Aiodic

Back to the table

116 Emerging Gn AGI Frontier

AGI

Artificial general intelligence.

What it is

AGI stands for artificial general intelligence: an AI system that could do most of the thinking work people do, about as well as people do it, across many different kinds of task rather than just one. It does not exist today by most definitions, although by September 2026 some company leaders had started to say otherwise, and the argument about that is part of the story. That is why this element sits on the frontier of the table.

The key word is "general", and it's easiest to understand by comparison with its opposite. Almost every AI system ever built is narrow. A chess program plays chess superbly and can do nothing else. A spam filter sorts email and cannot tell you what the email means. A satnav plans routes and cannot help you write a wedding speech. Each is a specialist, like a calculator: brilliant inside its box, helpless outside it.

A person is general. The same brain that learned to ride a bike can learn to cook, argue, fix a leaking tap, comfort a friend, and pick up a new job. More importantly, it can learn things nobody trained it for. Drop an experienced office worker into an unfamiliar company and within a few weeks they'll have worked out the software, the unwritten rules, and who actually makes decisions. That ability to handle the new and unexpected, rather than a long list of skills, is what most people mean by general intelligence.

Today's chatbots, built on language models, blur this line, which is why AGI went from a niche idea to a boardroom topic. A modern chatbot can write a poem, explain tax rules, draft code, and translate a menu, so it feels general. But it also has odd holes. It can hallucinate, inventing facts with confidence; it usually forgets everything between conversations unless given memory tools; it struggles with long, messy, real-world tasks; and it can fail puzzles a child solves. Researchers describe this as a "jagged" profile: superhuman on some things, below average on others.

Where the word came from

The phrase is younger than most people assume. The earliest known use is by the physicist Mark Gubrud, in a paper for the Fifth Foresight Conference on Molecular Nanotechnology in 1997, where he worried about machines "usable in essentially any phase of industrial or military operations" that could rival the human brain. The term spread later, and separately. According to the AI researcher Ben Goertzel, in 2002 he and Cassio Pennachin were stuck for a title for a book on human-level AI, asked colleagues for ideas, and Shane Legg (who later co-founded DeepMind) suggested "Artificial General Intelligence". Goertzel only learned of Gubrud's earlier use years afterwards. Legg told MIT Technology Review in 2023 that at the time he didn't have an especially clear definition in mind.

That vagueness has never gone away. In 2007 Legg and the computer scientist Marcus Hutter gathered around 70 published definitions of intelligence itself, and the list has only grown since.

Three families of definitions

Most definitions of AGI fall into three families, and it helps to know which one someone is using.

Some prominent people avoid the word altogether. In his October 2024 essay "Machines of Loving Grace", Anthropic's chief executive Dario Amodei wrote that he finds AGI "an imprecise term that has gathered a lot of sci-fi baggage and hype", and used "powerful AI" instead.

AGI is also not the same as superintelligence. AGI means roughly matching people across the board. Superintelligence means going far beyond the best humans. Some people expect one to follow quickly from the other; that is a separate argument, covered on its own page.

Why it matters

Jobs and money. If the economic definition is ever met, most kinds of paid work change at once, not one profession at a time. That is why governments, unions, and investors follow AGI claims closely, and why the claims move share prices. Even well short of AGI, the direction of travel matters for anyone planning a career or a business.

Safety. A narrow system can only fail within its box. A general system that can act in the world, as agents increasingly do, can fail in far more ways. This is why alignment, the work of making AI pursue what people actually intend, gets more urgent the more general systems become. It's also why the DeepMind levels paper pairs capability with autonomy: how much a system is allowed to do on its own should depend on how capable it is.

Power. Whoever builds the first systems that can do most knowledge work would hold enormous economic and political influence. OpenAI's charter itself talks about avoiding uses of AI that "unduly concentrate power". The sceptics' concern is the mirror image: that talk of AGI can be used to attract investment and deflect regulation, whether or not the technology arrives.

Timelines, and why they differ so much. Predictions range from "already here" to "decades away" to "never with today's methods". Some dated examples, each a view rather than a fact:

Why do smart people land so far apart? Three reasons do most of the work. They mean different things. In that same 2023 survey, when the question asked when every occupation could be fully automated, rather than every task, the 50 percent date moved to 2116, nearly seventy years later. Same people, different wording, wildly different answer. They extrapolate differently. Optimists point to steady trends, such as METR's finding that the length of task AI agents can complete has been doubling roughly every seven months (see Agent). Sceptics point out that trends bend, and that fluent language is not the same as understanding. They have different incentives. Lab leaders raise money partly on the promise of AGI; critics build careers on doubting it. Neither fact makes anyone wrong, but it's worth knowing.

