Fine-tuning is taking a model that already exists — usually a foundation model someone else built — and continuing to train it further on a narrower, more specific set of data or feedback so it performs better on one particular job, rather than starting from nothing. Canada’s Cyber Centre uses the word directly, listing it among the steps an organization should keep doing after an AI system is already in production: “Continuously fine-tune or retrain the AI system with appropriate external feedback to improve the quality of outputs.”
No Canadian statute regulates fine-tuning by name, but the activity lands squarely inside a distinction Canada’s privacy regulators already draw for a different reason. Their generative-AI principles describe a “Developer or Provider” as an organization that determines “how a generative AI system operates, how it is initially trained and tested, and how it can be used”, against an ordinary user who is simply “using a generative AI system as part of their activities.” An organization that only sends prompts to someone else’s chatbot sits in the second camp. The moment it fine-tunes the underlying model on its own data, it has started to determine how that model behaves — which is exactly the test the first definition uses, and exactly why the regulators warn that an organization’s role “might shift between or play multiple roles at once.”
A brokerage fine-tunes a general-purpose model on a year of its own past client correspondence so replies come back in the firm’s own voice instead of a generic one. That correspondence is personal information the firm already held and was already responsible for; fine-tuning does not remove it from that responsibility, it just puts it to a new use inside the same organization’s own systems. The foundation model underneath does not change; only the narrower layer trained on top of it does — which is also why a fine-tuned model still needs the same output checks as the general-purpose one it started from. See inference for what happens every time the finished, fine-tuned model is actually used afterward.
See also foundation model and inference.
This is one term in a plain-English glossary on how AI actually works and where it fits in a Canadian business.
