Supervised learning is the most widely used way an AI system is trained: it studies a large set of examples that already carry the correct answer — a photo already tagged “cat,” a transaction already flagged “fraud” — and adjusts itself until it can predict the correct answer for new examples it was never shown during training.
The mechanism has a precise technical description, from the United States’ National Institute of Standards and Technology (a US, not Canadian, source): labelled training data is given to a training algorithm, which adjusts the model against that data before it is ever deployed on new, unlabelled cases. Two of the most common resulting techniques are “CLASSIFICATION, in which the predicted labels or classes are discrete, and REGRESSION, in which the predicted labels or response variables are continuous”
No Canadian statute names “supervised learning” specifically. What Canadian law does reach is the label data itself, where it is personal information. The federal, provincial and territorial privacy regulators’ joint principles for generative AI instruct developers: “Ensure that any personal information used to train their generative AI models is as accurate as necessary for the purposes” — a rule that bites directly on supervised learning, because its entire method depends on the correctness of the labels a person or process assigned.
A lender building a tool to flag applications for manual review might assemble ten thousand past applications, each already labelled by a human underwriter as “approved” or “declined.” A supervised model studies that labelled history and learns which application features tend to accompany each outcome, so that it can score a brand-new, unlabelled application against the same pattern.
The model is only ever as sound as that label history: if the past decisions it learned from carried a bias, or were recorded inconsistently, the model reproduces the pattern it was shown rather than correcting it — which is exactly why the accuracy of the labels, not just the accuracy of the model’s output, is where a Canadian privacy review of the tool should start.
See also: unsupervised learning, reinforcement learning, machine learning.
Whether a proposed tool actually needs a custom-trained supervised model, or can be scoped around an existing one, is one of the first questions a build should settle — custom-ai-solutions covers how that scoping works.
