AGI Made Clear › AGI vs narrow AI
AGI vs narrow AI: the distinction isn’t as simple as it used to be
Narrow AI used to mean a system built to do one kind of thing. AGI meant a system that could do many. Today’s best AI systems have made that distinction much harder to use.
The traditional explanation of narrow AI is easy. A chess program plays chess. A recommendation system recommends products. A medical imaging system looks for particular patterns in scans.
Each can be extraordinarily good at its job, perhaps far better than a human, while being useless outside the task it was designed to perform. That is narrow AI.
AGI is supposed to be different. Instead of being built for one particular task, a generally intelligent system should be able to work across many different kinds of problems, adapt to unfamiliar situations and use what it knows in one area somewhere else.
For years, that gave us a reasonably clean dividing line. Then large language models arrived. And the line became much harder to draw.
Today’s AI doesn’t look very narrow
Open a leading AI system today and ask it to write some software. Then give it a photograph to analyse. Ask it to explain a scientific paper. Give it a spreadsheet. Ask it to research a subject. Have it translate something. Ask it to plan a project. Give it a mathematical problem. Ask it to use software or operate a computer.
None of those things proves that the system is generally intelligent. But describing the same system as “narrow AI” starts to sound strange.
It isn’t doing one task. It isn’t even operating within one obvious domain. The same underlying model can attempt an enormous range of intellectual work without a separate AI system being built for every task. That is a significant change from the narrow AI the term was originally useful for describing.
So we now have an awkward middle ground. Today’s frontier AI is clearly broader than traditional narrow AI. Whether it is broad enough, adaptable enough and reliable enough to call general is precisely what people are arguing about.
Broad isn’t necessarily general
This distinction matters. Being able to attempt lots of tasks isn’t automatically the same thing as possessing general intelligence.
A system might know an extraordinary amount about thousands of subjects because it was trained on enormous amounts of information. It might be able to apply patterns learned during that training to a huge variety of new prompts. That gives it remarkable breadth.
But general intelligence implies something more demanding. Can it deal with situations substantially different from those it has encountered before? Can it learn effectively from new experience? Can it work out what matters when the task hasn’t been neatly specified? Can it transfer what it learns in one situation into another? Can it recognise when its approach isn’t working and change it? Can it keep doing those things reliably over a long and complicated piece of work?
Those are much harder questions. And the evidence is mixed.
AI capability has a strange shape
One of the most useful ideas for understanding today’s AI is the jagged frontier. The phrase came from research involving Ethan Mollick and colleagues looking at how people performed when working with generative AI.[Mollick]
The researchers found something counterintuitive. AI wasn’t simply good at easy tasks and bad at difficult ones. It could perform extremely well on one task and then struggle with another that appeared, to a human, to be easier. Its strengths and weaknesses didn’t follow the pattern we expect from human intelligence.
That jaggedness remains visible in today’s much more capable systems. An AI can produce sophisticated code and then misunderstand an apparently straightforward instruction. It can reason through a difficult technical problem and then confidently make a basic factual mistake. It can demonstrate remarkable knowledge across dozens of fields while still struggling to learn effectively from repeated experience.
That makes it difficult to place current AI neatly on a line running from narrow to general. The shape is messier than that.
The systems are becoming broader
There is still clear evidence that the boundary is moving. ARC-AGI is one attempt to test whether AI can solve unfamiliar reasoning problems rather than simply reproduce patterns it has already encountered, and performance has risen dramatically. ARC-AGI-3 goes further by putting models into unfamiliar interactive environments where they have to explore what is happening, infer the rules and act.
In September 2026, GPT-6 Astra scored 62.7% using ARC Prize’s standard harness. With a provider-specific setup that preserved internal reasoning state between requests, it reached 99.9%.[ARC Prize]
That is remarkable progress. But it is also a perfect example of why “general” remains difficult to establish. Change the surrounding system and the score changes enormously. The model is clearly capable of doing something much more general than recognising a fixed set of patterns. But exactly where the intelligence sits, and how robustly it transfers, is still an open question.
