Fairness and bias management

Unfair bias is identified, tested for and reduced at each stage of a use of AI, from design through to live operation.

Why it matters

AI can reproduce or amplify unfairness present in its data, its design or how it is used, and the effect often falls on groups already at a disadvantage. Bias needs managing across the lifecycle because it can appear after go-live as data and users change.

Levels

0 Absent	Confirmed absent: the evidence shows nothing is in place. Not known is recorded as insufficient evidence, not 0.
1 Initial	Bias is not considered, or is assumed not to be present.
2 Developing	Bias is considered for some uses, usually at design, but testing is informal and live use is not checked.
3 Defined	A documented fairness approach defines groups and measures for each in-scope use, tests before release, records mitigations and proxies, and monitors fairness in live use.
4 Managed	Fairness breaches in live use lead to investigation, root-cause analysis and action, and fairness results are reported to the oversight forum.
5 Optimising	Fairness methods are improved from monitoring results and external developments, and lessons are shared across uses.

Level 3 criteria

Level 3 is met when every applicable criterion is met. Criteria marked with an overlay apply only to organisations with that characteristic (overlays).

RSK-06.1 A documented approach to fairness and bias is applied across the organisation.
RSK-06.2 For each in-scope use of AI, the groups at risk and the fairness measures that matter are defined.
RSK-06.3 Each in-scope use of AI is tested for fairness before release.
RSK-06.4 Fairness mitigations are recorded for each in-scope use of AI.
RSK-06.5 Fairness is monitored in live use through MON-02.
RSK-06.6 Proxies for protected characteristics are identified, and any use of them is recorded with reasons.

Assessment questions

For a given use of AI, which groups could be treated unfairly, and how was that worked out?
What testing for bias is done before release, and against what measures?
How would you know if a live use started producing unfair results?

Evidence an assessor would expect

The documented approach to fairness and bias
Bias test results and mitigation records for recent uses
Live fairness monitoring reports

How to reach Level 3

Write a fairness approach for the organisation explaining how groups at risk are identified, how fairness measures are chosen and how testing is done.
For each in-scope use, define the groups at risk and the fairness measures that matter, and identify proxies for protected characteristics, recording any use of them with reasons.
Test against the chosen measures before release and record the mitigations applied.
Hand the fairness measures to live monitoring (MON-02) so they are tracked in use.

Common pitfalls

Testing only overall accuracy, which can hide poor results for smaller groups.
Removing protected characteristics but ignoring proxies that carry the same information.

Starter artefacts

Fairness approach document
Per-use fairness test plan and results record

Ownership and sequence

Accountable: business owner
Contributors: data owner, delivery lead, risk lead
Depends on: RSK-03
Roadmap step: 5
Lifecycle: design, acquire_build, deploy, operate

Cross-references

Algorithmic Transparency Recording Standard	Tier 2 Development data specification partial; Tier 2 Risks, mitigations and impact assessments partial
EU AI Act	Art 26(4) partial; Art 9 partial; Art 10 partial
GDS AI Playbook	P2 partial
ISO/IEC 42001	A.5.4 partial
JSP 936 Part 1	para 49a partial; para 49d direct; para 52 partial; para 69 direct; para 70 partial; para 71 partial
MITRE AI Maturity Model	P1.3 direct; P4.2 direct; P5.3 partial; P6.2 partial
NIST AI 600-1	R6 direct
NIST AI RMF 1.0	MEASURE 1.1 partial; MEASURE 2.11 direct
UK Data and AI Ethics Framework	Fairness 1 direct; Fairness 2 direct; Fairness 3 direct; SAT 1.13 direct; SAT 2.7 direct; SAT 2.8 direct; SAT 3.1 partial; SAT 3.2 direct
UNESCO AI Maturity Framework	4.2 partial; 5.2 direct; 5.3 partial

direct: this control at Level 3 meets the item. partial: it covers part of it.