AGI Made Clear › How the people building it define AGI
How the people building it define AGI
Ask seven of the most influential people in AI what AGI means and you don’t get seven versions of the same answer. You get different tests entirely. That helps explain why one person can say AGI is already here while another thinks it is still years away.
If you are trying to work out when AGI might arrive, there is a problem you have to deal with first. What exactly are we waiting for?
There is no single accepted definition of artificial general intelligence.
Some definitions are about work. Could an AI perform most of the economically valuable tasks humans currently do? Others are about capability. Could it learn, reason, plan and adapt across a very wide range of problems? Some set the bar at human-level performance. Others are more interested in whether AI can operate independently in the real world. And some of the people who used to talk regularly about AGI have started avoiding the term altogether.
That is why two intelligent people can look at exactly the same progress and give very different answers to the question: how close are we to AGI?
They may not actually be disagreeing about the technology. They may be measuring it against different definitions.
Seven people, seven tests
Jensen Huang NVIDIAMar 2026
Their testCould AI build and run a billion-dollar company?
In March 2026, Lex Fridman put a deliberately practical definition to NVIDIA chief executive Jensen Huang. Imagine an AI capable of starting, growing and running a successful technology company worth more than $1 billion. How far away would that be?[The Verge]
Huang’s answer was striking: “I think it’s now. I think we’ve achieved AGI.”
That quote travelled quickly, but the definition matters. Huang wasn’t declaring that an AI now possesses every capability associated with human intelligence. He was responding to a particular economic test proposed during the conversation. He also added important caveats when discussing what today’s AI agents could actually accomplish.
Under some practical or economic definitions, the AGI threshold can look much closer than it does under broader human-level definitions.
Their position: In March 2026 he said AGI could be considered achieved under the definition being discussed.
Greg Brockman OpenAISep 2026
Their testCan AI outperform humans at most economically valuable work?
OpenAI has historically defined AGI in its charter as highly autonomous systems that outperform humans at most economically valuable work. That is an economic definition. It doesn’t require an artificial mind that thinks exactly like a human. The emphasis is on what the system can do.
When OpenAI launched GPT-6 Astra in September 2026, president Greg Brockman ended the briefing with “Welcome to the AGI era.” Asked whether Astra itself represented the arrival of AGI, he was more cautious: “I think it might be about this model.”[Axios]
Brockman also said he personally believes OpenAI has reached AGI, while leaving others to decide whether Astra meets their definition. That qualification is important.
A system could satisfy OpenAI’s economic definition while still falling short of somebody else’s idea of broadly human-like intelligence.
Their position: September 2026. He says he personally believes OpenAI has reached AGI, while acknowledging that whether Astra qualifies depends on the definition being applied.
Sam Altman OpenAIAug 2025
Their testOne definition may no longer be enough
Sam Altman’s position demonstrates another problem with the AGI debate. Even within an organisation that has a formal definition, the people leading it don’t necessarily find the term particularly useful.
Asked for his definition of AGI in 2025, Altman replied: “I have many, which is why I think it’s not a super useful term.”
That is a revealing answer. The capabilities grouped together under “AGI” may not arrive simultaneously. A system could become exceptional at coding before it becomes dependable at planning. It might outperform humans at many commercially valuable tasks while remaining surprisingly weak in areas of everyday judgement.
Trying to compress all of that into one moment called “AGI” may eventually tell us less than examining the capabilities separately.
The disagreement isn’t only about when AGI arrives. There is disagreement about whether AGI remains a useful single threshold at all.
Dario Amodei AnthropicOct 2024
Their testWhat happens when AI becomes extraordinarily capable across many fields?
Anthropic chief executive Dario Amodei generally prefers the term “powerful AI” to AGI. His description sets an extremely high capability bar.
He has described a future in which AI systems could outperform very capable humans across a huge range of intellectual fields, operate with substantial independence and be replicated at enormous scale. He has used the image of a “country of geniuses in a data centre” to convey what that could mean.
The distinction matters because Amodei is not particularly interested in establishing the exact day on which the letters A, G and I apply. He is interested in what happens when the capability becomes powerful enough to transform science, work, economics and society.
You can believe extremely powerful AI is close while rejecting AGI as the best term for describing it.
Their position: He has placed extremely powerful AI within the next few years, although the exact capability being forecast matters more than attaching an AGI label to it.
Mustafa Suleyman Microsoft AIDec 2025
Their testHuman-level performance across a very broad range of tasks
Mustafa Suleyman has generally described AGI in terms closer to broad human-level capability. That sets a different bar from asking whether AI can perform economically useful work.
