Skip to main content

COMPEL Glossary / equalized-odds

Equalized Odds

Equalized odds is a mathematical fairness criterion requiring that an AI system has equal true positive rates and equal false positive rates across different demographic groups, meaning the system is equally accurate for each group and distributes its errors fairly.

What this means in practice

Unlike demographic parity, equalized odds accounts for actual qualifications or conditions, requiring equal accuracy rather than equal outcome rates. For organizations choosing fairness metrics, equalized odds is often preferred in contexts like medical diagnosis or recidivism prediction where base rates differ legitimately between groups. In COMPEL, equalized odds is one of several fairness metrics discussed within the ethical evaluation framework, where the AITGP is expected to help organizations select appropriate metrics based on context, values, and legal requirements.

Why it matters

Equalized odds is a fairness criterion that accounts for actual qualifications or conditions, requiring equal accuracy across demographic groups rather than equal outcome rates. This distinction matters because in contexts like medical diagnosis, where base rates differ legitimately between groups, demographic parity could produce paradoxically unfair results. Choosing the right fairness metric is a governance decision with significant ethical and legal implications.

How COMPEL uses it

Equalized odds is discussed within the ethical evaluation framework where AITGPs help organizations select appropriate fairness metrics based on context, values, and legal requirements. During Model, the chosen fairness metric is documented as part of the ethical review process. The Evaluate stage measures model performance against equalized odds criteria where applicable, with results informing governance decisions about model acceptability.

Related Terms

Other glossary terms mentioned in this entry's definition and context.