Menu
AGIConcepts

Artificial general intelligence (AGI)

4 min readbeginnerUpdated 28 Sept 2026
1 · In one line

AGI is a proposed category for AI at least as capable as a person across most thinking tasks. Definitions differ; a 2023 Google DeepMind framework marked the level matching many definitions as not yet achieved.

1 · What it is

Artificial general intelligence, or AGI, is the idea of an AI system that can do most kinds of mental work at least as well as a skilled person. The key word is general. In its 2023 placements, the Levels of AGI paper classed systems including AlphaFold and AlphaZero as superhuman narrow AI, while placing frontier language models at the first general level. AGI would be one system that is broadly capable.

Researchers and labs use different definitions. OpenAI’s charter ties AGI to outperforming people at most economically valuable work, while a 2023 OpenAI post describes it more loosely as AI that is generally smarter than humans. Google DeepMind describes it as matching people at most cognitive tasks. The Levels paper notes that some researchers already regarded the large language models available when it was written as AGI. Because of this, a claim that AGI is close, or already here, only means something once you know which definition is being used.

A useful way through the confusion comes from a 2023 Google DeepMind paper, later revised. It rates systems on two separate scales: how broad their skills are, and how good they are compared with skilled adults. On that grid, specialist systems already reach the top level, while general systems were placed only at the first step, and the level the paper says fits many definitions of AGI is marked as not yet achieved.

The labs that aim to build AGI also publish how they think about its risks. Google DeepMind groups them into four areas: misuse, misalignment, mistakes and structural risks. That is one reason the definition matters: a shared, testable definition shapes how risks are assessed and compared.

2 · Why it exists

The word gets used for very different things.

Narrow strength is not generalityIn its 2023 placements, the Levels paper classed systems such as AlphaGo and AlphaFold as highly capable narrow AI, while placing frontier language models at the first general level.
No shared definitionLabs and researchers draw the line in different places, so the same system can be called AGI by one person and not by another.
Hard to measure progressWithout an agreed yardstick, claims that AGI is near or already here are hard to compare or check.
3 · How it works

One way to judge a system, in five questions.

ONE WAY TO TEST A CLAIM THAT A SYSTEM IS AGI An AI system the claim to check 2 · Many kinds of task writing, maths, coding, learning a new skill … no robot body needed 3 · Score each task against skilled adults, as a percentile 4 · Apply the bar better than half of skilled adults on most tasks? no yes Not yet AGI narrow, or only “emerging” AGI by this bar: “competent” 1 · Left out on purpose whether it thinks or feels like a person, and whether it is conscious 5 · Judged apart how much it acts on its own Based on the Levels of AGI framework (Google DeepMind). Other labs set the bar elsewhere: OpenAI’s charter ties AGI to outperforming people at most economically valuable work.
One way to test an AGI claim: breadth first, then skill against people, then a clear bar.
  1. 1 · capabilityJudge what the system can do, not whether it thinks or feels like a person.
  2. 2 · breadthAsk how wide a range of non-physical tasks it can handle, including learning new skills.
  3. 3 · depthFor those tasks, compare its results with skilled adults, from roughly unskilled level up to beating everyone.
  4. 4 · barThe paper's Competent bar: better than about half of skilled adults on most tasks. It says many existing definitions of AGI fit this level.
  5. 5 · autonomyTreat how much independence the system is given as a separate choice from how capable it is.

In the Levels paper's 2023 placement, narrow systems reached the top level, while general systems had not demonstrated breadth at a skilled-human level.

LEVEL NARROW GENERAL 5 SUPERHUMAN beats all humans AlphaFold, AlphaZero Superintelligence not yet achieved 4 EXCEPTIONAL top 1% Deep Blue, AlphaGo Exceptional AGI not yet achieved 3 EXPERT top 10% Grammar checkers Expert AGI not yet achieved 2 COMPETENT top half Smart speakers Competent AGI catch-all bar · not yet achieved 1 EMERGING ≈ unskilled person Rule-based systems Emerging AGI chatbots, as placed in 2023 exists not yet achieved Ranks compare with skilled adults. Placements are the Levels of AGI paper’s own.
Where the Levels of AGI paper placed real systems in 2023: narrow AI reached the top level, while frontier language models were placed at Emerging AGI.
4 · Where it's used
WhoWhat they askWhat it works with
AI lab“What exactly are we trying to build, and how will we know?”Its own published definition and mission
Research team“How does our new model compare with earlier ones?”A graded scale of breadth and skill
Policymaker“How do the risks of different models compare?”Shared, testable definitions of capability levels
Safety team“Which risks grow as systems get more capable?”Risk areas such as misuse and misalignment
5 · What it solves, and what it doesn't
solves
  • Gives a name to the long-standing goal of human-level, general-purpose AI.
  • A levelled scale lets people say how far along a system is instead of arguing yes or no.
  • Separating skill, breadth and independence makes it easier to discuss risk.
doesn't solve
  • AGI names a proposed capability category, not a particular product or technique.
  • Definitions still differ between organisations, so claims need to say which one they use.
  • It does not settle whether a machine can understand or be conscious; the Levels framework deliberately leaves that aside.
  • The Levels paper says unambiguous placement would require a standardized benchmark of tasks.
6 · Go deeper
Narrow AISuperintelligenceTuring testArtificial intelligenceLLM

Sources used

This explainer is written in original language. The links below support its factual claims.

  1. paperPosition: Levels of AGI for Operationalizing Progress on the Path to AGI, Google DeepMind (Morris et al., arXiv, revised 2025) · read 27 Sept 2026
  2. officialOpenAI Charter, OpenAI · read 27 Sept 2026
  3. officialTaking a responsible path to AGI, Google DeepMind · read 27 Sept 2026
  4. officialPlanning for AGI and beyond, OpenAI · read 27 Sept 2026
  5. paperAn Approach to Technical AGI Safety and Security, Google DeepMind (arXiv, 2025) · read 27 Sept 2026
  6. paperUniversal Intelligence: A Definition of Machine Intelligence, Legg and Hutter (Minds and Machines, 2007) · read 27 Sept 2026
Suggest a correction