An AI model is the trained artifact — the thing that was fit to data — sitting inside a larger AI system. the United States’ NIST AI Risk Management Framework treats model-building as its own phase of work, carried out by “machine learning experts, data scientists, developers, third-party entities, legal and privacy governance experts, and experts in the socio-cultural and contextual factors associated with the deployment setting”, separately from the interface, rules and data pipeline wrapped around it.
Canada’s Directive on Automated Decision-Making treats the model as a distinct, auditable component with its own obligation, separate from the automated decision system as a whole. Before any such system goes into production, the department responsible must test “the data, information, and underlying model for accuracy, unintended biases and all factors that may unintentionally or unfairly impact the outcomes or violate human rights and freedoms”. That wording only makes sense if the model is something you can point at, swap, and test on its own — not a synonym for the software product around it.
The Federal Court draws the same object-level distinction for a different reason: the Federal Court’s Notice on the use of AI in court proceedings limits its AI Declaration rule to “a computer system capable of generating new content and independently creating or generating information or documents, usually based on prompts or information provided to the system” — describing the trained, generative artifact itself, not the surrounding application. A fixed rule wrapped around that same model is not what either source is describing.
A document-intake tool used by a Canadian firm might keep its interface, its rules for routing files and its user permissions completely unchanged while the classifier at its centre — the model — is replaced with a newer one trained on more recent data. To the person using the tool, nothing visibly changes. To the department responsible for it under the Directive’s own testing requirement, that swap is exactly the event that triggers re-testing “the data, information, and underlying model for accuracy, unintended biases” again, because the model — not the system around it — is what changed.
See also machine learning, deep learning and algorithm.
This is one term in a plain-English glossary on how AI actually works and where it fits in a Canadian business.
