narrow AI
/ NAR-oh AY-EYE /
Narrow AI is artificial intelligence that is brilliant at one specific job and useless outside it. Think of a star athlete who is the best sprinter in the world but has never learned to swim, cook, or do arithmetic. A program that beats grandmasters at chess, a model that transcribes speech, an app that flags tumors in a scan — each is a champion in its single lane and has no idea anything else exists.
More precisely, narrow AI is built and trained for a bounded task (or a small family of related tasks) and has no understanding, goals, or competence beyond it. Even today's most impressive systems, including chatbots that seem to talk about anything, are narrow in this sense: they are doing one thing — predicting plausible text — extremely well, not thinking across domains the way a person does. The system has no inner life and does not 'know' what it is doing; it has simply been shaped by examples to behave a certain way within its lane.
This matters because almost every AI you will ever meet is narrow AI. It is the opposite of the science-fiction idea of a machine that can do anything a human can. Recognizing the difference keeps expectations honest: a tool that is superhuman at one task can fail in childishly simple ways the moment you nudge it outside its training. Narrow does not mean weak — narrow AI can be wildly more capable than any human at its specialty. It just cannot transfer that skill to a new kind of problem on its own.
A model trained to read chest X-rays can outperform radiologists at spotting pneumonia — yet hand it an X-ray of a hand, or ask it whether the patient should be hospitalized, and it has nothing to offer. Its skill does not stretch one millimeter past the exact task it was trained on.
Superhuman in one lane, blank everywhere else — the signature of narrow AI.
Every AI in commercial use today is narrow AI; there is no exception. A system feeling 'general' (like a chatbot) is not the same as being general — fluency across topics is still one narrow skill (producing plausible language), not genuine cross-domain understanding.