Control · C-075 · Fairness & Customer Outcomes
Fairness requirements and metric selection
For each system that affects people, the relevant groups, the legal test, the fairness measures and their thresholds are defined and justified before testing begins.
- Control type
- Directive
- Risks addressed
- 5
- Audience
- Heads-of · Practitioner
- Frameworks
- Section 19 · Article 10 · MEASURE 1
- Responsible AI domain
- Lifecycle stage
- Reviewed
- 2026-11-11
What it does
It decides what fair means for a particular system before anyone measures it. A short, documented analysis identifies who could be disadvantaged, which legal tests apply, which measures correspond to the harm in question, what difference would be unacceptable, and what data will be used to test. The business owner is accountable for the choices; legal, compliance and data science contribute. It is done at design and revisited when the use or the population changes.
How to implement it
- Start from the impact assessment: who is affected and what harm they could suffer.
- Identify the protected characteristics and other groups that matter, including combinations.
- Choose measures that reflect the harm, such as wrongful refusal rates for a credit decision.
- Set thresholds in advance, with reasons, and avoid borrowing a rule from another legal system.
- Decide how group membership will be established, and settle the lawful basis for that data.
- Record what trade-offs between measures were accepted and by whom.
Evidence it operates
Design: a fairness requirements template. Operation: a completed statement for each system above the threshold, dated before testing; evidence of legal and compliance input; test reports that use the stated measures; a revision following a change of use.
Where it fits
At design, before data preparation and testing. It is the primary control for mis-specified fairness measurement and missing data to test for bias, and with bias testing and fairness monitoring for inadequate fairness testing before deployment. It relies on data minimisation and lawful basis review for the data question.
Limits
No set of measures captures fairness completely, and different measures conflict, so the statement records a judgement that others may dispute. Legal tests turn on justification and context, which a number cannot settle. The requirement is only as good as the understanding of those affected that went into it.
Risks this mitigates
Primary mappings
- Inadequate fairness testing before deployment Medium
- Sets the groups, metrics and thresholds the tests must meet. An AI system that affects people is released without testing whether its performance and outcomes differ between groups.
- Mis-specified fairness measurement Medium
- Requires the choice of groups, metric and threshold to be justified against the legal test and the harm. Fairness is tested with a metric, threshold or grouping that does not match the legal test or the harm, so a system passes while remaining unfair.
- Missing data to test for bias Medium
- Identifies the data needed for testing and how it will lawfully be obtained. The organisation does not hold, or believes it may not use, the data on protected characteristics needed to detect discrimination, so bias goes unmeasured.
Supporting mappings
- Proxy discrimination via correlated features High
- Requires proxy analysis for features likely to correlate. A model produces discriminatory outcomes through features that correlate with a protected characteristic, even though the characteristic itself is not an input.
- Training data bias and discriminatory outcomes High
- Defines which groups and measures matter for this system. Bias in the data a model learns from, whether from history, sampling or labelling, is reproduced or amplified in its outputs and in the decisions that rely on them.
Framework and clause references
| Framework | Clause | Title |
|---|---|---|
| Equality Act 2010 | Section 19 | Indirect discrimination The legal test. |
| EU AI Act | Article 10 | Data and data governance Bias examination. |
| NIST AI Risk Management Framework (AI RMF 1.0) | MEASURE 1 | Appropriate methods and metrics are identified and applied Selection of metrics. |
Related terms
How to cite
Axiom Verity (2026). Fairness requirements and metric selection. Reference Library, C-075. https://axiom-verity.com/controls/fairness-requirements-and-metric-selection. Reviewed 2026-11-11.
Found an error or a better source? Report a correction.