AGI vs AI: The Labs' Own Definitions, Side by Side

In brief

AI is the whole field, and AGI is a hypothetical kind of AI that would match humans across most tasks. OpenAI's charter defines it as highly autonomous systems that outperform humans at most economically valuable work. Google DeepMind's paper rates today's chatbots Level 1 of 5 and lists Level 2 general AI as not yet achieved.

What is the difference between AGI and AI?

AI is the umbrella for software that does tasks that normally take human intelligence, and AGI is the claim that one system could do most of them as well as a person. The table splits the four terms people mix up.

A small glass cube on a low black plinth beside a tall stepped staircase of five glass blocks that rises into the dark, with a thin thread running from the cube to the first step, in a charcoal studio under white rim light.
TermWhat it meansDefined byExists today?
AIThe whole field, from spam filters to chatbotsNo single ownerYes
Narrow AIStrong at one clearly scoped taskDeepMind's paper: "Narrow" columnYes. Stockfish and AlphaFold sit at Level 5 narrow
AGIBroad ability across many tasks, at a stated levelOpenAI: outperforms humans at most economically valuable work. DeepMind: Levels 1 to 4 generalDeepMind's paper lists Level 2 to 4 general AI as not yet achieved
ASIGeneral ability that beats every humanDeepMind: Level 5 Superhuman, outperforms 100% of humansDeepMind's paper lists it as not yet achieved

So the two are not rivals. AGI is a subset of AI that doesn't exist at full strength yet, and ASI would sit above it.

How do OpenAI and Google DeepMind define AGI?

OpenAI defines AGI by economic output, and Google DeepMind defines it by percentiles of skilled adults. The two sentences don't describe the same finish line.

OpenAI's charter says its mission is to ensure that artificial general intelligence, "by which we mean highly autonomous systems that outperform humans at most economically valuable work", benefits all of humanity. That is a jobs test. Nothing in it says the system must think like a person.

DeepMind's paper, Levels of AGI by Meredith Ringel Morris, Shane Legg and six co-authors, first appeared on 4 November 2023 and reached version 5 on 24 September 2025. It treats AGI as a ladder, not a line. Depth is how well a system performs, and breadth is how many kinds of task it covers.

LevelNamePerformance barGeneral example in the paper
0No AINoneHuman in the loop tools such as Amazon Mechanical Turk
1EmergingEqual to or somewhat better than an unskilled humanChatGPT, Bard, Llama 2, Gemini
2CompetentAt least the 50th percentile of skilled adultsNot yet achieved
3ExpertAt least the 90th percentileNot yet achieved
4ExceptionalAt least the 99th percentileNot yet achieved
5SuperhumanOutperforms 100% of humansASI, not yet achieved

Level 4 was first called Virtuoso and was renamed Exceptional in version 5. The paper placed chatbots at Level 1 because they reached Competent on some tasks, such as short essays and simple coding, but only Emerging on others, such as math and factuality. Its table dates from 2023, so check the current version before you quote a model's level.

The paper also sets six principles. Two matter most here: define AGI by what a system can do, not how it works, and define levels of progress, not one endpoint.

What is the difference between AI, AGI and ASI?

AI exists, AGI is a goal that labs define differently, and ASI is a level above every human. In DeepMind's framing the same scale covers all three: narrow systems at Level 5 already beat every human at one thing, while general systems stop at Level 1.

That is the point most explainers miss. A chess engine is superhuman at chess, and DeepMind's paper lists Deep Blue and AlphaGo at Level 4 narrow and Stockfish and AlphaFold at Level 5 narrow. None of them is AGI, because none can do anything else. ASI is the same superhuman bar applied across the board.

Hassabis put the honest limit on this in March 2025. CNBC reported that "no one really knows" when artificial superintelligence will arrive, Hassabis said, in a story on NBC News.

What are examples of AGI vs AI?

Three real cases and one thought experiment show where the line sits. None is a prediction.

