- Narrow AI does one thing. Brilliantly, often superhumanly — and nothing else. Every AI system in commercial use has been this.
- AGI does anything a person can do. Not one task well; any task adequately, including tasks nobody prepared it for.
- ASI does everything better than every person. Not "very smart" — beyond the best human at essentially all cognitive work.
- We have narrow AI. We do not have AGI. ASI does not exist. That is the uncontroversial version.
- The contested part: today's large models are so broad that "narrow" fits them badly — which is exactly why the AGI argument is so heated.
Three acronyms, used loosely, describing things separated by enormous distances. ANI, AGI, ASI — or narrow, general, super.
Getting them straight takes about five minutes and permanently improves your ability to read AI news, because a great many confusing headlines are simply someone using one word while meaning another.
The three levels at a glance
| Level | In one line | Everyday comparison | Do we have it? |
|---|---|---|---|
| Narrow AI (ANI, "weak AI") |
Superb at one defined task, useless outside it | A calculator. Unbeatable at arithmetic, cannot plan your week. | Yes — decades of it |
| AGI (general, "strong AI") |
Handles essentially any mental task a person can, including unfamiliar ones | A capable colleague. Worse than the calculator at sums; can be handed anything. | No — disputed, but no |
| ASI (superintelligence) |
Substantially better than the best humans at essentially everything | No honest comparison exists. That is rather the point. | No — entirely hypothetical |
The single word that separates the first two is general. Not better. Not faster. General — meaning it copes with the unfamiliar.
Narrow AI: one thing, extremely well
Narrow AI is built for a specific task and cannot do anything else.
The important and counter-intuitive part: narrow does not mean weak. Narrow AI is routinely superhuman. A chess engine destroys every human who has ever lived. A spam filter classifies faster than any person could. Neither can do anything but its one job.
Examples you have used today, almost certainly:
- Spam filtering in your email
- The route your maps app picked
- Recommendations on any streaming service
- Face unlock on your phone
- Fraud detection on your card
Every one is superhuman within its lane and helpless one inch outside it. Ask the fraud model to write a birthday message and there is no answer — not a bad answer, no answer at all. The capability does not exist.
This is the category every commercially deployed AI system has belonged to. Whether it still describes today's large language models is the argument in section five.
AGI: anything a person can do
AGI means an artificial intelligence that can handle essentially any mental task a human can — including ones it was never specifically built for.
The emphasis belongs on that last clause. Generality is not a long list of skills. It is the ability to face something genuinely unfamiliar and work out an approach.
A calculator beats you at arithmetic every time. You cannot ask it to draft an email, plan a trip, or work out why a customer is upset.
A competent colleague loses to the calculator at arithmetic. But you can hand them almost anything — a spreadsheet, a difficult client, a problem they have never seen — and they will figure something out.
That flexibility is what "general" means. Not brilliance at one thing. Adequacy at nearly everything, including the unfamiliar.
By that standard we do not have AGI. Today's systems are extraordinarily broad — they will attempt almost any request — but breadth is not the same as reliability, and they fail in ways a person would not, including being confidently wrong about things they appear to know.
I have written the full version of this argument, including the three competing definitions and why the same model can count as AGI under one and clearly not under another, in what AGI actually means.
ASI: better than all of us, at all of it
Artificial superintelligence means a system substantially more capable than the best humans at essentially every cognitive task.
Not "smarter than average". Not "smarter than you". Better than the best specialist alive, in every field simultaneously — better than the finest mathematician at mathematics while being better than the finest biologist at biology.
Three things worth being precise about:
- It does not exist, and nothing resembling it exists. ASI is entirely hypothetical. Anyone claiming otherwise is either confused or selling something.
- It is not "AGI but faster". Faster human-level thinking would be transformative on its own, but ASI implies qualitatively better reasoning — solutions people would not reach with unlimited time.
- It is the level the serious safety debate is about. Concerns about controllability are largely about this step, not about current systems.
Honest caveat: because nothing like ASI exists, everything written about it — including this — is reasoning about a hypothetical. Treat confident detail about superintelligence with the scepticism you would give confident detail about any unbuilt thing.
So where does today's AI actually sit?
Formally: narrow. Honestly: the word has started to strain.
The classic definition puts every current system in the narrow bucket, and by the strict test that is right — no model today reliably handles arbitrary unfamiliar tasks.
