In 1980, a philosopher named John Searle published a paper called Minds, Brains, and Programs. Buried inside it was a distinction that has since been stretched, misquoted and repackaged by marketing departments around the world: the split between strong AI and weak AI.
Searle meant something narrow and technical. “Strong AI” was the claim that a correctly programmed computer doesn’t merely simulate understanding, it actually has a mind. “Weak AI” was the modest position: the machine is a model, a lab instrument for studying cognition, nothing more. Forty-five years later, almost nobody uses the terms that way. In everyday tech writing, strong AI vs weak AI usually means human-level general intelligence versus single-purpose tools.
Both readings are worth understanding, because the muddle between them causes real mistakes: bad product expectations, wasted budgets, and a lot of breathless headlines about machines that are nowhere near as capable as the copy suggests.
What the two terms mean in practice
Strong AI describes a system that can handle essentially any intellectual task a person can. It would transfer knowledge between domains, pick up new skills without being retrained, and reason sensibly about situations it was never prepared for. Nobody has built one. There is no agreed benchmark for one, and no consensus on how you would know if you had.
Weak AI, also called narrow AI, is built for a defined task. It can beat a grandmaster at chess and still be unable to tell you what a chess piece is. It can transcribe your voice with startling accuracy and have no idea what the words mean. Nearly everything sold as AI today lives in this bucket.
That includes the flashy stuff. When a generative model writes a passable paragraph of marketing copy, it is doing a narrow task very well, not demonstrating general intelligence. The line between generative and general systems gets blurred constantly, which is why it helps to read up on the difference between AGI and generative AI before you start assuming one implies the other.
Weak AI is already threaded through your day
Start counting and it gets faintly ridiculous. Your spam filter. The route your maps app picks at 8:15am. Autocomplete on your phone keyboard. The fraud check that freezes your card in Lisbon because you were in Leeds yesterday. Shazam. Face unlock. None of these systems knows what a car or a face is. Each one is extremely good at a very thin slice of pattern matching.
Creative tools follow the same pattern. An image generator like Fooocus, for instance, produces genuinely useful pictures from a text prompt, and it will also cheerfully mangle hands and spelling because it has no model of anatomy or language. It’s narrow, and the failure modes give it away.
The strength of these systems is that they’re cheap, fast, and often superhuman inside their lane. AlphaFold’s protein structure predictions beat decades of lab work in some cases. A chess engine has not lost to a human world champion in years. Narrow does not mean weak in the everyday sense of the word.
A three-question test for any system you meet
Forget the marketing copy. Ask these instead:
- Can it handle a task outside its training distribution? A strong system would improvise. A narrow one usually fails quietly or confidently invents something.
- Does it carry context across domains? If a model can help you debug code and then apply the same reasoning to a legal contract, that’s a signal. If it starts over each time, it isn’t.
- Can it notice it’s failing and change approach? Not retry the same broken output, but genuinely re-plan. Very few systems do this reliably.
Three noes means you’re looking at narrow AI. That’s not an insult. It’s a spec sheet.
Why “weak” is a terrible name
No vendor on earth puts “weak AI” on a landing page. The word carries a whiff of failure, so the industry quietly dropped it and started saying “AI” for everything, which is how a scheduling widget and a frontier research model end up in the same category on a conference slide.
The label also flatters the other side. “Strong AI” sounds like something just around the corner, when the honest description is a research programme with no timeline and no agreed test. Plenty of researchers think the framing is unhelpful altogether, since general and narrow capability exist on a spectrum rather than two shelves.
Where agentic tools actually sit
AI agents complicate the picture. An agent chains many narrow capabilities together and runs them in a loop: read the request, plan, call a tool, check the result, try again. That looks a lot like generality from the outside, and sometimes it even works.
But you’ve probably noticed the wobble. The same agent that nails a complex task on Tuesday produces a mess on Wednesday with identical inputs, because the prompt path drifted somewhere in step four. Consistency is the giveaway. Evaluating whether a system repeats its success is a much better use of your attention than admiring a single good demo, which is the argument in why repeatability separates real agents from lucky ones. A system that succeeds once proves very little; one that succeeds fifty times in a row is telling you something.
What this means if you’re buying or building
On the buying side, ask two questions before you sign anything: what exact task is this system trained for, and what does it do when it hits something outside that boundary? Tools that are honest about their edges tend to be the ones you keep. A diagram generator such as Napkin AI is a decent example of a narrow tool with a clear job. It takes your text and returns a diagram, quickly. It will not write your strategy or judge whether the diagram makes sense, and expecting it to is how disappointment gets manufactured.
On the building side, the practical skill is no longer writing AI from scratch. It’s understanding what these models do internally well enough to design around their limits. Working through a hands-on course like Fast.ai’s free deep learning curriculum gives you that grounding faster than most paid alternatives, and it’ll make the strong versus weak distinction feel obvious rather than abstract.
The part of Searle’s argument that still stings
Here’s the uncomfortable bit. Even if someone built a system that passed every behavioural test you could throw at it, Searle’s original question would remain open. In his Chinese Room thought experiment, a person who doesn’t speak Chinese follows rules to produce flawless Chinese replies. The output looks like comprehension. Inside, there’s rule-following.
You can scale that room up to a data centre and the question doesn’t dissolve. Fluent output is not evidence of understanding, and no amount of capability settles what’s happening in the machinery. That’s the real reason strong AI vs weak AI is worth learning properly: the conversation isn’t only about what a system can do. It’s about what you’re entitled to assume it’s doing. Most of the industry’s confident claims, in both directions, would survive a lot better if more people kept that question in view.