully automatable, the same community gave a 50% chance only by 2116.

Those two questions describe nearly the same event, yet the answers are about 70 years apart. The authors say they do not know what accounts for the gap. For anyone reading an AGI headline, that is the honest summary: researchers expect big changes, but even they are unsure what the milestone looks like and when it comes.

Why the definition matters to you

• Headlines: "X has reached AGI" means little until you know which definition is used. Ask: human-level at what, measured how, and by whom?
• Safety: OpenAI's charter says that if "a value-aligned, safety-conscious project comes close to building AGI" first, OpenAI will "stop competing with and start assisting" it. Commitments like this only work if the finish line is clear.
• Money and power: as the Microsoft and OpenAI agreement shows, the word now decides who owns what.
• Everyday trust: today's SI systems are not AGI, and they still make confident mistakes. Treat their answers as drafts to check, not as the work of a general intelligence.

You do not need AGI to be fooled

A system does not have to be generally intelligent to produce a convincing fake photo, a fluent essay or a cloned voice. Narrow tools that are excellent at one job are enough. That is why checking content matters now, not after some future milestone. When an image or a piece of text makes you wonder, run it through is this SI, our free checker for images, text, video and music, and read the result as one clue among others. For text in particular, the SI text detector page explains what such tools can and cannot tell, and our article on the dead internet theory looks at how much of the web is already machine-made.