Abstract: AI algorithms increasingly inform decisions in credit, hiring, healthcare, criminal sentencing, and public administration. In Canada, governing these tools is no longer only a technical matter; it is also a matter of public policy. Yet policy debate still often asks whether an algorithm is “fair” as if that question could be answered at the level of the model alone. In many high-stakes settings, however, algorithmic outputs are filtered through human judgment, institutional rules, and competing fairness objectives before they shape final outcomes. This article draws five policy-relevant insights from the academic literature. Fairness has multiple competing definitions, and they can pull in different directions. Better predictive performance does not necessarily reduce disparity. A widening disparity after adoption does not by itself identify the source of unfairness. How fairness risks are framed can influence organizational response. And evaluating the model is not the same as evaluating the decision process built around it. These points matter for how policymakers define fairness objectives, interpret post-deployment disparities, communicate fairness risks, and govern the translation from model output to institutional action.