An AI might transform entire industries and still fail a definition requiring it to match the breadth and adaptability of human intelligence.
Suleyman has also pushed back against treating AGI as a simple race with a finish line and a winner. That makes his position particularly useful here.
The more capabilities you require to arrive together before using the term AGI, the further away the threshold can appear.
Their position: He has resisted giving the concept a simple arrival date and has questioned the usefulness of the race framing itself.
Demis Hassabis Google DeepMindJun 2026
Their testCan AI demonstrate the full breadth of human cognitive capability?
Demis Hassabis uses one of the broader definitions. His conception of AGI centres on a system exhibiting the cognitive capabilities humans possess, rather than merely completing enough economically valuable tasks.
That is a demanding test. It requires us to think beyond benchmarks and individual areas of superhuman performance. A system might be a better mathematician than almost every human and still fail this definition if it lacks capabilities humans routinely use elsewhere.
Hassabis has said he believes AGI could arrive around 2030, plus or minus roughly a year.[Fast Company]
A broad cognitive definition produces a very different threshold from an economic one.
Their position: Around 2030, plus or minus roughly a year.
Yann LeCun AMI LabsJan 2026
Their testGenuine human-level intelligence
Yann LeCun represents perhaps the clearest challenge to the assumption underneath many AGI forecasts. His argument is not simply that AGI will take longer. It is that today’s dominant AI architectures are missing important ingredients required for human-level intelligence.
Large language models can be extraordinarily capable, but LeCun argues that human-level AI will require major conceptual breakthroughs beyond simply making today’s systems bigger.
That makes his disagreement deeper than a difference over dates. If he is right, the route itself has to change.
Any AGI forecast depends partly on whether you believe today’s technology can actually get us there.
Their position: Human-level intelligence requires breakthroughs that current architectures have not yet delivered.
The disagreement starts to make sense
Put these positions beside one another and something important becomes clearer.
Imagine a future AI system that can write software, conduct research, analyse complex information, use computers, communicate persuasively and complete a large amount of professional work with limited supervision. Has AGI arrived?
Under an economic definition, perhaps. Under a definition requiring the complete breadth, robustness and adaptability of human cognition, perhaps not. Under a definition that requires fundamentally different mechanisms of intelligence, definitely not yet.
Same system. Different test. Different answer. That is why AGI timelines can look wildly contradictory even when the people giving them have access to similar evidence about what current systems can do.
Economic AGI and cognitive AGI
A useful way of understanding the divide is to separate two broad families of definition.
These ask what the system can do. Can it perform most economically valuable knowledge work? Can it carry out the work of a remote employee? Could it operate significant parts of a business? The advantage is that it gives us something relatively concrete to observe. The weakness is that economic usefulness and intelligence aren’t necessarily the same thing. A system could become enormously economically important without possessing anything resembling the full breadth of human cognition.
These ask something closer to what the system is capable of understanding and learning. Can it reason flexibly? Adapt to unfamiliar situations? Learn genuinely new tasks? Transfer knowledge from one area into another? Does it have the breadth and robustness of human intelligence rather than exceptional performance on particular categories of task? That is a much higher and much harder threshold to measure.
This is one reason somebody using an economic definition might say we are entering the AGI era while somebody using a cognitive definition says we are still years away.
There is another group: people giving up on the word
Something else has been happening as AI capabilities have advanced. Some prominent people in AI have become less enthusiastic about the term AGI itself.
That isn’t necessarily because they think progress has slowed. The opposite can be true. As AI systems acquire capabilities unevenly, the idea of one clean threshold becomes harder to defend.
Does AGI arrive when AI surpasses humans economically? When it matches us cognitively? When it can learn anything? When it can work autonomously for days? When it can improve AI research itself? When it can run a company?
Those could happen at very different times. So saying “AGI arrives in 2028” sounds precise while potentially concealing a much more important question: which version of AGI?
This is why we don’t give you one definition
It would be convenient for this site to put a box at the top of every page saying: AGI means this. It would also make many of the arguments on these pages artificially simple.
Instead, when somebody makes a claim about AGI, we want to know which threshold they are using. When somebody gives a date, we want to know what they expect to exist by that date. When two experts disagree, we want to know whether they are actually making conflicting predictions or simply answering different questions.
Because “when will AGI arrive?” isn’t really one question. It is several questions hiding inside the same three letters. And until we separate them, the timelines will continue to look far more contradictory than they really are.
How would we know if AGI had arrived? →