Article · 4 min read

What is AGI, really?

Several of the world’s biggest tech companies say they are building it. Yet there is no definition everyone agrees on. Here is the idea behind one of the great races of our time, explained from the ground up.

Short answer

AGI, artificial general intelligence, is an AI system that can learn and do roughly any intellectual work a person can. There is no shared definition, no test that settles it and no agreement on when, or whether, it will happen. AGI is not the same as superintelligence, but it may be the step before.

From narrow to general

Almost all AI in history has been narrow. Deep Blue beat the world chess champion in 1997 but could not play noughts and crosses. AlphaGo defeated one of the world’s best Go players in 2016 but could not write a sentence. Each system was built for a single task.

General intelligence is something else: the ability to take on new tasks nobody prepared you for, and to learn them. Humans do it all the time. The same brain that learns to ride a bike also learns Spanish, cooking and tax returns.

Today’s large language models are the most general systems ever built. The same model can write, translate, program and solve maths problems, and in 2025 AI systems performed at gold-medal level at the International Mathematical Olympiad. Yet the same models can make mistakes few people would make. In a 2023 study in which 758 consultants at Boston Consulting Group worked with GPT-4, the researchers described a jagged frontier: tasks that look equally hard can fall on opposite sides of the line of what AI can do.

Why nobody can agree on a definition

The trouble starts with intelligence itself, which is hard to define. So different people measure different things.

  • Capability. The most common idea: an AGI system handles most thinking tasks at least as well as a person.
  • Economic value. OpenAI’s charter from 2018 defines AGI as “highly autonomous systems that outperform humans at most economically valuable work”.
  • Learning. The AI researcher François Chollet argues that intelligence is not about what a system can already do but about how efficiently it learns something new. His ARC test consists of puzzles in which every task is new, so the answers cannot be learned by heart.
  • Levels. In 2023, researchers at Google DeepMind proposed a five-step scale, from “emerging” to “superhuman”, that measures both how well and how broadly a system performs. Chatbots such as ChatGPT landed on the lowest step. Narrow systems such as AlphaFold already counted as superhuman, but only in their own field.

All four views capture something important, and sometimes they point in different directions. A system can be economically valuable without learning particularly efficiently, and the other way round.

How close are we?

Nobody knows, and the answer depends on the definition you choose. Three yardsticks still give a picture.

  • In one of the largest surveys of its kind, with 2,778 AI researchers in 2023, the combined estimate gave a 50% chance that unaided machines will outperform humans at every task by 2047.
  • The research organisation METR measures the length of tasks AI agents can complete on their own, measured by how long the same task takes a person. By measurements published in 2025, that length has doubled roughly every seven months since 2019.
  • Several heads of AI companies have said it could happen within a few years. Many researchers are more cautious, and some doubt that today’s methods will get all the way there.

What is clear is that the question is no longer just a thought experiment. It decides where companies and governments put their money.

AGI is not superintelligence

AGI means roughly human level. Superintelligence means an intellect that vastly outperforms the best humans in almost every field. Yet the step between the two may turn out to be short.

The mathematician I. J. Good explained why as early as 1965: a machine that can improve machines can improve itself. An AI system as good as the best AI researchers could also work around the clock, in thousands of copies at once. Whether that leads to a slow transition over decades or to an intelligence explosion in days is one of the big open questions. Try the dynamics yourself in our simulator.

Why the definition matters

It sounds like an academic question, but it has had very concrete consequences.

  • Money. In the agreement between Microsoft and OpenAI, AGI was long a legal line: the day OpenAI reached AGI, Microsoft’s rights to the new technology would end. According to The Information, the agreement defined AGI as a system able to generate around 100 billion dollars in profit. In 2025 it was decided that an independent panel of experts would judge the matter, and in April 2026 the agreement was rewritten to run to fixed dates instead.
  • Rules. The EU AI Act does not talk about AGI but about general-purpose AI models. The most powerful face special requirements: a model trained with more than 1025 floating-point operations is presumed to carry systemic risk.
  • Safety. The closer systems come to general ability, the more alignment matters: that they want what we actually want. It is easier to prepare for something you will recognise when it arrives.

Sources

  1. OpenAI, OpenAI Charter, 2018.
  2. Morris et al., Levels of AGI for Operationalizing Progress on the Path to AGI, Google DeepMind, 2023.
  3. François Chollet, On the Measure of Intelligence, 2019.
  4. Grace et al., Thousands of AI Authors on the Future of AI, 2024.
  5. METR, Measuring AI Ability to Complete Long Software Tasks, 2025.
  6. Dell’Acqua et al., Navigating the Jagged Technological Frontier, Harvard Business School, 2023.
  7. Microsoft, The next phase of the Microsoft-OpenAI partnership, 27 April 2026. The 100 billion dollar figure: The Information, December 2024.
  8. European Union, Regulation (EU) 2024/1689 on artificial intelligence, Article 51.
  9. I. J. Good, Speculations Concerning the First Ultraintelligent Machine, 1965.

Terms in this article

How we work: the articles are written with the help of AI, and every fact is checked against the sources listed above. Found a mistake? Write to hej@sibolaget.se.

More questions?

New to AI? Start with What is AI? The glossary explains the terms, and the front page has straight answers to the most common questions. Missing something? Write to us.