The computer scientist Melanie Mitchell, writing in Science in March 2024, highlighted a deeper split: many AI practitioners treat intelligence as something that can be separated from a body and a social world, while many people who study human and animal minds doubt that. That is one reason some researchers think world models, embodied AI, or continual learning will be needed before anything deserves the name.

Real-world examples

Google DeepMind proposes "levels" instead of a finish line

In November 2023 a Google DeepMind team led by Meredith Ringel Morris, with Shane Legg among the authors, published "Levels of AGI". Their argument was that "is it AGI, yes or no?" is the wrong question, in the same way that "is this car self-driving?" is less useful than the numbered levels the car industry uses.

The paper grades performance in five steps: Emerging (equal to or somewhat better than an unskilled person), Competent (at least the 50th percentile of skilled adults), Expert (90th percentile), a fourth level first called Virtuoso and later renamed Exceptional (99th percentile), and Superhuman (better than every human). Each level has a narrow column and a general column. A chess engine is superhuman but narrow. The authors placed chatbots of the day, including ChatGPT and Google's Bard, at "Emerging AGI": general, but at the level of an unskilled person, and said no general system had reached a higher level. They also listed principles, such as judging what a system can do rather than how it works, and leaving out physical tasks.

Whatever you think of the exact boundaries, the idea is a useful habit for readers. When someone says a system "is AGI", you can ask: at which level, across how many tasks, compared with which people?

ARC-AGI: a test designed to be easy for people

In 2019 the Google engineer François Chollet published a paper arguing that intelligence isn't a pile of skills but the efficiency with which you acquire new ones. To test it, he made the Abstraction and Reasoning Corpus, now ARC-AGI: small coloured-grid puzzles, each with its own hidden rule, shown through a couple of examples. Most people find them fairly easy. Memorising the internet doesn't help, because each puzzle is new.

In 2024 Chollet and the entrepreneur Mike Knoop launched the ARC Prize to reward progress. The puzzles became a public yardstick: in December 2024 OpenAI's o3 scored highly on the first version, at great computing cost (the details are in Inference). The foundation responded in March 2025 with ARC-AGI-2. It tested the new puzzles on more than 400 people; every task was solved by at least two of them, and the average person scored about 60 percent. At launch, leading chatbots scored close to zero. By December 2025 the best commercial model the foundation listed scored 37.6 percent, a refined system built on Google's Gemini 3 Pro scored 54 percent, and the $700,000 grand prize for reaching 85 percent within the competition's limits was still unclaimed. The foundation's own summary asked whether we had AGI and answered: "Not yet."

A third version, ARC-AGI-3, replaced static puzzles with small interactive games where the player has to explore and work out the goal. The organisers have always said that beating their test would not by itself prove AGI. It measures one important ingredient, learning something new from very little, and nothing else.

AGI written into a contract

AGI stopped being only a research idea when it became a clause in a business deal. When Microsoft invested in OpenAI in 2019, OpenAI said it would license its "pre-AGI" technology to Microsoft, which implied that AGI technology would be treated differently. In December 2024 the tech publication The Information reported that, under an agreement reached in 2023, the two companies had defined AGI for contract purposes as OpenAI developing systems able to generate at least $100 billion in profits. That's not a scientific definition at all, and it showed how much money could hang on the word.

On 28 October 2025 the companies announced a restructured deal: once OpenAI declared AGI, "that declaration will now be verified by an independent expert panel", and Microsoft's rights to OpenAI's technology were extended through 2032, including models after AGI. Then, on 27 April 2026, a further amendment said OpenAI's revenue-share payments to Microsoft would continue through 2030 "independent of OpenAI's technology progress". Commentators, including the developer and writer Simon Willison, read that as the end of AGI as a trigger in the contract. The lesson for readers: when a company talks about AGI, it may also be talking about money, and the definition may be doing legal work.

"The AGI era" and the pushback, September 2026

On 3 September 2026 OpenAI released GPT-6 Astra, a model it described as state of the art at using computers, coding, and science. Speaking to reporters before the launch, OpenAI's president Greg Brockman said: "I think it's not unreasonable to feel that we are now in the AGI era." Nvidia's chief executive Jensen Huang posted that AGI had arrived.

The response shows how to read such a claim. The ARC Prize Foundation published its own test of Astra on ARC-AGI-3 the same day. The near-perfect score OpenAI had highlighted came from a setup that used OpenAI's own features; on the foundation's standard, provider-neutral setup, Astra scored 62.7 percent. The foundation called the result meaningful progress but wrote: "we are not claiming that it is AGI." Academics quoted by the Australian publication Information Age, including Toby Walsh of UNSW Sydney, were sceptical that the model matched all human thinking abilities. Notice that Brockman's own wording was careful: "not unreasonable to feel". A strong model, a hedged claim, and a disputed test result can all be true at once.

