The short answer
AGI, or artificial general intelligence, is a hypothetical AI system that can match or exceed people across most kinds of work or thinking, rather than at one task. There is no agreed definition. OpenAI's charter describes highly autonomous systems that outperform humans at most economically valuable work, Google DeepMind uses AI at least as capable as humans at most cognitive tasks, and ARC Prize defines it as matching human learning efficiency. No lab has formally declared AGI and no independent body has confirmed it.
| Stands for | Artificial general intelligence |
|---|---|
| Agreed definition | None; each lab uses its own |
| OpenAI's charter | Highly autonomous systems that outperform humans at most economically valuable work |
| Google DeepMind | AI at least as capable as humans at most cognitive tasks |
| ARC Prize | A system that can match the learning efficiency of humans |
| Has it been reached? | Not by any agreed test as of November 2026 |
What AGI means, in plain English
Today's AI tools are powerful but uneven. The same assistant that writes a strong contract summary can miscount the days in a month. The idea behind AGI is a system without those gaps: one that can learn and do most things a capable person can, across subjects, without being built for each job.
The trouble is that "most things" and "capable person" mean different things to different people. TechCrunch's AI glossary, updated in September 2026, calls AGI a nebulous term and notes that even leading AI researchers are unsure what it means. That is not a side detail. It is the reason AGI predictions range from "almost here" to "decades away": people are often predicting different things.
How the main players define AGI
| Who | Definition | What it measures |
|---|---|---|
| OpenAI (charter) | Highly autonomous systems that outperform humans at most economically valuable work | Economic output |
| Sam Altman, OpenAI CEO (as quoted by TechCrunch) | The equivalent of a median human you could hire as a co-worker | Usefulness as a colleague |
| Google DeepMind (as quoted by TechCrunch) | AI at least as capable as humans at most cognitive tasks | Breadth of thinking ability |
| ARC Prize Foundation | A system that can match the learning efficiency of humans | How quickly it learns something new |
The definitions pull in different directions. An economic definition could be met by a system that is very good at office work but poor at learning anything new. ARC Prize focuses on learning instead: its benchmarks are built on the principle "Easy for Humans, Hard for AI", with puzzles most people solve quickly that AI systems still find difficult. Its newest benchmark, ARC-AGI-3, runs alongside an ARC Prize 2026 competition.
What leaders said about timing in 2026
At the World Economic Forum in Davos in January 2026, two of the most influential lab leaders disagreed in public, as reported by diginomica:
- Demis Hassabis, CEO of Google DeepMind, gave a 50% chance of a system matching all human cognitive capabilities by the end of the decade.
- Dario Amodei, CEO of Anthropic, broadly stood by his earlier, faster prediction, saying he did not think it was going to turn out to be that far off.
Two cautions apply. First, these are forecasts by people who run companies that benefit from the attention, so treat them as informed opinion rather than evidence. Second, they are using their own definitions, so "AGI by 2030" from one person is not the same claim as from another. Outside the labs, ARC Prize takes a more cautious line: its site says AI can automate many valuable tasks, but that open-ended innovation and discovery are still shown only by humans. The absence of an agreed test is exactly why the argument continues.
The argument flared again in September 2026. When OpenAI launched GPT-6 Astra, its president Greg Brockman told reporters he thought it was not unreasonable to feel that we are now in the AGI era, and Nvidia's chief executive Jensen Huang posted that AGI had arrived, according to Information Age, the news site of the Australian Computer Society. Neither was a formal declaration. The ARC Prize Foundation, which tested Astra, said it saw meaningful progress but was not claiming that it is AGI, and academics quoted in the same report disagreed with the AGI label.
Why it matters at work
For almost everyone, the practical answer is: plan for the AI you have, not the AI that is promised. Three reasons:
- Vendor claims. "Near-AGI" in a sales deck is marketing. Ask what the product does on your tasks, with your data, and how often it is wrong. Our guide on checking if AI is telling the truth helps.
- Skills. Waiting for AGI before learning AI tools is the costliest option, because the tools you can use today already change how work gets done. Our free lessons start there.
- Risk and policy. Debates about AGI drive regulation and company policy. Knowing that the term has no fixed meaning helps you read headlines and internal strategy papers critically. See AI safety and the EU AI Act.
An example you can try in five minutes
- Go to arcprize.org/play and try the daily puzzle. It shows a few example grids that change according to a hidden rule, and you work out the rule.
- Note how long it took you. Most people find these quick.
- Describe the same puzzle in words to your usual assistant, or upload a screenshot, and ask it to solve it. Compare how it does.
- Then ask the assistant the first prompt below. A good answer will give several definitions with their sources and admit the disagreement. A poor one will state a single definition as fact.
You will come away with a feel for the gap ARC Prize is measuring, and a quick test of whether your assistant handles contested topics honestly.
Common misconceptions about AGI
- "AGI is a switch that flips on one day." Capabilities arrive unevenly, task by task. Different definitions will be met at different times, if at all.
- "There is an official AGI test." There is not. Benchmarks such as ARC-AGI measure specific abilities, and labs choose their own measures.
- "AGI means the AI is conscious or has feelings." None of the definitions above mention consciousness. They are about what a system can do.
- "Today's chatbots are basically AGI already." They still hallucinate, struggle with some puzzles people find easy and need checking. Whether they count as AGI depends entirely on which definition you pick, which is the point.
- "AGI is science fiction and not worth thinking about." OpenAI's charter is built around it and Google DeepMind and Anthropic leaders discuss it in public. You do not have to accept anyone's timeline to take the question seriously.
Related terms
- Generative AI: today's AI, which creates text, images and code.
- Large language model: the technology most AGI claims are about.
- Reasoning model: the step labs point to as progress towards general ability.
- AI agent: AI that acts on its own, part of most AGI visions.
- Benchmark: how claims of progress are measured.
- AI safety: the field focused on keeping powerful AI under control.
To see how today's assistants actually compare on real work, start with ChatGPT vs Claude and Claude vs Gemini.
Prompts to copy
Test how your assistant handles a contested term
Define AGI. Give at least four definitions used by different organisations or researchers, name who uses each and where it was stated, and explain how they disagree. Then tell me which claims about AGI timelines are forecasts rather than evidence. If you are unsure of a source, say so instead of guessing.
You get: A balanced overview and a check on your assistant's honesty
Cut through an AGI claim in a vendor pitch
A vendor told us their product is [quote the claim, for example close to AGI for customer service]. Write five questions I should ask them that turn this into testable claims about our tasks, error rates and the human review needed. Then suggest a two-week pilot that would show whether the claim holds.
You get: Practical questions and a pilot plan instead of hype