Definitions
The clause defines the algorithm, the system, bias, the bias profile, the stakeholder types, protected and sensitive attributes, and the three drift terms.
What clause 3.1 asks
Several terms carry a fixed meaning. An algorithm is a finite set of steps for solving a problem, and an algorithmic system applies one or more of them to data to produce a result. The document calls the system it covers an autonomous intelligent system, or AIS, and includes symbolic, subsymbolic and mixed designs. Bias means a consistent difference in how some objects, people or groups are treated compared with others. Bias that serves the system's purpose is wanted; bias that damages the purpose or harms people is unwanted. The bias profile is where the work and its results are kept. A stakeholder is a person or thing that the system affects or that affects the system; an impacted stakeholder receives the result, and an influencing stakeholder shapes the system. A group is a set of objects linked by shared attributes. A protected attribute has legal protection in the setting, and a sensitive attribute is either protected or carries social or cultural weight there. Data drift is a fall in predictive accuracy as live data changes. Concept drift is a shift in the decision boundary that lowers accuracy even when the data is stable. Model decay is a fall in the chosen performance measures, and it may signal drift.
Why it matters
The definitions are where most arguments about scope are settled. If a team reads bias as statistical disparity only, it will miss the wanted bias that makes the system work. If it reads stakeholder as user only, it will miss the people affected by a decision they never see. If it reads drift as one thing, it will monitor the wrong signal. The definitions are also the only place where the standard permits a defined term to stay in its exact form, because a defined term has to mean the same thing on every page.
What good looks like
Build a glossary from these terms and use it in every clause document. Where the organization already has a definition, map it to the standard's term rather than replacing one with the other. For protected and sensitive attributes, record the jurisdiction that supplies the protection, because the same attribute can be protected in one market and not in another. For drift, keep the three terms separate in the monitoring plan.
Evidence an assessor asks for
The assessor will test whether the team uses the terms consistently. A data map that calls every attribute sensitive makes the protection analysis useless.
- Glossary The defined terms with the organization's map and the jurisdiction for protection
- Attribute register Protected and sensitive attributes, with the law or context that makes them so
- Drift definitions The three drift terms as they will be measured in the monitoring plan
- Term map A note where an internal term differs from the standard's term
Where it commonly fails
- Bias is used as a synonym for unfairness, so wanted bias is treated as a defect
- Only users are counted as stakeholders
- Data drift and concept drift are monitored with the same test
What to do before publication
Agree the glossary before the clause 4 workshop. The workshop will otherwise spend its time on definitions.
Read across
- 1.4 Field of application
- 5.1 Purpose
- 6.1 Purpose
- 7.4.2.1 Collection and creation conditions
- 9.3.4.2 Each iteration of the ongoing evaluation