But "narrow" was coined for systems like chess engines and spam filters, and it fits a modern large model badly. A model that will attempt code, poetry, translation, image description and — with GPT-6 Astra, which operates desktop software directly — even drive the applications on your screen. That is not narrow in the way a fraud detector is narrow.
Today's frontier models are broad but unreliable, rather than narrow.
They attempt almost anything, which looks general. They fail unpredictably on subsets of it, which is not. Astra reports 72.6% on OSWorld 2.0 for operating a computer — three tasks in four complete, one does not.
A colleague who silently botched a quarter of their work would not be described as generally capable. That gap between breadth and reliability is where the whole AGI argument lives.
So the truthful answer to "is ChatGPT narrow AI?" is: yes by the textbook definition, and the textbook definition is showing its age.
Why the boundaries are genuinely blurry
Three reasons the levels resist clean lines — worth knowing so you are not surprised when experts disagree.
- There is no agreed test. No regulator, no threshold, no scoreboard that flips from "not AGI" to "AGI". The ARC Prize team built the best-known AGI-branded benchmark and then explicitly refused to call passing it a threshold.
- Capability is jagged, not a single number. Current models beat most humans at some professional tasks and fail at things a child manages. "Human level" assumes one dial when there are hundreds.
- The definitions do different jobs. A capability definition asks what it can do; an economic definition asks what work it can replace. Under the economic one, the argument for being close is far stronger — which is exactly the definition companies reach for when making announcements.
The jump problem: why AGI to ASI might be fast
Here is the part that makes the hierarchy more than trivia.
The distance from narrow AI to AGI is a hard research problem that has taken decades and is unsolved. You might assume AGI to ASI is a similarly long road.
It might not be. The argument runs: a system that can do any human mental task can also do AI research. Which means it can improve itself. Which means the next version arrives faster, and improves itself faster again.
Whether that actually happens is genuinely open. The compressing step might be blocked by compute limits, by physical experiments that take real-world time, or by diminishing returns nobody has hit yet.
But it explains something otherwise puzzling: why safety researchers treat AGI as the important threshold rather than ASI. If the second step is fast, the last moment you have meaningful control is at the first one.
FAQ
What is the difference between AGI and ASI?
AGI matches human capability across essentially all mental tasks — as flexible as a competent person, including on unfamiliar problems. ASI substantially exceeds the best humans at essentially every cognitive task simultaneously. AGI is a comparison to people in general; ASI is a comparison to the best specialists alive, in every field at once. Neither currently exists.
Is ChatGPT narrow AI or AGI?
Narrow AI by the textbook definition, though the term fits awkwardly. Large language models attempt an enormous range of tasks, unlike classic narrow systems such as spam filters that do exactly one thing. But they fail unpredictably in ways a person would not, so they are better described as broad but unreliable rather than general. GPT-6 Astra reports 72.6% on the OSWorld 2.0 computer-use benchmark, meaning roughly one task in four still fails.
What is an example of narrow AI?
Spam filtering, route planning in maps, streaming recommendations, face unlock, and card fraud detection — all systems most people use daily. Each is superhuman at its single task and completely incapable outside it, which is the defining feature: narrow does not mean weak, it means specialised.
Does artificial superintelligence exist?
No. ASI is entirely hypothetical, and nothing resembling it has been built. It is also not simply a faster version of human-level intelligence — it implies qualitatively better reasoning, reaching solutions people would not find given unlimited time. Any confident description of ASI is reasoning about something unbuilt.
Why do experts disagree about whether we have AGI?
Because there is no agreed test, capability is jagged rather than a single level, and the competing definitions answer different questions. Under a capability definition we clearly do not have AGI; under an economic definition — can it do economically valuable work — the argument is far stronger. Announcements tend to use the economic definition without saying so.
Will AGI lead to ASI quickly?
Possibly, and that possibility is why AGI is treated as the critical threshold. A system able to perform any human mental task could also perform AI research, and therefore improve itself — each version arriving faster than the last. Whether that compounding actually occurs is unresolved; it could be limited by compute, by experiments that take real-world time, or by diminishing returns.
- Jayant Solanki, what AGI actually means, and who gets to decide — the three competing definitions in full
- Jayant Solanki, when AGI arrives: every forecast and who makes it
- Jayant Solanki, GPT-6 Astra: what shipped and what real tests show — the OSWorld 2.0 figure
Jayant Solanki
AI-Ready SEO, GEO & AIO strategist based in Indore, India. I write explainers that define their terms in plain words, so you can read the next AI headline without needing anyone to translate it.
More about Jayant →