AGI Made Clear › How would we know if AGI arrived?
How would we know if AGI arrived?
There is no agreed test for AGI. And as AI gets better at the tests we do have, an awkward question keeps returning: are we measuring general intelligence, or simply getting better at the test?
If somebody announced tomorrow that AGI had arrived, how would we check?
That sounds like it should have a straightforward answer. Run the system through a sufficiently difficult test. Compare it with humans. See whether it passes.
The problem is that we have been testing AI for years. And the better AI gets, the more complicated the testing becomes.
Benchmarks that once looked extraordinarily difficult have been beaten. New ones have been created specifically to address the weaknesses of the old ones. Some of those have then been beaten too. But every impressive score comes with another question: what exactly did we just prove?
A benchmark can tell us something important without telling us everything
Benchmarks aren’t useless. Far from it. Without them we would be left comparing demonstrations, anecdotes and claims from the companies building the systems. A good benchmark lets us ask the same question of different models and see how their performance changes.
But a benchmark measures what it was designed to measure. AGI is supposed to describe something much broader.
We usually associate general intelligence with some combination of reasoning, learning, adaptability, planning, knowledge, problem solving, judgement and the ability to deal with situations that haven’t been encountered before. No single benchmark captures all of that.
And increasingly, some of the most interesting evidence comes from looking at the gaps between different tests.
Consider what happened to ARC
ARC was created around a deceptively simple idea. Give an AI unfamiliar visual reasoning problems and see whether it can infer the underlying rule. The tasks were deliberately designed to make memorisation less useful.
Then models got much better at them. ARC-AGI-2 raised the difficulty considerably. Models got much better at that too.
So ARC-AGI-3 changed the problem again. Instead of simply presenting a puzzle to solve, it places the AI in unfamiliar interactive environments. The system has to explore, work out what the environment is doing, learn from what happens and then act. That sounds much closer to something we would associate with general intelligence.
And then GPT-6 Astra arrived.
The same model scored 62.7% and 99.9%
This is one of the most useful examples of why AI benchmarks need to be read carefully.
On ARC-AGI-3, GPT-6 Astra scored 62.7% using ARC Prize’s standard harness. Using a Provider Adapter that preserved the model’s internal reasoning state between requests and compacted longer conversations, it scored 99.9%.[ARC Prize]
Neither number is necessarily wrong. They are measurements of the system operating under different conditions. And that distinction matters enormously if we are trying to use benchmark scores as evidence for AGI.
Because once the scaffolding around a model can change its score from 62.7% to 99.9%, the question isn’t simply how intelligent the model is. It also becomes: what exactly are we measuring when we measure an AI system? The underlying model? The model plus memory? The model plus tools? The model plus a particular reasoning setup? The complete system?
That question is going to become harder, not easier, as AI systems become more agentic.
Then there is the problem of genuinely new knowledge
Mathematics gives us another unusually useful test. Most AI benchmarks have answers. Somewhere, somebody already knows the solution.
That creates an unavoidable problem. Even if the exact test question wasn’t in a model’s training data, the concepts, methods and related problems may have been. So how do you distinguish between an AI that is extremely good at using existing human knowledge and one that can genuinely discover something new?
One answer is obvious: give it problems nobody has solved.
68 problems with no known solution
In September 2026, Epoch AI introduced FrontierMath Erdős. It contains 68 significant mathematical problems associated with Paul Erdős that were still open as of August 2026. The answers therefore weren’t sitting somewhere in the training data waiting to be retrieved. A system has to produce a proof or disproof that can be formally verified.[Epoch AI]
GPT-6 Astra solved two. That is a score of 3%. Other tested systems scored zero under the same evaluation budget.
Three percent doesn’t sound impressive beside the enormous scores appearing on mature AI benchmarks. But those two solutions matter. The system contributed genuine solutions to problems that were previously open.
At the same time, Epoch reports Astra at around 98% on FrontierMath Tier 4, an exceptionally difficult benchmark containing problems that already have solutions.
These are performance on exceptionally hard problems with known solutions, against performance on problems that were genuinely open at the benchmark’s cutoff. That is not the same as retrieving versus producing: a model solving an already-solved problem may well be reasoning its way there independently. What the gap does show is that today’s best systems can be extraordinarily capable at exceptionally hard mathematics while remaining far less capable when the answer genuinely has to be discovered. Both numbers matter.
