In plain words
Your phone recognises your face in a moment. A maps app predicts traffic. Your bank flags an odd payment. Each of these is a separate AI system trained for one job.
That is narrow AI, sometimes called weak AI. It can beat the best humans at its task: a chess engine will win against any grandmaster. But ask it anything outside that task and it has nothing. The chess engine cannot tell you the time.
Almost all AI that exists is narrow. It is the opposite of artificial general intelligence, the still-hypothetical idea of one system that can learn almost anything.
Modern chatbots blur the line. One language model can write, translate, summarise and code. Many researchers still see them as short of general, because their abilities are uneven and limited to what training covered.
Everyday example
A railway ticket machine
The ticket machine at the station is excellent at its job: pick a destination, pay, get a ticket, faster than any clerk. Ask it which platform the Rajdhani leaves from, or whether it will rain, and it has no idea. It was never built for that. Narrow AI is like this machine: superb in its lane, blank outside it.
Where the example breaks: A ticket machine follows fixed rules. Narrow AI usually learns its skill from data, so within its task it can handle situations nobody programmed in advance.
How it actually works
Narrow is about the breadth of an ability, not about how impressive it is.
One task, one model
Most AI in use is trained on data from a single task, such as photos labelled with faces or past payments marked as fraud. The model learns patterns that only make sense for that task.
Superhuman inside a box
Within its task, a narrow system can beat any person. IBM's Deep Blue beat world chess champion Garry Kasparov in May 1997, evaluating 200 million positions per second, and could do nothing else.
Why it is brittle
Outside the conditions it trained on, performance can drop sharply. A road-sign reader trained on clear days may fail in fog. That is why narrow systems are tested and monitored for their exact use.
"Weak AI" first meant something else
Philosopher John Searle's 1980 "weak AI" meant using computers to model the mind without claiming they truly understand. Today "weak AI" is often used to mean narrow AI, but the original was about understanding, not breadth.
- Also called
- Weak AI, artificial narrow intelligence (ANI)
- Famous example
- Deep Blue beat Garry Kasparov 3.5 to 2.5 in May 1997
- Everyday examples
- Face recognition, spam filters, voice assistants, fraud checks, recommendations
- Opposite
- AGI, still hypothetical as of October 2026
Questions people ask
What are examples of narrow AI?
Face recognition on phones, spam filters, recommendations on streaming and shopping apps, voice typing, translation, fraud checks on card payments, chess engines and navigation apps that predict traffic. Each is trained for one kind of task.
Is ChatGPT narrow AI?
What is the difference between weak AI and strong AI?
In philosopher John Searle's 1980 terms, strong AI claims a suitably programmed computer truly understands and has a mind; weak AI only claims computers are useful tools for studying the mind. Today the words are often used loosely for narrow versus general AI.
Is narrow AI less risky than general AI?
Not automatically. Narrow systems already cause real harm when they fail or are misused, such as biased decisions, wrong face matches or deepfakes. The risks are different in kind, not absent.