What metric is often used in GenAI risk models to evaluate bias?

GenAI risk models evaluate bias by using quantitative fairness metrics. These metrics are specifically designed to measure disparities in model performance across different demographic groups (e.g., race, gender). Metrics like demographic parity (or statistical parity) ensure that the proportion of positive outcomes is similar across groups, while equalized odds ensures that the true positive rate and false positive rate are equal across groups. These are standard, technical measures for identifying and quantifying systemic bias in a model's predictions, which is a critical component of responsible AI governance and risk management.

A. Accuracy rate without considering demographics is incorrect because high overall accuracy can mask severe underperformance and bias against specific minority subgroups.

C. Computational efficiency during training measures resource consumption (time, memory) and is unrelated to the fairness or ethical implications of the model's outputs.

D. The number of parameters in the model relates to its complexity and capacity, not its fairness. A model can be biased regardless of its size.

1. National Institute of Standards and Technology (NIST). (2023). AI Risk Management Framework (AI RMF 1.0). NIST AI 100-1. In the discussion of the "MAP" function

the framework emphasizes measuring for bias: "Approaches to test

evaluate

verify

and validate (TEVV) AI systems for bias include... statistical and conditional statistical parity..." (p. 29). This directly supports the use of fairness metrics.

2. Verma

S.

& Rubin

J. (2018). Fairness Definitions Explained. In Proceedings of the 2018 ACM/IEEE International Workshop on Software Fairness (FairWare). This paper provides a formal overview of fairness metrics

stating

"Statistical Parity (a.k.a. Demographic Parity)... requires that the prediction... be independent of the protected attribute" (Section 3.1

p. 2). It also defines Equalized Odds (Section 3.2). DOI: https://doi.org/10.1145/3194770.3194776

3. Stanford University. (2021). CS224N: Natural Language Processing with Deep Learning - Lecture 18: Fairness

Accountability

Transparency

and Ethics in NLP. The lecture notes explicitly discuss fairness metrics for bias evaluation

covering "Group Fairness: Equalized Odds" and "Group Fairness: Demographic Parity" as key methods to assess model bias across different populations (Slides 35-40). Available at: https://web.stanford.edu/class/cs224n/slides/cs224n-2021-lecture-18-FATE.pdf