The only form of AI that currently exists. Systems designed and trained for a specific, limited task — such as playing chess, recognizing faces, or translating text — with no ability to generalize beyond that domain.
In Depth
Narrow AI, also called Weak AI, is any AI system designed to perform one specific task — and only that task. Despite the word 'weak,' these systems can be extraordinarily powerful within their domain. AlphaGo can defeat any human at Go but cannot play checkers. GPT-4 can generate fluent prose but cannot drive a car. Every AI product in use today falls into this category.
The 'narrowness' refers to generalization, not capability. A Narrow AI system learns a statistical mapping from inputs to outputs within a well-defined problem space. If the problem space shifts even slightly, performance degrades. This is why a face-recognition model trained on one demographic can fail on another, and why a model fine-tuned for legal documents may produce nonsense when asked about cooking.
Understanding Narrow AI is foundational to understanding the hype cycle around AI. Many claims about AI 'thinking' or 'understanding' describe systems that are, in practice, sophisticated pattern matchers within a constrained domain. Recognizing these constraints helps set realistic expectations about what current AI can and cannot do.
Every AI system deployed in the real world today is Narrow AI — extraordinarily capable within its specific domain, but unable to transfer that capability to anything outside it.