AI Fairness and Bias: Testing for bias in a clinical population

NIST AI RMF MEASURE 2.11 is the outcome on AI fairness and bias. It is short to read and easy to underestimate, because satisfying it means producing evidence rather than holding an intention.

What the framework says

Fairness and bias – as identified in the MAP function – are evaluated and results are documented.

That is the text of NIST AI RMF MEASURE 2.11. Read it closely: it describes an outcome to achieve, not a product to buy, which is why two practices of the same size can satisfy it in different ways and both be right.

What this looks like in practice

In an assessment, NIST AI RMF MEASURE 2.11 is not one question. It resolves into several, and each one is really asking for different evidence. These are among the questions asked about it:

• Has this AI system been evaluated for fairness and harmful bias, with the results documented?
• Has harmful bias in this generative AI system's generated content been measured in context, through methods such as field testing with the groups and languages it serves, red-teaming with counterfactual and low-context prompts, or sampling production output for denigrating content, with disparities in quality of service quantified?
• Are use-case-appropriate benchmarks used to quantify bias, stereotyping, denigration and hateful content in this generative AI system's outputs, with the benchmarks' assumptions and limitations documented?

Ownership usually sits with the data science team, though the evidence is often produced by someone else, which is where the trail breaks. A control that works but has no owner tends to stop working the month the person who quietly maintained it changes roles.

Who this applies to

NIST AI RMF is voluntary rather than binding, and that changes how it should be used. Adopting it does not create a legal obligation, and declining it is not a violation. It earns its place when a customer contract names it, when it gives structure to AI governance work you already owe under a law or a contract, or when a recognized practice is worth having on record. Treat it as a way of organizing that work, not as a second regime.

Building it, in order

The sequence below is the order that produces evidence as a by-product rather than as a separate documentation exercise:

1. Select fairness metrics suited to the system's use and affected groups.
2. Evaluate performance across relevant groups.
3. Record disparities found and mitigations applied.
4. Identify the groups, languages and dialects relevant to the system's users and to the people affected by its outputs.
5. Measure harmful bias in context using one or more methods: field testing with those groups to estimate how often they are exposed to generated content showing harmful bias, red-teaming with counterfactual prompts (the same request with a group attribute changed) and low-context prompts that invite stereotypes, or sampling a fraction of production output and annotating denigrating content.
6. Compare quality of service across groups and languages and, where the system feeds categorical or numeric decisions, apply fairness metrics such as demographic parity or equalized odds to the outcome.

Where else this maps

The questions behind NIST AI RMF MEASURE 2.11 also map to NIST AI 600-1 2.6, which is voluntary guidance rather than law; following it does not by itself establish compliance with any law. That matters for scoping: the same evidence can be written once and referenced from each, rather than rewritten per framework.

Evidence an assessor asks for

For this outcome the evidence types that satisfy it are record and report. The distinction matters more than it looks: a policy states what you intend to do, a procedure states how, and a record proves it happened on a date. Auditors ask for all three, and a practice that has written the first two often has nothing for the third.

Date every document and keep the superseded versions. The framework sets no retention period, so this is our recommendation rather than a requirement: keep each version for as long as the system it describes is in use and for a period after it is retired, because a question about a past decision is answered from the version that was in force when it was made.

Where this usually goes wrong

The usual gap is a control that exists in practice and nowhere in writing. Someone does the work, reliably, and it has never been written down, so the organization cannot demonstrate it and cannot notice when it stops. The fix is not more control, it is a dated record and a named owner.

What to do about it

• Produce the evidence, not the intention. For NIST AI RMF MEASURE 2.11 that means dated records of each occurrence, not a description of the process and a dated report with the period it covers and who reviewed it. If the only evidence is that someone knows how this works, you have a practice and not a record.
• Put a name on it. The data science lead should be recorded as the owner. A control with no owner stops working the month the person who quietly maintained it changes roles, and nobody notices.
• Set a review cadence and keep the evidence of each review. An annual date with a dated output is a common floor; what fails an assessment is a control implemented once and never re-checked against a changed environment.
• Do not treat this as a legal obligation. It is a voluntary framework. Adopt it where a customer contract or a regulator's guidance names it, or where it gives structure to AI governance work you owe for other reasons, and say in your own documentation which of those reasons applies.

Work backward from the evidence. Decide what evidence would prove AI fairness and bias to someone who does not already trust you, then build the control that produces it. Programs built in that order tend to stay current, because the evidence goes stale visibly.

Related outcomes

• AI Security and Resilience: Testing an AI system against adversarial input
• AI Risk Tracking: Tracking AI risks that resist measurement
• AI System Performance: Measuring performance on your own population

References

• MEASURE 2.11, in NIST AI Risk Management Framework 1.0 (NIST AI 100-1). This is a voluntary framework, not a legal requirement. The quoted statement is NIST's own, from the Core tables of NIST AI 100-1.
• Crosswalk references in this post are the mappings in our question bank, and are not verified against the published text of those standards.
• This post describes what a framework outcome asks for and what evidence documents it. It is not legal advice, and your counsel owns the conclusion. State law may impose additional obligations of its own.