  1. AlphaFold and Stockfish. DeepMind's paper rates both as Level 5 narrow AI. They beat every human at one job, protein structure and chess. Neither can book a flight or write a letter, so neither is general.
  1. A chatbot writing an essay. The same paper rated chatbots Competent on short essays and Emerging on math and factuality as of 2023. One system, very uneven results. That unevenness is why the paper uses a ladder.
  1. Wozniak's coffee test. Steve Wozniak proposed that a real AGI could walk into an unfamiliar house and make coffee, finding the machine, the beans and the water. Zapier cites it. It tests physical, open world ability, which DeepMind's paper says is a plus but not a requirement.
  1. Zapier's AGI project manager. This one is a thought experiment, not a product. Zapier imagines a system that takes a vague goal, such as launching a product line by Q3, and then builds the plan, assigns the tasks and settles team conflicts without human input. Nobody has shown a system doing that.

Software agents are a fifth case people confuse with AGI. An agent is a system that takes actions toward a goal, and it can be narrow. Agentic AI vs generative AI covers the split.

Who claims what about AGI, and when?

The four lab leaders below disagree on the date, the definition and even the word. Each row is dated and quoted from a page we read on 21 November 2026.

SaidWhoClaimSource
Oct 2024Dario Amodei, AnthropicCalls AGI "an imprecise term that has gathered a lot of sci-fi baggage and hype". Says powerful AI "could come as early as 2026, though there are also ways it could take much longer."Machines of Loving Grace
9 Feb 2025Sam Altman, OpenAICalls AGI "a weakly defined term". Generally means a system that can tackle increasingly complex problems, at human level, across many fields (our paraphrase of the rest of his sentence).Three Observations
17 Mar 2025Demis Hassabis, Google DeepMindSays AGI will start to emerge in the next five or 10 years. "We're not quite there yet."CNBC via NBC News
22 Jan 2026Yann LeCun, AMI LabsCalls the belief that scaling LLMs reaches human level intelligence "simply false". Says it "is going to take a while".MIT Technology Review

Hassabis defined AGI to CNBC as "a system that's able to exhibit all the complicated capabilities that humans can." That is broader than OpenAI's jobs test and closer to DeepMind's ladder.

Amodei avoids the word and uses "country of geniuses in a datacenter" for his version, a system smarter than a Nobel Prize winner across most fields. Altman's dated calls are scored one by one in Sam Altman's AGI predictions, and Amodei's in Dario Amodei's predictions. We make no forecast of our own.

How would we know AGI has arrived?

Nobody agrees on one test, and each camp tests something different. The table lists the tests we could open at their own pages.

WhoWhat it testsSource
OpenAIWhether systems outperform humans at most economically valuable workOpenAI Charter
Google DeepMindThe percentile of skilled adults a system beats across a broad set of tasks, scored on a "living benchmark"Levels of AGI
ARC PrizeLearning efficiency: "AGI is a system that can match the learning efficiency of humans."ARC Prize
Steve Wozniak, via ZapierWhether a system can make coffee in a house it has never seenZapier

ARC-AGI is the best known public benchmark of the group. François Chollet introduced it in 2019 in the paper On the Measure of Intelligence, and ARC Prize describes its tasks as easy for humans and hard for AI. It defines intelligence as how efficiently a system picks up new skills on unknown tasks, not how many skills it already holds. That is a third definition, and it differs from both lab definitions above.

DeepMind's paper says an AGI benchmark should be "a living benchmark", with a way to agree on new tasks over time. It also says the tasks should include metacognitive ones, such as learning a new skill and knowing when to ask for help. The paper does not propose the benchmark itself. So no single score settles the question, and a claim of AGI should name the test it passed.

Will AGI take jobs, and is it dangerous?

That depends on the definition, which is why it belongs here. OpenAI's version is a work test: systems that outperform humans at most economically valuable work. By that wording, what AGI means and what it does to jobs are close to the same question.

On risk, the abstract of DeepMind's paper says it discusses how the levels interact with deployment considerations such as autonomy and risk. Zapier's explainer says AGI could be self directed where today's AI works under human set rules. Both describe what AGI might be, and neither reports harm. We make no forecast on either. Will AI take over the world takes the risk question on its own.

Is ChatGPT AGI?

