
AGI (artificial general intelligence) is AI that can understand, learn and carry out most intellectual tasks as well as a capable human, rather than excelling at just one. A chess engine is narrow AI. An AGI could learn chess, then write a legal brief, debug software and plan a research project, adapting the way a person does.
That’s the simple version. The hard part is that nobody agrees exactly where the line is, which is why some tech leaders say AGI is already here while most researchers say it isn’t. This guide explains the main definitions, how AGI is measured, whether it exists in 2026, and when it might arrive.
Looking for AGI in taxes? That’s adjusted gross income, an unrelated term. This page is about artificial general intelligence.
AGI definition: what the major labs mean
AGI has no official definition. The companies building it each use their own, and the differences matter:
| Who | Definition |
|---|---|
| OpenAI (Charter) | “Highly autonomous systems that outperform humans at most economically valuable work” |
| Shane Legg, Google DeepMind co-founder | “An artificial agent that can do the kinds of cognitive things that people can typically do” |
| Demis Hassabis, Google DeepMind CEO | “A system that’s able to exhibit all the complicated capabilities that humans can” |
| Dario Amodei, Anthropic CEO | Prefers “powerful AI”: smarter than a Nobel Prize winner across most fields, a “country of geniuses in a datacenter” |
| Jensen Huang, Nvidia CEO | Informal test (2026): could an AI start and grow a company worth $1 billion? |
OpenAI’s version is economic (can it do the work?), DeepMind’s is cognitive (can it think like people?), and Amodei’s sets the bar much higher than “average human.” Which definition someone uses largely decides whether they think AGI has arrived. Even Sam Altman has called AGI “a very sloppy term.”
Narrow AI vs AGI vs ASI
AGI sits in the middle of the three standard levels of AI capability:
| Narrow AI (ANI) | General AI (AGI) | Superintelligence (ASI) | |
|---|---|---|---|
| Scope | One task or a narrow set | Most cognitive tasks | Virtually every domain |
| Level | Can beat humans at its one task | Roughly human level overall | Far beyond the best humans |
| Exists? | Yes, everywhere | Disputed | No |
| Examples | Spam filters, chess engines, recommendation feeds | No agreed example | None (hypothetical) |
The step after AGI is artificial superintelligence. We cover that jump in AGI vs ASI and explain how AI, AGI and SI relate in super intelligence vs AI.
The levels of AGI
Instead of a yes/no line, Google DeepMind researchers proposed levels of AGI in 2023, rating systems on how well they perform and how general they are:
| Level | Performance (compared with skilled adults) |
|---|---|
| Level 1: Emerging | Equal to or somewhat better than an unskilled human |
| Level 2: Competent | At least 50th percentile of skilled adults |
| Level 3: Expert | At least 90th percentile |
| Level 4: Virtuoso | At least 99th percentile |
| Level 5: Superhuman | Better than all humans |
When the paper was published, the authors placed chatbots like ChatGPT at Level 1, “Emerging AGI,” and said no general system had reached the higher levels. The framework is useful because it replaces “is it AGI?” with “how good, and at how many things?”
Has AGI been achieved in 2026?
This is the most argued-about question in AI right now.
The case that it has (or nearly has):
- In March 2026, Nvidia CEO Jensen Huang told Lex Fridman that AGI “has already been achieved,” using his billion-dollar-company test.
- Frontier models have saturated many benchmarks built to resist them. The ARC Prize’s results page shows OpenAI’s GPT-6 scoring 95% on ARC-AGI-2 (September 2026), a puzzle test designed to measure flexible reasoning rather than memorised knowledge.
- AI agents now carry out long, multi-step tasks in coding and research with little human help.
The case that it hasn’t:
- Today’s AI is “jagged”: superhuman at some tasks and well below humans at others. A Center for AI Safety team scored GPT-5 at about 57% on its own AGI scale.
- Models still make basic errors a competent person wouldn’t, and struggle to reliably complete open-ended jobs from start to finish.
- Skeptics such as Yann LeCun argue current methods lack key abilities like true world models, and won’t reach human-level intelligence through scaling alone.
The fair summary: AI in 2026 is more general than anything before it, but there’s no consensus that it meets the bar of AGI. If you use OpenAI’s economic definition strictly, it doesn’t yet do most economically valuable work. If you use a looser definition, a case can be made.
Is ChatGPT AGI?
Not by most definitions. ChatGPT is general-purpose (it writes, codes, translates and reasons across topics), which is why it’s often called a step toward AGI. But it still can’t reliably do most jobs end to end without human supervision, and it fails in ways a capable person wouldn’t. In DeepMind’s framework it was rated “Emerging AGI”: general, but not yet at skilled-human level across the board.
