Narrow AI

The spam filter in your inbox, the fraud model at your bank, the route optimizer in a delivery app, the voice assistant that sets timers, and the model that spots tumors in scans are all narrow AI. Each does one thing, often better than people, and nothing else. Move it outside its task and it fails, sometimes silently.

Large language models blur the line. One model can draft an email, write code, summarize a contract, and explain a spreadsheet formula, which feels general. But it is still a system trained to predict text, with no goals of its own, no memory between sessions unless one is added, and no ability to learn a new skill on the job. It is better described as a very broad narrow AI than as general intelligence.

The distinction is useful because it sets expectations. Narrow systems are reliable inside their lane and unreliable outside it. A language model that writes beautifully may add numbers wrong. A vision model trained on daytime photos may fail at night. Knowing the lane, and testing at its edges, is most of the work of using AI well.

Example at work

A retail manager uses a demand-forecasting tool that has been accurate for two years. A new product category launches, and the forecasts are badly off for months. The tool was trained on the old categories and has no basis for the new one. She overrides it with manual estimates until enough history accumulates.

Why it matters

Understanding that every tool you use is narrow keeps you from over-trusting a system just because it is impressive at one thing. Ask what a tool was built for, test it on your edge cases, and keep a person responsible for the situations it was never designed to handle.