No lab or paper cited here calls today's chatbots full AGI. DeepMind's table lists ChatGPT, Bard, Llama 2 and Gemini as Level 1 Emerging general AI, and lists Levels 2 to 4 as not yet achieved. Hassabis said "We're not quite there yet" in March 2025.

Whether a chatbot counts as narrow or general depends on the definition. By breadth it is general, since one model writes, codes and translates. By depth it is uneven. That is why older explainers that call all current AI narrow now read as out of date, and why the paper uses a ladder. For how language models work, see what is an LLM.

The AI vs AGI line also moves with the speaker. Under OpenAI's economic test, the question is how much valuable work a system does without help. Under DeepMind's, it is percentiles. Neither is a settled standard, and no outside body certifies either one.

What can you ask an AI about AGI?

Paste any of these into a chatbot with web search on, and ask it to link every quote. Open the links before you trust the answer.

  1. "Quote the AGI definition from OpenAI's charter exactly, then give the URL. If you can't open the page, say so."
  1. "List the six levels in Google DeepMind's Levels of AGI paper with the percentile for each, and mark any you are unsure of."
  1. "Compare how OpenAI, Google DeepMind and Anthropic each define AGI or its nearest term. Use only their own pages, and say not stated where a page is silent."
  1. "Give me the dated AGI timeline claims of [name of a lab leader], each with the original source. Leave out any claim you can't source."
  1. "Explain AGI vs ASI to a [age] year old using one real example from DeepMind's paper."
  1. "Which of these claims about [system name] being AGI could a third party test, and which can't be tested?"

Where do the definitions fall short?

No definition above has an independent referee. OpenAI's test depends on what counts as economically valuable work, and DeepMind's depends on who counts as a skilled adult. Both are set by the labs that build the systems.

Several popular explainers add numbers that no one can source. One page ranked high for this query states the share of narrow AI as 95%, then nearly 90%, then about 85%, with no source for any of them. Treat unsourced percentages as decoration. Others describe superintelligence in the intro and never define it.

Our sources also have limits. DeepMind's table dates from 2023, and its authors work at Google. Hassabis, Altman and Amodei all lead companies that sell AI, and LeCun now leads a company that sells a different approach. Later 2026 interviews exist for some of them, so check the date before you quote a number.

Where does Whizi fit?

Whizi is one app instead of a pile of single purpose tools, and the models in it are today's chatbots. Pro carries Claude, GPT and Gemini, and its first week is $1.99. The free plan gives 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card.

Web search is on by default per chat, so the prompts above can fetch the primary pages. Side by side Compare, on Pro and Powerhouse, runs two models on one prompt so you can see where their definitions of AGI differ. Whizi loses to the primary sources themselves: for a quote, read the charter and the paper, not any chatbot's summary.

Workflow checklist
  • Name which definition of AGI the claim uses before you weigh it
  • Ask for both depth (how good) and breadth (how many tasks)
  • Treat "may", "could" and "as early as" as part of the prediction
  • Quote only words you read on the original page
  • Check the date: an AGI claim from 2023 is not a claim about 2026
Common questions

Frequently asked questions

Is AGI better than AI?

AGI isn't a better product, it is a different target. AI today beats humans at narrow jobs such as chess and protein folding, and AGI would be one system that does most jobs at a stated level. DeepMind's paper lists general AI above Level 1 as not yet achieved.

What is the difference between AGI and ASI?

ASI is AGI taken to the top of the scale. In DeepMind's Levels of AGI paper, Level 5 Superhuman means a system that outperforms 100% of humans, and the paper lists artificial superintelligence as not yet achieved. AGI covers Levels 1 to 4.

Does AGI exist yet?

Not by the labs' own sources we read. DeepMind's table lists Competent, Expert and Exceptional general AI as not yet achieved, and Hassabis said "We're not quite there yet" in March 2025. Whether it counts depends on the definition, which no outside body certifies.

Is an AI agent the same as AGI?

No. An AI agent is software that takes actions toward a goal, and it can be narrow, such as an agent that only files expense reports. AGI is about breadth of ability across tasks. See what is an AI agent.