Testing the AI model with a curated and representative sample data set allows auditors to directly evaluate the fairness and bias of model decisions. This approach is aligned with best practices outlined in the AAIA™ Study Guide, as it enables quantifiable analysis of model behavior across different demographics or input scenarios.

“To assess fairness, auditors should use controlled data sets to evaluate whether model outputs disproportionately impact specific groups. This empirical testing provides stronger evidence than qualitative methods.”

While metadata (A) and developer interviews (C) can supplement findings, only B provides objective, reproducible evidence. Option D may reflect real-world interactions but lacks the control and consistency required in an audit.