What Is AGI? A Plain Answer Without the Hype

What is AGI? The letters stand for artificial general intelligence, and the cleanest short definition I've found comes from Stanford's Institute for Human-Centered AI, which describes it as an AI system with general, human-level or beyond ability to learn, reason and apply knowledge across a wide range of tasks and domains. The word carrying the weight is general. Every AI you use today is strong at a few things and hopeless at the rest. AGI would be one system that could take on almost any thinking task a person can, including tasks nobody trained it for.

Does it exist yet? Not under that definition, and not under the scale Google DeepMind published for measuring progress toward it, which placed chatbots like ChatGPT and Gemini on the lowest rung. The complication is that the answer moves when the definition moves, and there are at least four definitions in serious use. Most explainers skip that. This one is mostly about it.

AGI Versus the AI You Already Use

Start with the difference, because it's the question people ask right after the definition.

The AI in your phone's camera, the spam filter on your email and the chatbot you tried last week are all what researchers call narrow AI. Each was built for a scoped job. A spam filter can't draft your reply. An image generator can't do your taxes. The chatbot can write an essay and it can write code, and that breadth is new, but it still can't pick up a skill on its own the way a new hire does, and it still fumbles tasks a careful adult would handle without thinking. If the umbrella term itself is fuzzy for you, the piece on AI versus machine learning sorts the vocabulary out first.

AGI is the label for the system that closes that gap. The difference is breadth. DeepMind's scale puts Deep Blue, the chess machine, at the 99th-percentile rung for its one narrow task, and nobody calls it AGI, because it did exactly one thing.

So AGI isn't a separate technology. It's a bar. Today's models are attempts to reach it, and whether they're close depends on how high you set the bar, which is where the definitions start to disagree.

Four Definitions and Why They Disagree

Here are the four I'd want a beginner to know, with what each would need as proof.

Read across the rows and you can see why arguments about AGI go in circles. OpenAI's version is about money. DeepMind's is about percentiles. Chollet's is about learning speed. Stanford's is the everyday meaning, and it admits it can't be tested cleanly. A system could clear one bar and miss the other three, and two people quoting different bars will talk past each other for an hour.

The DeepMind paper, by Morris, Sohl-Dickstein, Fiedel and colleagues, also set out six principles for a workable definition, and the first one settles a lot of dinner-table arguments. It says to focus on capabilities rather than processes. Under that principle, whether a machine "really thinks" or "really understands" is set aside. What matters is what it can do, measured against people. I think that's the right call for a beginner too. The philosophy is interesting, but you can't check it, and you can check a task.

Does AGI Exist Yet? Same Question, Four Answers

Put the four definitions against the same question and the answers line up, which is more than you'd expect.

So the plain answer is no, with a date attached. As of mid-September 2026, no source I could load places a public system at or above any of these bars, and no lab page I could load claims otherwise. Google DeepMind's own ten-year AlphaGo retrospective from March 2026 still describes its work as on the path to AGI and its goal as "on the horizon." The distance between "on the horizon" and "here" is where the marketing lives, so I'd keep that phrase in mind the next time a headline says a lab built AGI.

Where Today's Chatbots Sit on DeepMind's Scale

The Levels of AGI paper is the most useful thing I've read on this topic, because it replaces "is it AGI, yes or no" with "how far along is it." The scale has six rows for performance and two columns for breadth, narrow and general. Here is the general column, the one that matters for AGI, in plain words.

Two things about this table matter more than the labels.

The placement is dated. The authors wrote that as of September 2023, frontier chatbots showed Competent performance on some tasks, such as short essays and simple coding, but were still Emerging on most, and they named maths and factual accuracy as the weak areas. The paper was revised in September 2025 and kept that wording. Models released since then have not been placed on the scale by the authors, so the accurate reading is "Level 1 as of the last published check." Anyone telling you a newer model has climbed the ladder should be able to point at a measurement.

The bar for Level 2 is the median skilled adult across a wide range of non-physical tasks, including learning new skills. That's a high bar, and it's the level the paper says best corresponds to what most people have historically meant by AGI. When someone says a model is at PhD level, ask on which tasks and against whom. A model can be Expert at one narrow thing and Emerging in general at the same moment, and the paper says so directly.

The factuality weakness is the one you'll run into first. If a chatbot invents a citation for you, that's the "tasks involving factuality" line showing up in your own work, and I wrote a whole piece on why hallucinations happen if you want the mechanism.

The Tests People Have Proposed

Before anyone had a scale, they had tests. Three come up constantly.