Learning exposes one of the remaining gaps
If there is one thing we would expect from general intelligence, learning from experience seems fairly fundamental. Humans don’t need to be rebuilt every time they learn something. We encounter a situation, try something, make mistakes, remember what happened and change our behaviour.
Current AI can mimic parts of that process by storing information in its context or external memory. But evidence that models themselves are becoming substantially better at learning through repeated experience is much weaker.
In July 2026, Epoch AI tested frontier models on an unfamiliar cooperative board game over 30 playthroughs. The models were allowed repeated experience. They did not meaningfully improve. Newer models tended to start from a higher level of performance, but Epoch found no improvement across the repeated games themselves.[Epoch AI]
That distinction is important. Building a better model is not the same thing as building a model that can keep learning once it has been released. A genuinely general system might eventually need much more of the second.
So is today’s AI narrow or general?
Neither label describes the situation particularly well.
Calling GPT-6 Astra or another frontier multimodal model “narrow AI” puts it in the same conceptual bucket as a chess engine or a spam filter. That hides how much has changed.
But calling it AGI creates the opposite problem. It suggests we have already established that these systems possess the breadth, adaptability, learning ability and reliability associated with general intelligence. We haven’t.
A better description of today’s frontier AI is simply that it is becoming increasingly general-purpose. That doesn’t settle the AGI argument. It describes the direction without pretending we have agreed on the destination.
This is why the word “general” matters
The interesting part of artificial general intelligence isn’t really “artificial”. We already have artificial intelligence. It isn’t even necessarily “intelligence”, because we have spent decades arguing about what that means. The difficult word is general.
How broad does an intelligence have to become before it counts as general? Breadth of knowledge? Number of tasks? Ability to learn? Ability to adapt? Transfer between domains? Autonomy? Reliability? Human-level performance? Economic usefulness?
Different definitions of AGI put different weight on each of those things. And that is why the narrow-versus-general distinction has become much more important just as it has become much harder to draw.
There may not be a clean line
We may eventually discover that thinking about AI as two categories was the wrong model.
Perhaps there isn’t one boundary separating them. Instead, systems may become more general along several dimensions at different speeds. One model becomes extraordinarily broad in knowledge. Another becomes better at learning. Another becomes more autonomous. Another becomes more reliable. Another becomes better at transferring knowledge into unfamiliar situations.
Eventually enough of those capabilities may converge that describing the system as general becomes difficult to dispute. But that doesn’t mean there has to be one moment when narrow AI suddenly becomes AGI. The transition may be gradual.
And that possibility helps explain one of the strangest features of the current AGI debate.
Two people can look at the same AI and see something completely different
Someone focusing on breadth can look at today’s systems and say: how can this possibly still be called narrow AI? They are right that the old label fits badly.
Someone focusing on learning, robustness and adaptability can look at exactly the same system and say: how can this possibly be called AGI? They have a point too.
Those positions only appear contradictory if we assume “narrow” and “general” are the only two boxes available. The technology has moved into the space between them. And that space is getting more interesting very quickly.
Where this leaves “narrow AI”
The term still has a useful job. A chess engine is narrow AI. A system designed specifically to detect fraud is narrow AI. A recommendation algorithm optimised for one particular task is narrow AI.
The UK government’s own frontier-AI terminology, for example, still defines narrow AI as a system that performs a single task or narrow set of tasks.[GOV.UK]
That definition remains perfectly useful for those systems. It just doesn’t describe frontier general-purpose models particularly well any more.
Narrow AI hasn’t disappeared. The frontier has simply moved beyond the point where “narrow” tells us enough.
So where are we now?
We have AI systems that operate across an extraordinary range of tasks. They can reason, write, code, analyse, research and increasingly act. They can perform at very high levels in areas that once looked like useful tests of generality.
But they remain uneven. They struggle to learn from experience in ways we would expect from humans. Their performance can depend heavily on the tools and scaffolding around them. They can be astonishingly capable one moment and unexpectedly brittle the next.
That doesn’t give us a neat label. It gives us a much more useful question. Not “is this narrow AI or AGI?” but: which parts of general intelligence are already emerging, and which are still missing?
That is the question we will keep following.
How would we know if AGI had arrived? →