Glossary
What is narrow AI (weak AI)?
In short
Narrow AI, also called weak AI, is a system that solves one task or a narrow range of tasks and does not carry its skills over to other areas.
AI basics · narrow AI
MMOOZZGG.com editorsChecked: 2 min read
For a child
You have a remote-control car: it drives very well, but it cannot sing or draw. An alarm clock wakes you up on time, but it will not tell you a story. Each thing is good at its own job. Almost all smart programs around us are like that: one recognizes a voice, another picks cartoons, a third translates words from a foreign language. People call them weak AI.
In plain words
Weak, or narrow, AI is AI trained for one specific job. A program reads car number plates in a parking lot but will not understand a question about the weather. The translator on your phone translates phrases but does not play chess. “Weak” does not mean “bad”: in its own task such a system can work more precisely and faster than a person. Its weakness is elsewhere: it does not learn a new job by itself, and it gets lost when conditions differ a lot from those it was trained on. Almost all AI that exists today belongs to this class.
For business
Companies buy narrow systems: document recognition, demand forecasts, call analysis, answers to typical questions. They are easier to check because the task has a clear measure of success: the share of fields read correctly, the forecast error, the response time. The price depends on whether a ready solution exists on the market or the model must be trained on your data. The main risk is transfer: a system trained on one branch or season makes mistakes on another. So you measure quality on fresh data from the site where it will run, and only then expand. A pilot with one task and one metric is a good way to practice this order.
For a specialist
The term goes back to John Searle’s paper “Minds, Brains, and Programs” (Behavioral and Brain Sciences, 1980). Searle separated weak, or cautious, AI, for which the computer is a powerful tool for studying the mind, from strong AI, for which a correctly programmed computer itself has cognitive states. In engineering practice another pair became common later: narrow AI, which solves a predefined class of tasks, and artificial general intelligence (AGI). Morris and co-authors (ICML 2024) rate systems on depth of performance and on breadth (generality) separately, so a narrow system can beat humans at one task and still stay narrow. Large language models blur the line: they solve many different tasks without retraining, and the debate about whether to call them narrow goes on.
Example
A chain of clinics introduces a program that finds signs of a fracture on X-ray images. It helps the doctor go through a stream of images, but it will not fill in the patient’s record or answer a call: each of those tasks needs its own system.
Related terms
Questions and answers
How is weak AI different from strong AI?
Is ChatGPT weak AI?
Sources
Where companies use it
Not sure whether your company needs this?
AI solutions map