An everyday scenario: the AGI headline at work

Imagine Dana, who runs a small bookkeeping firm. Her business partner forwards a headline, "AGI is here", and asks whether they should stop hiring. Dana asks three questions. Which definition is being used? (An executive's feeling, not a test.) What was measured? (A benchmark score, disputed by the people who run the benchmark.) What can the tool do for their clients this month? She spends an afternoon testing the new model on ten of their real jobs, finds it excellent at categorising receipts and unreliable at reconciling a messy account, and keeps hiring while giving the receipts to the AI. The headline didn't change her plan. Her own test did.

Try it yourself

This exercise helps you form your own view instead of borrowing someone else's. You need a chatbot and about 45 minutes.

  1. Write your own AGI checklist. List ten abilities you'd expect from something you'd call generally intelligent. Mix easy and hard. For example: learn a new card game from the rules alone; remember what you told it last week; admit when it doesn't know; plan a multi-day task and carry it through; notice when its own answer is wrong; understand a joke in context; do simple physical reasoning (will this stack of books topple?); handle a situation nobody warned it about; explain a local custom it hasn't seen written down; and learn from a mistake so it doesn't repeat it.
  2. Ask a chatbot to challenge your list. Paste this:
Here is my list of ten abilities I think an AI would need before I'd call it
"general intelligence": [paste your list].
Tell me which items are about knowledge, which are about learning something
new, and which are about acting in the world. Point out anything important
I've left out, and any item that is too vague to test. Don't tell me whether
AGI exists; just help me sharpen the list.
  1. Test three items. Pick three abilities you can test in a chat. For learning something new, try:
I'm going to teach you a made-up game. Rules: a word scores 1 point for each
vowel, doubled if the word starts with a consonant, and zero if it has a
letter repeated. Score these words and show your working: banana, tiger,
orbit, fleece, oak.

Check the answers yourself. (Banana scores zero because letters repeat; tiger scores 4.) For admitting ignorance, ask about something obscure but checkable in your own life or town, and see whether it guesses or says it doesn't know. For novelty, try a few puzzles at arcprize.org/play first yourself, then describe one to the chatbot in words.

  1. Score honestly. For each of the three, write pass, partial, or fail, and one sentence on why. You'll probably find the jagged profile: impressive on some items, oddly weak on others.
  2. Write down what would change your mind. This is the most important step. Finish these two sentences: "I'd believe AGI is here when an AI can…" and "I'd believe AGI is far off if, by [year], AI still can't…". Make them specific and testable.
  3. Use it as a filter for headlines. Next time you read an AGI claim, paste it in with this prompt:
Here is a news article claiming progress toward AGI: [paste text].
List: (1) which definition of AGI it uses, if any; (2) what was actually
measured and by whom; (3) whether the test was run independently;
(4) who benefits if readers believe the claim; (5) what the article
leaves out. Quote the article for each point.

Then check its quotes against the article, because the chatbot can make mistakes too.

Common mistakes and misunderstandings

Treating AGI as one clear line. There is no agreed definition, test, or authority. When someone says AGI has or hasn't arrived, the first question is always "by which definition?"

Hearing a prediction as a fact. Every timeline in this article is an opinion from a named person or survey, on a stated date. Survey answers shifted by decades depending on how the question was worded. Treat dates as a range of views, not a schedule.

Thinking "it can talk about anything" means "it can do anything". Fluent conversation is impressive, but broad knowledge isn't the same as the ability to learn new things, stay reliable over long tasks, or act sensibly in the physical world. Benchmarks such as Humanity's Last Exam test knowledge; ARC-style tests and long real-world tasks test other things.

Trusting a benchmark score on its own. Scores depend on setup, cost, and who ran the test, as the Astra example showed. Look for independent results and for what the test does not measure.

Assuming AGI means robots or consciousness. Most definitions, including DeepMind's, are about thinking tasks, and leave out physical ability and inner experience. Whether a system is conscious is a separate and even less settled question.

Mixing up AGI and superintelligence. AGI is roughly human level across the board; superintelligence is far beyond it. Arguments about the second don't settle the first.

Waiting for AGI before taking AI seriously. Narrow and "jagged" systems are already changing work. The practical question for most readers isn't when AGI arrives, but which of your tasks today's tools do well, which they do badly, and how you'd tell. That's what evals are for.

Related elements