Turing's paper opens with "I propose to consider the question, 'Can machines think?'" and then, famously, declines to answer it head-on. He swaps it for a game an interrogator can run. That move, from an unanswerable question to a checkable one, is the same move DeepMind made seventy-odd years later with its levels. It's also why I'd distrust any AGI claim that isn't attached to a test someone else can rerun.

The ARC Prize page states Chollet's definition in one sentence, that the intelligence of a system is a measure of its skill-acquisition efficiency over a scope of tasks. Efficiency is the word to notice. A system that needed a million examples to learn what a child learns from three would score low even if it got the answers right.

Is ChatGPT AGI?

No, and this is the question where the disagreement is loudest, so precision helps.

A 2023 paper from Microsoft Research titled Sparks of Artificial General Intelligence said of an early version of GPT-4 that "it could reasonably be viewed as an early (yet still incomplete) version" of an AGI system. That sentence gets quoted with the hedges removed. The authors said early and incomplete in the same breath, and the paper is an argument by its authors rather than a measurement against a published scale. My concern with how it gets used is that "sparks" became "AGI" somewhere between the abstract and the headlines.

DeepMind's scale gives the more careful answer. ChatGPT is a Level 1 general system on that scale, Emerging AGI, which is a real category and a low one. The same paper takes on the argument that breadth alone is enough. Chatbots can discuss almost anything, and some researchers say that already makes them AGI, but the authors' reply is that breadth has to be paired with performance, and on most tasks the performance was Emerging.

So the tidy version is that ChatGPT is a large language model, which is one kind of AI, and you can read what an LLM actually is for the mechanics. AGI is a bar it hasn't cleared under any of the four definitions above. Both things are true at once, and most of the confusion comes from people using "AGI" when they mean "impressive."

When Do Experts Think It Will Arrive?

The largest survey I know of asked 2,778 researchers who had published at top AI venues. The headline estimate was a 10 percent chance of unaided machines outperforming humans in every possible task by 2027, and a 50 percent chance by 2047. That 2047 figure was 13 years earlier than the same group's answer only a year before.

I'd take two things from that and no more.

Experts moved their estimates a long way in a single year, which tells you the forecasts are soft. And the median still sits about two decades out, which is a long way from "next year," whatever a launch event implies. Well-known researchers and executives have offered shorter timelines in interviews, and the survey is a better anchor than any one of them because it averages over thousands of people who don't share a marketing department.

Is AGI possible at all? The survey respondents are betting yes, and nobody has produced a proof that it isn't. The honest position for a beginner is that it's an open engineering question with a wide spread of informed guesses, and that the definition you're using changes what "arrive" would even mean.

Who Came Up With the Term

Ben Goertzel, who edited a 2005 book with the phrase in its title, says the name was suggested by Shane Legg around 2002, when Goertzel emailed friends looking for something better than "Real AI." Goertzel also notes that a researcher named Mark Gubrud had used "artificial general intelligence" in a 1997 article, so the term is older than its popular use. DeepMind's paper credits Legg and Goertzel with spreading it among computer scientists in the early 2000s, and Legg's name is on the Levels of AGI author list, which closes the loop neatly.

AGI on Your Tax Return Is Something Else

If you searched this phrase around tax season, you may want a different AGI entirely. On US tax forms, AGI means adjusted gross income. The IRS defines it as your total income from all sources minus certain adjustments listed on Schedule 1 of Form 1040, and it appears on Form 1040, line 11. Nothing to do with artificial intelligence. The two acronyms only collide because English ran out of three-letter combinations. If that's your question, the IRS page is the right next stop, and the rest of this article won't help.

How to Read the Next AGI Headline

The term will keep showing up, usually with no definition attached. Here's the checklist I'd run.

Ask which definition. Economic work, percentile against adults, learning efficiency, or the everyday meaning. If the article doesn't say, the claim can't be checked.

Ask for the test. A claim tied to a benchmark someone else can rerun is worth more than a demo, and a demo is worth more than an adjective.

Ask for the date. DeepMind's placement of chatbots at Level 1 is dated September 2023 in the paper's own text. A 2026 model needs a 2026 measurement, and until someone publishes one, "Level 1 as last measured" is the accurate phrase.

Separate the model from the product wrapped around it. A chatbot that searches the web and runs tools is a system built around a language model, and the tools change what it can do without changing what the model is. The explainer on AI agents covers that layer, and DeepMind's paper has a matching scale for it, with autonomy levels running from AI as a tool up to AI as an agent.

AGI, then, is a bar with several published heights. Nothing you can use today clears the lower ones by the accounts of the people who set them, the experts' median guess sits decades out, and the gap between the everyday meaning and the measurable ones is exactly where the arguments come from. Knowing that much puts you ahead of most of the people using the word.