AGI examples: what would count?
Because there’s no agreed AGI yet, examples are tests rather than products. Researchers have proposed many:
- The coffee test (attributed to Steve Wozniak): could a robot walk into an unfamiliar home and make a cup of coffee?
- The employment test: could an AI do most jobs as well as the people who do them?
- The ARC benchmarks: can AI solve novel puzzles it has never seen, learning the rule from a few examples?
- The Turing test: can it pass as human in conversation? Modern chatbots arguably pass, which is why most researchers no longer treat it as a meaningful AGI test.
Today’s systems pass some of these and fail others, which is exactly the “jagged” picture.
When will AGI arrive?
Forecasts have moved much earlier in the past few years:
| Source | Forecast |
|---|---|
| Sam Altman, OpenAI (Jan 2025) | “We are now confident we know how to build AGI as we have traditionally understood it” |
| Dario Amodei, Anthropic (Oct 2024) | Powerful AI “could come as early as 2026,” though it could take much longer |
| Demis Hassabis, Google DeepMind (2025) | AGI in “five to 10 years” |
| 2,778 AI researchers (survey 2023) | 50% chance machines outperform humans at every task by 2047; 10% by 2027 |
Lab leaders tend to forecast earlier than academic researchers, and they also have commercial reasons to be optimistic. Our superintelligence timeline tracks these predictions as they change.
Why AGI matters
AGI would be the biggest change in technology since the computer, for three reasons:
- Economic impact. A system that can do most cognitive work could transform jobs, productivity and science.
- Acceleration. AGI could speed up AI research itself. That’s why many experts think superintelligence could follow AGI quickly, a jump often called the technological singularity.
- Risk. A general system that acts on its own is harder to oversee than a narrow tool. Making AGI’s goals match human intent, known as alignment, becomes far more important at this level. See will AI take over the world? for what experts estimate.
Our take
AGI means AI that matches people across most thinking work, and in 2026 we’re in the grey zone. Today’s models are general and often impressive, but uneven. The real argument is no longer whether AGI is possible. It’s where to draw the line, and how fast systems are moving past it.
For what comes after AGI, read our guides to what super intelligence is and the difference between AGI and ASI.
Enjoyed this? Get the SI Briefing
One short email a week: what changed on the road to super intelligence, and why it matters. No spam, unsubscribe anytime.
Frequently asked questions
What is AGI in simple terms?
AGI, or artificial general intelligence, is AI that can learn and do most intellectual tasks as well as a capable human, instead of being good at only one thing. It could switch between writing, coding, research and planning the way a person can.
What does AGI stand for?
In AI, AGI stands for artificial general intelligence. In US taxes, AGI means adjusted gross income, which is unrelated.
Is ChatGPT AGI?
Not by most definitions. ChatGPT and other frontier models are general-purpose and very capable, but they still make basic mistakes, struggle with long autonomous tasks and can't reliably do most jobs end to end. Some leaders argue a form of AGI is already here; most researchers disagree.
Does AGI exist yet?
It is disputed. Nvidia's Jensen Huang said in March 2026 that AGI has been achieved, using his own test of whether AI could build a billion-dollar company. Most researchers say current AI is still jagged: superhuman at some tasks and well below humans at others.
What is the difference between AI and AGI?
Most AI today is narrow: built or trained for specific tasks such as recommendations, translation or image recognition. AGI would be general: able to learn and perform a wide range of tasks at human level without being rebuilt for each one.
When will AGI be achieved?
Forecasts vary. Sam Altman and Dario Amodei have pointed to the late 2020s, Demis Hassabis said five to ten years in 2025, and a 2023 survey of 2,778 AI researchers gave a 50% chance that machines could outperform humans at every task by 2047.
11 Sources
- OpenAI Charter (definition of AGI)
- Morris et al., Google DeepMind, “Levels of AGI” (2023)
- MIT Technology Review, Google DeepMind's levels of AGI (2023)
- Sam Altman, “Reflections” (January 2025)
- Dario Amodei, “Machines of Loving Grace” (October 2024)
- NBC News, Hassabis: AI matching humans at any task in five to 10 years (2025)
- Fortune, Jensen Huang says ‘we’ve achieved AGI’ (30 Mar 2026)
- ARC Prize, ARC-AGI results
- Grace et al., “Thousands of AI Authors on the Future of AI” (2024)
- Yann LeCun (Wikipedia)
- Artificial general intelligence (Wikipedia)