A spam filter flags a phishing email in about 40 milliseconds. A large language model writes a passable sonnet, then confidently invents a court case that never existed. Neither system could swap jobs without being rebuilt from scratch.
That trade is the whole narrow AI vs general AI argument in miniature. One is astonishingly good at a bounded task and useless one step outside it. The other is a system nobody has built yet, and might not for a long time.
Narrow AI is everything you have actually used
Narrow AI, sometimes called weak AI, is software trained to do one job or a tight cluster of jobs against a fixed objective. It can beat humans inside that boundary and fail absurdly just outside it.
The examples are everywhere once you start looking:
- Recommendation engines. YouTube has said recommendations drive roughly 70% of what people watch. The model predicts one thing: which thumbnail-and-title pairing you will click next.
- Machine translation. Google Translate covers more than 200 languages and moves over 100 billion translated words a day. It has no working idea of what a word is; it maps patterns between languages.
- Protein structure prediction. AlphaFold produced predicted structures for around 200 million proteins, a job that would have kept experimental labs busy for decades.
- Fraud detection. Payment networks score transactions in milliseconds at peak volumes above 65,000 messages a second, flagging the ones that look wrong.
What unites them is architecture, not cleverness. Each one is trained on a dataset, tuned against a metric and deployed to repeat that judgement. None of them can reframe the problem. A fraud model cannot decide it would rather work on climate modelling. If you want the machinery underneath explained properly, this plain-English guide to how artificial intelligence actually works is a good place to start.
General AI is a target, not a product
Artificial general intelligence describes a system that can learn and perform any intellectual task a human can, including new ones it was never trained for. There is no agreed definition, no standard test and no working example.
Ask five researchers what counts and you get five answers. DeepMind has framed it around human-level performance on most economically valuable work. OpenAI’s charter talks about highly autonomous systems that outperform humans at most economically valuable work. Others insist on embodiment, or curiosity, or the capacity to set its own goals.
One common mix-up is worth clearing up: general-purpose AI in regulation is not AGI. The EU AI Act’s general-purpose AI rules cover foundation models such as GPT-4 or Claude, which are flexible but still narrow. A model can answer questions about contract law, organic chemistry and 1990s sitcoms without being general, because it still cannot learn on the job.
The goalposts keep moving
Deep Blue beat Kasparov in 1997. AlphaGo beat Lee Sedol in 2016. Each time the reaction was identical: fine, but it cannot do the thing I actually care about. That is Moravec’s paradox playing out in public. Tasks that feel trivial to a person, like catching a ball or reading a room, turn out to be brutal to formalise.
There is a serious point buried in the scepticism. Building a system that reasons across unfamiliar problems is a different project from building one that dominates a single game or dataset. The cognitive AI systems that reason through a problem instead of just retrieving an answer are where much of that work now happens.
Three things narrow AI genuinely cannot do
Learn without being retrained
A deployed model’s weights are frozen. You can paste a new fact into a 200,000-token context window, but the model itself has learned nothing durable. Retrain it on fresh data and it may lose older skills; catastrophic forgetting is still unsolved at scale.
Carry a skill into a new domain
AlphaGo’s architecture does not drive a car. A model trained to spot tumours in chest X-rays will not spot corrosion on a pipeline, even though both tasks reduce to “find the anomaly in an image”. People do that transfer constantly, and cheaply.
Know where the edge of its knowledge is
Ask a narrow model something outside its training distribution and it does not shrug. It answers, fluently and wrongly. Calibration, the ability to say “I don’t know, and here is my confidence level”, is one of the cleanest dividing lines between today’s systems and anything you would call general.
The data gap makes the point. A child hears somewhere between 10 and 15 million words a year and ends up fluent. Frontier models train on more than 10 trillion tokens and still need reminding not to invent citations.
Why the line feels blurrier than it did in 2020
Modern chatbots pass bar exams, write working code and hold a conversation for hours. That is a lot of surface area. It is also, mechanically, a very large next-token predictor with an enormous amount of training behind it.
Two developments have muddied the picture. First, benchmark saturation: MMLU, a long-standing test of broad knowledge, now sits near the top of its range, so researchers have moved to harder tests such as GPQA, FrontierMath and ARC-AGI, where humans still lead comfortably. Second, reasoning models that spend extra compute thinking before answering have posted real gains on multi-step problems. OpenAI’s o3 reportedly hit 87.5% on ARC-AGI-1 in late 2024, though at a cost per task that would sink most products.
Progress like that is why serious people have pulled AGI timelines in from “centuries” to “decades”, and why others have left the field’s biggest labs to chase it directly. The story of how Ilya Sutskever left OpenAI to pursue superintelligence is the clearest example.
Where the distinction actually bites
Outside research papers, “general” is mostly a marketing word, and it costs people money.
- Vendor claims. A tool that summarises contracts is narrow. When a sales deck implies it will also draft policy, price risk or handle HR calls, ask what it was trained on and what it does when a request falls outside that.
- Regulation. Two systems that both answer questions in natural language can sit under completely different obligations depending on capability and use. Capability claims drive compliance scope.
- Risk assessment. A narrow AI failing recommends a bad film. A general system failing in the same confident tone could approve a loan, a diagnosis or a hire. Blast radius scales with generality, not with eloquence.
- Build decisions. Most “we need AI” briefs are really “we need a classifier” or “we need search over our own documents”. Framing them as AGI projects burns budget on scope nobody asked for.
For a grounded view of what current systems handle well and where they still break, this piece on what AI can do, what it can’t, and where it’s actually going is worth your time.
Five questions that separate the two in practice
Before calling any system general, run these:
- Can it learn from a single example and still hold that learning tomorrow?
- Can it take a skill from one field and apply it in an unrelated one, unprompted?
- Does it know when it doesn’t know, and say so?
- Can it set its own sub-goals across a multi-day task without a human steering every step?
- Can it improve itself without a fresh training run?
Current systems fail at least three of those, usually all five. That is not a knock on them. A well-built narrow system beats a general one on cost, accuracy and auditability for almost every job a business actually needs done today. The mistake is buying the story instead of the tool.
So the useful question is not “is this AGI yet”. It is “which one am I looking at, and does that match the job I need it for”.