AIPCH17 — Bias-Controlled & Fairness-Measured
“Where fairness obligations apply, bias and fairness are defined, measured, governed, and acted upon.”
What AIPCH17 is really asserting
AIPCH17 does not claim that every AI Product has the same fairness surface.
Fairness obligations depend on the Product’s intended use, affected populations, intelligence effects, risk, domain policy, and Governance Obligation Profile.
Where fairness is applicable, it must be an evidence-backed governed property, not a producer claim or one-time training check.
Applicability
AIPCH17 may be Not Applicable where the approved Product use has no meaningful people/group allocation, ranking, evaluation, treatment, representation, or comparable fairness concern.
Examples may include some infrastructure-oriented symbolic reasoners or machine-control Products whose approved effects do not create a relevant fairness surface.
N/A requires a rationale tied to Product context and governance evidence.
N/A means the fairness characteristic is not relevant to the assessed Product/use context; it does not mean fairness was ignored.
Positive criteria when applicable
AIPCH17 is Met when:
- relevant populations/attributes/outcomes are identified;
- appropriate fairness criteria and thresholds are defined;
- evaluation frequency matches the Product’s mechanism and risk;
- material fairness signals are observable and included in trust/governance evidence;
- threshold breaches have defined mitigation, escalation, or restriction paths;
- fairness evidence remains current for the assessed Product version and intended use.
Continuous runtime measurement is required only where the Governance Obligation Profile or Product dynamics make it necessary. Other Products may require periodic or event-triggered evaluation.
Partial / Not Met
Partial:
- relevant fairness dimensions are known but monitoring, evidence currentness, or mitigation is incomplete.
Not Met:
- fairness is applicable but no relevant criteria exist;
- only generic/unrelated metrics are used;
- known disparities have no governed response;
- evidence is stale for a materially changed Product/use context.
Evidence signals
- Governance Obligation Profile;
- intended-use and affected-population analysis;
- fairness metrics/evaluations;
- subgroup outcome evidence;
- mitigation/escalation records;
- AIPS-DPP and assurance currentness.
Agent decision rule
Determine applicability first. If fairness is applicable, mark Met only when relevant fairness risks are evidence-backed, current, monitored at an appropriate cadence, and governed through explicit response mechanisms.