Measuring Bias
Annex D is informative. It discusses metrics, reference values, the granularity of testing and further resources.
This annex is informative. It explains and illustrates; it does not create a requirement. A finding cannot rest on it alone. The normative clauses are 4 to 9.
What annex D asks
Annex D is an informative annex on measuring bias. It covers metrics, reference values, the granularity of testing and additional resources. It supports clause 9.2.4.3, which requires a metric and reference values, but it does not set a threshold and it does not make any metric mandatory. The standard deliberately leaves the choice to the context of use.
Why it matters
The annex is a menu of measurement options, and the menu is not neutral. Different metrics encode different fairness assumptions, and a metric chosen for convenience can hide the harm the team cares about. Reference values are the organization's decision, not the annex's. The granularity discussion is a reminder that an aggregate metric can conceal a small group's experience.
What good looks like
Choose the metric against the requirement, not against the data. Record the fairness assumption the metric encodes and the alternative metrics considered. Set the reference value with the business owner and the affected-group representative. Test at the granularity the stakeholder set requires, and record any group too small to test. Then record the interpretation rule before the test.
Evidence an assessor asks for
The assessor will ask why this metric rather than another. A metric chosen because it was available is not a decision the organization can defend.
- Metric decision The metric, the fairness assumption it encodes and the alternatives considered
- Reference value The value, who set it and the basis
- Granularity record The groups and subgroups tested, and any group too small to test
- Interpretation rule The rule fixed before the test
Where it commonly fails
- The metric is chosen after the result
- The reference value is a vendor default
- Only the aggregate is tested
What to do before publication
Decide the metric and the reference value in the clause 9.2.4.3 protocol, before the run.
Read across
- 5.2.5 Evaluation in the life cycle
- 9.2.3 Evaluation of design and outputs: outputs
- 9.2.4.3 Testing
- 7.4.2.2 Data exploration
← C Guidance for Risk and Impact AssessmentE Cultural Aspects →