Definition
Narrow AI is designed, authorized, or validated for a bounded task and operating domain rather than broad transferable competence. The boundary may specify the input type, users, environment, decisions, tools, time horizon, and conditions under which the system must abstain or escalate.
"Narrow" does not mean weak. A chess program, protein-structure predictor, fraud classifier, or industrial vision system can outperform people within its task while remaining unable to transfer that competence to unrelated work. The term describes scope, not quality.
A system built on a broadly trained foundation model can still be narrow. If the application exposes only invoice extraction, validates only invoices, and grants no authority outside that workflow, its operational claim should stay inside that boundary even when the underlying model can discuss thousands of subjects.
Distinguish it from nearby terms
- Artificial general intelligence aims at broad competence and transfer across domains, though its exact threshold is contested.
- A foundation model is broadly reusable. A particular system built on it may remain narrow by design and evidence.
- A specialized model has been trained or adapted for a bounded task. Narrow AI describes the system-level scope, which can include a general-purpose model.
- Automation can be narrow without using AI, as with a fixed rule or script.
Operational significance
Narrow scope is a control only when the system enforces it. Document allowed inputs and actions, reject or route out-of-domain cases, test boundary conditions, and prevent broad model capability from silently expanding permissions. Product marketing should not substitute for an operating envelope.
Check your understanding
A medical model answers questions across specialties, but it has been evaluated only for radiology report drafting. Its authorized use is narrow even if the underlying model produces plausible answers elsewhere.