How does the NIST AI RMF address fairness and harmful bias in AI agents?

The NIST AI RMF addresses fairness and harmful bias in AI agents by integrating characteristics of trustworthy AI, including fairness, into organizational practices and by requiring the identification of potential impacts to individuals and groups.

Specifically, the framework includes the following controls:

• NIST-GOVERN-1.2 mandates that a risk-management culture is in place and that characteristics of trustworthy AI, such as fairness, are integrated into organizational practices.
• NIST-MAP-5.1 requires the identification of potential positive and negative impacts to individuals, groups, and society, which includes identifying data-sensitivity and regulated-data exposure.
• NIST-GOVERN-6.1 addresses risks from third-party models, datasets, and tools, including provenance and licensing, which can be sources of bias. This cross-maps to OWASP LLM03/LLM05 (supply chain).
• NIST-MEASURE-2.8 requires mechanisms to log decisions and trace AI behavior, which can help in identifying and understanding biased outcomes.
• NIST-MEASURE-3.1 outlines approaches for tracking identified and emergent risks over time through monitoring, logging, and drift detection, which can help in detecting and addressing bias as it emerges.