Equalized odds is a group fairness criterion for classification. It is satisfied when people with the same actual outcome face the same model error rates across protected groups, meaning both true-positive rates and false-positive rates are equal. This differs from demographic parity, which compares overall selection rates without conditioning on the correct label. Equalized odds can be useful when errors have serious and different consequences, but it relies on trustworthy ground-truth labels and may conflict with calibration or other fairness definitions. Teams should report uncertainty, examine intersectional groups, investigate why disparities exist, and consider whether changing thresholds addresses the underlying harm or only adjusts the measured metric.
Anthropic unveils a revised Responsible Scaling Policy with a Frontier Safety Roadmap, regular Risk Reports, and clearer separation between company commitments and industry recommendations.