Miller's chapter on model performance introduces a concept that surprises many first-time AI builders: aggregate accuracy is often the wrong metric to optimise for. A credit-scoring model that is 95% accurate but systematically wrong for a specific demographic group will produce discriminatory outcomes at scale. A medical triage model that is 99% accurate but fails on the most severe cases will miss the patients who need it most. The relevant performance question is not 'how often is the AI right?' but 'when is the AI wrong, and what are the consequences?' Good AI performance means understanding the distribution of errors — which types of cases fail, at what rate, and with what consequences — not just the aggregate accuracy number. This is more sophisticated analysis, but it is the analysis that determines whether an AI deployment is actually safe to use.