Understanding Fairness Metrics

There is no single number for "fair". Different metrics encode different — sometimes mathematically incompatible — definitions of fairness. The report computes several so you can see the whole picture.

The metrics in your report

• Demographic parity — do all groups receive positive outcomes at similar rates? See Demographic Parity Explained.
• Equalized odds — are error rates (false positives *and* false negatives) similar across groups? See Equalized Odds Explained.
• Equal opportunity — the recall half of equalized odds: do qualified people from every group get selected at similar rates?
• Disparate impact ratio — the ratio of selection rates between groups. The classic "four-fifths rule" from US employment law flags ratios below 0.8; it's a useful screening threshold in the EU context too.

Why they conflict

If your groups differ in base rates, you mathematically cannot satisfy demographic parity and equalized odds at once. That's not a bug in the report — it's the trade-off your team must decide on and document.

What the EU AI Act expects

For high-risk systems, Art. 10 demands data governance including examination for possible biases, and Art. 15 demands accuracy monitoring. A periodic bias report with your chosen metrics, thresholds, and rationale is exactly the kind of evidence that satisfies "we examined it".