Narrow AI, sometimes called weak AI, is designed for a defined problem: ranking a feed, detecting a tumour, transcribing speech. It can exceed human performance inside that boundary and has no competence outside it. General AI aims at the boundary itself, carrying its own ability to acquire new skills.
How to tell them apart
Give the system a task from a different domain and change the format of the input. A narrow system degrades sharply, often without recognising that it has. A general system notices the mismatch and adapts. The second behaviour is the hard one.
Why nearly everything today is narrow
Because specialisation is how you get reliability. A system tuned for one distribution can be measured, audited and trusted. Generality trades that certainty for range, and the trade is not yet favourable for work with real consequences.
Why the line is blurry
Large multimodal models are narrow in mechanism and broad in apparent function, which makes the categories feel outdated. The distinction still holds where it matters: on the first problem nobody anticipated.