Learning may be an even bigger test
There is another capability that feels fundamental to the idea of general intelligence. Can the system actually learn from experience?
Humans don’t arrive in the world with a fixed model and then remain unchanged. We try something. It doesn’t work. We adjust. We remember what happened. We get better.
Current AI systems can appear to do this within a conversation because information can be added to their context. But that is not necessarily the same thing as learning.
Epoch AI tested this in 2026 using an obscure cooperative board game. Models were allowed to play repeatedly and could keep notes between attempts. You might expect them to improve. They largely didn’t. Newer models performed better than older ones, but they tended to start better rather than becoming substantially better through repeated experience.[Epoch AI]
That is a significant distinction. Building a better model is not the same capability as a model becoming better through experience after it has been deployed. If that changes, it will be worth paying very close attention.
And then there is reliability
AI capability stories usually focus on the best thing a system can do. But in the real world, another question matters just as much: how often can it do it?
A model demonstrating a capability once is interesting. A model performing it reliably enough that a person or business can depend on it is something different.
This is particularly important as AI moves from answering questions to carrying out work. If an agent is asked to complete a task that takes hours or days, one impressive successful run doesn’t tell us enough. Can it stay on track? Can it recognise when something has gone wrong? Can it recover? Can it avoid taking shortcuts that technically satisfy an evaluation while defeating its purpose? Can it complete the task repeatedly?
AGI, however we eventually define it, seems unlikely to be very useful if its intelligence only appears on its best attempt.
Knowing when you’re wrong matters too
There is another human capability that gets less attention than benchmark performance. Knowing when you don’t know.
A system can become more accurate overall while still remaining poorly calibrated about individual answers. That creates an unusual problem. The AI looks increasingly capable, so we are more inclined to trust it. But when it does make a mistake, it may still present that mistake confidently.
For low-stakes tasks, that is irritating. For medicine, law, scientific research, financial decisions or autonomous systems, it is much more serious.
So one of the signs of genuinely more mature AI may not be another spectacular capability. It may be something much less dramatic: the system gets better at recognising the limits of its own knowledge.
What about doing real work?
This is where AGI testing leaves the laboratory. If AI is becoming generally capable, we should eventually see that reflected in what it can accomplish in the real world.
Not just coding exercises. Not just benchmark questions. Not just neatly specified tasks with everything required supplied in advance. Actual work.
Work where the instructions are incomplete. Where priorities change. Where information is missing. Where somebody else doesn’t reply. Where the first approach fails. Where something unexpected happens halfway through. Where success requires knowing what to do next rather than merely answering the question in front of you.
That is a much harder test. And it may ultimately be far more revealing than another percentage on a benchmark.
So what would convince us?
There probably won’t be one score. Instead, we would expect several things to start happening together.
No single one of these would prove AGI. Together, they would make the argument much stronger.
The benchmark problem isn’t going away
There is an irony here. The better AI becomes, the harder it may become to test.
Every fixed benchmark eventually risks becoming part of the environment models are developed against. Developers know the test exists. Researchers investigate why systems fail it. New techniques are created. Scores rise. Then we need a new test.
That doesn’t mean the progress is fake. A system going from unable to solve a class of problem to routinely solving it is still progress. But passing a benchmark and possessing the underlying general capability aren’t automatically the same thing.
That is why the next generation of AI evaluation increasingly needs to look beyond static question sets. Unfamiliar environments. Long-running tasks. Learning from experience. Open problems. Real-world work. Independent replication.
Those are harder to reduce to one headline number. They are also much closer to the thing we are actually trying to understand.
So, would we recognise AGI if we saw it?
Probably not from one benchmark. And perhaps not immediately.
We may instead see several boundaries moving at once. AI that can work for longer. Learn more effectively. Recover from mistakes. Handle unfamiliar situations. Know when it is uncertain. Make discoveries that weren’t already available somewhere in human knowledge. And do all of that reliably rather than occasionally.
At some point, the argument may stop being about whether a particular benchmark has been beaten. It will become increasingly difficult to identify important areas of cognitive work where humans retain an obvious general advantage.
Whether we call that point AGI will still depend on the definition. But the evidence will be much harder to argue with.
The real test of AGI probably won’t be whether AI can pass one of ours. It will be whether we keep having to invent new tests because the old distinctions between what humans and machines can do no longer hold.
Seven definitions of AGI, and what each one would require →