What is AGI in AI?
AGI stands for artificial general intelligence: a system that can match or beat humans across a wide range of cognitive tasks, not just one. There is no agreed test for it. Google DeepMind's 2023 framework instead grades AI on two axes — performance depth (Emerging to Superhuman) and generality (narrow vs. general).
Why — the first-principles explanation
The "G" is the whole word. Today's AI is stuffed with narrow superhuman systems: a chess engine crushes every human alive and cannot tell you why the sky is blue. A protein-folding model outperforms decades of lab work and cannot add two numbers reliably. Each one is a spike — enormous height, no width. AGI names the other shape: not a taller spike, but a wide plateau that covers most of what human minds do, including tasks nobody trained it for.
Why is this so hard to pin down? Because "intelligence" was never a single number to begin with. The moment you try to write a test, you discover you're really asking three separate questions: how well does it perform, across how many domains, and how much autonomy does it have while doing it. Google DeepMind's Morris et al. paper, "Levels of AGI for Operationalizing Progress on the Path to AGI" (first posted November 4, 2023), treats intelligence as "a continuous, multidimensional construct" and splits it exactly that way. Performance gets six tiers: No AI, Emerging (equal to or somewhat better than an unskilled human), Competent (50th percentile of skilled adults), Expert (90th percentile), Virtuoso (99th percentile — later renamed Exceptional), and Superhuman (beats all humans). Each tier then splits into narrow and general. So a chess engine is Superhuman-Narrow. Today's frontier chatbots are usually placed around Emerging-General — broad, but not yet reliably at the level of a skilled adult across the board.
That grid explains why AGI arguments never resolve. Two smart people saying "we have AGI" and "we're nowhere close" are often both right, because they're pointing at different cells. One is impressed by the width — a single model that codes, translates, diagnoses, and writes poetry is genuinely unprecedented. The other is unimpressed by the depth and reliability — that same model still fabricates citations and fails tasks a competent adult wouldn't. There is no scoreboard, so the debate runs on vibes.
The last dimension is the one that actually determines risk. The DeepMind framework separates capability from autonomy on purpose, because they're independent: a highly capable model used as a consulted tool is a different animal from the same model running unsupervised with real permissions. Capability sets what's possible; autonomy sets what's at stake. Most of what people fear about AGI is really a fear about the autonomy axis.
An example that makes it click
Think about a decathlon versus the 100-meter dash. Usain Bolt is superhuman at one event — nobody on Earth beats him. But drop him into the pole vault, the discus, and the 1500 meters and he's an ordinary athlete having a bad afternoon. He's a spike: unbeatable height, no width.
A decathlete is the opposite. He loses the 100 meters to Bolt every single time, and he's not the best in the world at any of the ten events. But he can do all ten at a serious level, and if you added an eleventh event tomorrow, he'd probably be decent at that too. That's generality — and that transfer to the untrained event is the part that matters. AGI is asking for the decathlete, not the sprinter. And the reason people argue about whether we've arrived is that today's AI is a strange decathlete: it writes a passable poem, a passable diagnosis, and a passable legal brief, then confidently invents a court case that never existed. Wide, but you can't yet stop watching it.
Key facts
- AGI = artificial general intelligence: broad, human-level-or-better competence across many cognitive tasks, as opposed to narrow AI that excels at one.
- Google DeepMind's "Levels of AGI for Operationalizing Progress on the Path to AGI" (Morris, Sohl-Dickstein, Fiedel, Warkentin, Dafoe, Faust, Farabet, Legg) was submitted November 4, 2023 and last revised September 24, 2025.
- The framework's six performance tiers are No AI, Emerging, Competent (50th percentile of skilled adults), Expert (90th percentile), Virtuoso/Exceptional (99th percentile), and Superhuman (outperforms all humans).
- The framework grades on three separate axes — performance, generality, and autonomy — and treats intelligence as "a continuous, multidimensional construct" rather than a single threshold.
- There is no standardized, agreed-upon test that certifies a system as AGI; DeepMind's paper explicitly addresses the unsolved benchmarking problem.
- A 2012–2013 survey of AI experts by Bostrom and Müller found a median 50% confidence that human-level machine intelligence would arrive by 2040–2050.
▶ The 60-second explainer (script)
AGI means artificial general intelligence — a system that matches or beats humans across a wide range of tasks, not just one. The key letter is the G, for general. Because today's AI is full of narrow superhuman systems. A chess engine beats every human alive and can't tell you why the sky is blue. Each one is a spike: huge height, no width. AGI is asking for a plateau instead — broad coverage, including tasks nobody trained it for. So why can't anyone agree whether we have it? Because intelligence was never one number. Google DeepMind's Levels of AGI paper, first published in November 2023, splits it into three axes: how well it performs, across how many domains, and how much autonomy it has. Performance alone gets six tiers, from Emerging — roughly an unskilled human — up through Competent at the fiftieth percentile of skilled adults, Expert at the ninetieth, Virtuoso at the ninety-ninth, and Superhuman. Then each tier splits into narrow and general. So a chess engine is Superhuman-Narrow. Today's best chatbots usually get placed around Emerging-General: impressively wide, not yet reliably skilled-adult-level across the board. That's why two smart people can both be right when one says we have AGI and the other says we're nowhere close. They're pointing at different cells in the same grid. And notice the third axis — autonomy. Capability sets what's possible. Autonomy sets what's at stake.
What authoritative sources say
People also ask
Do we have AGI right now?
Not by most definitions. Frontier models are broad but still fall short of a skilled adult's reliability across the board — on DeepMind's grid they're usually placed at Emerging-General, well below Competent-General.
What's the difference between AGI and narrow AI?
Narrow AI is superhuman at one thing and useless outside it — a chess engine, a spam filter. AGI would perform well across many unrelated tasks, including ones it was never specifically trained on.
Is AGI the same as the singularity?
No. AGI is a capability level. The singularity is a hypothesized event where AI improving itself triggers runaway change. AGI is usually treated as a possible cause; the singularity is the claimed consequence.
When will AGI arrive?
Nobody knows, and expert forecasts vary by decades. A 2012–2013 Bostrom and Müller survey found a median 50% confidence by 2040–2050 — but forecasts like this have a long history of being wrong in both directions.
Who decides when we've reached AGI?
No one, currently. There's no accepted certification test — which is exactly the gap DeepMind's Levels of AGI paper tried to close by grading progress on axes instead of declaring a finish line.