Ask ten people where the line sits between AI and AGI and you’ll get ten answers. Most will be confident. Several will contradict each other. That happens because “AI” has quietly turned into a product category while “AGI” has turned into a prophecy, and the two words are now doing very different jobs in the same sentence.
Here’s the version that holds up under scrutiny.
AI is a category. AGI is a claim about a ceiling.
Artificial intelligence covers any system that does something we’d call intelligent if a person did it. A spam filter qualifies. So does a route-planning algorithm, a chess engine, and the model drafting your emails. The bar is deliberately low because the category is deliberately wide.
Artificial general intelligence describes something far narrower in scope and vastly bigger in ambition: a system that can pick up and perform essentially any cognitive task a human can, including ones it was never trained for, and improve at them without a team of engineers rebuilding it from scratch.
The defining variable isn’t how smart the system seems. It’s breadth plus adaptability. Deep Blue beat Garry Kasparov in 1997 and could not tell you what a pawn is.
The three-tier framing that actually helps
Narrow AI
One task, occasionally superhuman. Fraud detection, wake-word spotting, tumour segmentation on a scan. These systems don’t generalise and don’t need to. Most AI you interact with in a given day lives here, and it works precisely because the problem is bounded.
General-purpose, but not general
Frontier language models sit in this middle band, and it’s where nearly all the confusion lives. A current top-tier model can draft a contract clause, debug Python, and explain a cardiology paper inside one session. That feels general.
Then you notice it was trained once, frozen, and shipped. It doesn’t learn from your corrections in any lasting way. Close the chat and everything it absorbed about your project evaporates. Wrapping a model in a workflow is now cheap, too. Open-source platforms like Dify make building AI apps an afternoon’s work, which moves the bottleneck from access to judgment.
AGI
The hypothetical end state. It would need to acquire new skills on its own, transfer them to unfamiliar problems, and hold a goal across days rather than turns. Nobody has shipped one. Several labs claim to be close, and the word “close” is doing heavy lifting in those claims.
Where the two diverge in concrete terms
- Scope. Narrow AI operates inside a domain boundary. AGI, by definition, wouldn’t have one.
- Learning. Today’s models are trained, then frozen. Genuine AGI implies continual learning that doesn’t wipe out earlier skills, the technical problem known as catastrophic forgetting.
- Transfer. A vision model trained on daylight photos degrades on night shots. A general system shouldn’t collapse the moment the input shifts.
- Horizon. Chat turns last minutes. AGI claims usually involve multi-day autonomous work.
- Error profile. 92% accuracy is excellent for a suggestion engine and unacceptable for a system acting on its own.
The label is doing marketing work
Ask five labs to define AGI and you get five definitions, each suspiciously aligned with whatever that lab is currently building. One camp measures it in economic output: systems that can do most remote work. Another measures it against cognitive benchmarks. A third insists on embodiment, or autonomy, or something closer to consciousness.
That disagreement isn’t academic nitpicking. It’s the reason credible timelines range from two years to never, and why a useful explainer on what artificial general intelligence really means ends up being mostly about the arguments rather than the answer.
The goalposts move, too. Tasks once treated as AGI-hard in 2019 — fluent essay writing, passing a bar exam, code that compiles on the first try — are routine now, so they get quietly reclassified as “just narrow AI.” Every time a benchmark saturates, a new one appears.
Benchmarks trying to put a number on the gap
ARC-AGI tests novel puzzle solving from a handful of examples. Humanity’s Last Exam and GPQA push expert-level reasoning. SWE-bench measures whether a model can fix real bugs in real repositories. The interesting signal isn’t a headline score. It’s the shape of the curve: models climb fast on problems with abundant training data and crawl on problems that require genuine on-the-fly reasoning.
Announcements like OpenAI’s claimed generational leap with GPT-6 Astra get read as AGI progress when what’s usually being described is a bigger, better narrow-plus system with a wider surface area.
Why the distinction matters when you’re buying software
Almost everything you can pay for today is narrow AI, and that isn’t an insult. The real question is whether the tool is excellent at its one job.
A writing assistant like HyperWrite’s free tier is genuinely handy for rewriting a paragraph or drafting a product description. It will not set your pricing strategy. Napkin AI turns a block of text into a diagram in seconds. One trick, and a good one.
Where people go wrong is mismatching expectations in both directions. Judging a narrow tool by general standards (“it can’t reason about my whole business”) produces disappointment. Judging a frontier model by narrow standards (“it got one fact wrong, so it’s useless”) throws away real value. Calibrate to the category you’re actually holding.
A test you can run in any conversation
If someone tells you a system is “basically AGI,” ask three things. What did it fail at last week? Confident claims rarely come with failure logs. What happens after you correct it? If the fix doesn’t persist, you’re looking at stateless narrow AI with a wide surface. And can it handle a task it has never seen, with a tool it doesn’t know, and zero examples? Watch what breaks first, and in what order.
What would actually convince me
Forget the demos. Four signals would move this from speculation to something worth planning around.
- Continual learning that sticks. A model learns a new internal API once and still remembers it a month later, without fine-tuning and without losing the ability to write.
- Sample efficiency near human level. A person learns a board game from the rulebook. Current systems need thousands of games.
- Multi-day autonomy. Not a forty-step agent that stalls, but something that works for three days, notices it has drifted, and corrects course.
- Honest transfer. Performance that holds when the format changes and the domain shifts.
Until a couple of those land, the practical answer stays boring and genuinely useful: AI is a set of sharp tools you deploy on purpose, and AGI is a research bet you read about. Confusing the first for the second wastes money. Assuming the second has already arrived wastes attention.