Narrow AI is an AI system designed, trained and evaluated against one class of task, with no expectation it works outside it. the United States’ NIST AI Risk Management Framework defines every AI system this way — as operating “for a given set of objectives” — which means that, by the standard Canadian frameworks cite, every deployed AI system is narrow by definition; nothing fielded today qualifies as anything broader.
Even Canada’s own guidance for the most attention-grabbing category of AI stays task-bound. Canada’s Voluntary Code of Conduct on generative AI commits developers of “a generative AI system with general-purpose capabilities” to specific risk measures, but describes what those systems actually do in task terms: it states that “These advanced systems may be used to perform many different kinds of tasks — such as writing emails, answering complex questions, generating realistic images or videos, or writing software code.” Even a system marketed as general-purpose is regulated, and described, one task at a time. The Federal Court reaches for the same distinction from the opposite direction: the Federal Court’s Notice on the use of AI in court proceedings says its AI Declaration rule “does not apply to AI that only follows pre-set instructions, including programs such as system automation, voice recognition, or document editing” — the plainest Canadian statement of what sits on the narrow side of the line.
Canada’s Directive on Automated Decision-Making’s Appendix C impact-assessment levels work the same way: a department scores risk by the specific decision a system is supporting and the specific data it uses, not by how broadly “intelligent” the underlying technique is claimed to be. A system that sorts intake documents and a system that screens loan files are assessed separately even if both run on the same underlying technique.
A computer-vision system trained to grade lumber by visual defects can outperform a person at that one task and still be narrow AI in the fullest sense: point the same system at a document scan or a customer email and it does nothing useful, because it was never trained on that objective. That is not a limitation to be fixed later — it is what “for a given set of objectives” means in practice, and it is true of every AI system a Canadian business is likely to buy or build today.
See also artificial general intelligence, AI model and computer vision.
This is one term in a plain-English glossary on how AI actually works and where it fits in a Canadian business.
