What is AGI? Artificial general intelligence explained

AGI, short for artificial general intelligence, is a hypothetical kind of AI that could learn and perform most of the intellectual tasks a person can, across many fields, rather than one task it was built for. Today's systems are impressive but uneven. What AGI means exactly, how to test for it, and when or whether it will arrive are all still argued. (In US taxes, AGI means adjusted gross income, a different thing.)

In short

• AGI names a goal, not a product: AI as broadly capable as people, able to pick up new tasks without being rebuilt for each one.
• There is no agreed definition or test. Labs and researchers define it by economic value, by levels of skill, or by how efficiently a system learns.
• Forecasts of when it could arrive range from a few years to many decades, and some doubt current methods will get there at all.

What does the "general" in AGI mean?

Most AI ever built is narrow. A chess engine plays chess, a spam filter sorts email, and a system that reads X-rays does nothing else. Each is trained for one job and fails outside it. The "general" in AGI means the opposite: one system that can move between tasks, learn a new one from a little experience, and transfer what it knows from one field to another, the way a person who learns to drive a car can soon drive a van.

Artificial intelligence as a field set out with that general aim in the 1950s. The narrow systems came because they worked. The term "artificial general intelligence" was first used by Mark Gubrud in a 1997 article on military technology, and Shane Legg and Ben Goertzel spread it among computer scientists in the early 2000s, to mark the difference between the original goal and the narrow tools the field had become good at, according to a 2023 paper by researchers at Google DeepMind.

How is AGI different from today's AI?

The large language models behind modern chatbots blur the old line. One model can write code, summarize a contract, translate, and explain physics. That is far more general than a chess engine. But such models still make basic errors a careful person would not, struggle with some tasks that need long chains of planning, and can't reliably keep learning from experience after training ends. Whether they are early AGI or something else is itself part of the argument.

	Narrow AI	General-purpose models (2020s)	AGI, as usually defined
Example	A chess engine, a spam filter	A chatbot built on a large language model	Not yet built
Range of tasks	One	Many, mostly through text, images and code	Most tasks people do with their minds
Learns new tasks	Only by retraining	Partly, from examples in the prompt	Quickly, from little experience
Reliability	High inside its task	Uneven: strong in places, brittle in others	At least as reliable as a skilled person

Why is there no agreed definition?

Because people mean different things by intelligence. The definitions in use measure different qualities, and a system could meet one and fail another. Four of the best known, three definitions and one proposed test, show the spread:

Who	Definition, in brief	What it measures
OpenAI's charter	"Highly autonomous systems that outperform humans at most economically valuable work"	Economic output
Google DeepMind, 2023	Five levels, from "emerging" (as good as an unskilled person) to "superhuman", judged by both skill and breadth	Performance across many tasks
François Chollet, 2019	Intelligence is how efficiently a system acquires new skills, not the skills it has	Learning, not knowing
Mustafa Suleyman, 2023	"Artificial capable intelligence", a step short of AGI, judged by a "modern Turing test": turn $100,000 into $1,000,000 over several months	One open-ended task in the real world

The DeepMind framework is a useful middle path. It sets "competent" AGI at the 50th percentile of skilled adults across a wide range of tasks, and "expert" at the 90th. In its 2023 version, the authors placed the chatbots of the day, ChatGPT, Bard and Llama 2 among them, at level 1: "emerging AGI." Chollet's view points the other way. He argues that a system trained on vast data can look skilled without being able to adapt, so he built the ARC puzzles to test how quickly a system solves kinds of problems it has not seen.

No single test settles whether a system is AGI. The Turing test, the oldest candidate, asks only whether a machine can pass for human in conversation, and the DeepMind authors call it insufficient for the job now that chatbots can pass some versions of it.

When could AGI arrive?

Nobody knows, and forecasts disagree by decades. Three kinds of view stand out, each from a named source.

Researchers, surveyed: in late 2023, Katja Grace and colleagues asked 2,778 researchers who had published at top AI venues. Taken together, their answers gave a 50% chance that unaided machines could outperform humans at every task by 2047, 13 years earlier than the same question had produced a year before. The same respondents put full automation of all human jobs much later, with a 50% chance by 2116. The gap between those two answers shows how much depends on the wording.

Lab leaders, writing: some heads of AI companies expect it sooner. In an October 2024 essay, Dario Amodei, the chief executive of Anthropic, wrote that "powerful AI", his preferred term, "could come as early as 2026, though there are also ways it could take much longer."

Skeptics, arguing: others doubt that scaling up today's methods leads there at all. Chollet's argument is one version: more data and computing buy more skill, but not the ability to learn new things efficiently, which he takes to be the heart of intelligence.

Still argued

Whether AGI is near, far, or even a coherent goal is contested. Optimists point to how quickly general-purpose models have improved. Skeptics point to how they fail and to the lack of a clear test. Surveys show wide spread among experts themselves, and forecasts have moved by more than a decade from one year to the next. This note sets out the positions; it doesn't predict a date.

Why does AGI matter?

Because a system that could do most mental work would change work, science and power at once. Its supporters picture faster medical research and abundant cheap expertise. Its critics and many of its builders worry about mass job loss, misuse, and systems more capable than the people overseeing them, the debate laid out in whether AI is an existential risk. Some go further, to the idea that AI which improves itself could set off a technological singularity. These hopes and fears shape billions in investment and much of the debate over how to regulate AI, whether or not AGI is ever built.

Check yourself

• What makes an AI system "general" rather than narrow, in your own words?
• Why can a chatbot be very capable and still not count as AGI under some definitions?
• Why do forecasts of AGI's arrival differ so much, even among researchers?

Sources

• Levels of AGI for Operationalizing Progress on the Path to AGI — Morris et al., Google DeepMind, 2023
• Thousands of AI Authors on the Future of AI — Grace et al., 2024
• On the Measure of Intelligence — François Chollet, 2019
• Machines of Loving Grace — Dario Amodei, 2024