AI Standard: Fairness, Bias, and Justice

1. Purpose

Define enforceable requirements to identify, measure, mitigate, and govern fairness and bias risks in AI systems, especially where systems influence high-impact decisions.

2. Ethical Mapping

A1 Unity & Social Cohesion: prevent structural inequity and discriminatory burdens
A2 Human Dignity & Moral Agency: preserve meaningful agency and non-coercive treatment
A3 Justice, Due Process, and Remedy: support contestability, auditability, and correction
A4 Truthfulness & Trustworthiness: ensure claims about fairness are evidence-based

3. Scope

This standard applies to:

AI systems used in Tier 1–3 contexts (Tier 0 is encouraged where feasible), and
both automated decision-making and decision-support systems when they materially influence outcomes for affected parties.

This standard covers:

training-time and deployment-time fairness risk management,
dataset provenance and representativeness,
evaluation, monitoring, and remediation obligations,
prohibited practices that undermine fairness and justice.

4. Definitions (only if required)

The following definitions are specific to this standard and supplement 00_foundations/definitions.md.

Bias: a systematic tendency in data, measurement, modeling, or operations that produces errors or distortions that can lead to unequal outcomes across groups or contexts.
Discrimination: differential treatment or outcomes that is unlawful, unjustified, or otherwise impermissible under the applicable normative and legal regime; discrimination may be direct (explicit) or indirect (via proxies).
Fairness (operational): a set of context-specific constraints and evaluation criteria that limit unacceptable disparities and ensure consistent treatment across relevant groups or individuals, selected and justified for the decision context.
Equity: a focus on distribution of benefits and burdens, accounting for baseline disadvantage and structural context; equity assessments consider whether the system increases or reduces unjust disparities.
Statistical fairness: fairness criteria assessed using statistical measures (e.g., error rates, calibration, selection rates) disaggregated across relevant groups.
Contextual fairness: fairness criteria assessed against the real-world decision context, including how a system is used, the stakes, available remedies, and the social meaning of errors.

5. Normative Requirements

MUST

AI-FJ-1 (Fairness Objective & Context). For Tier 1–3 deployments, operators MUST document:

the decision context and stakes (including whether it is a high-impact decision),
whether the system is decision-support or decision-making,
the fairness objectives selected for the context, and
which affected parties and groups are in scope for evaluation.

AI-FJ-2 (Bias vs Discrimination Analysis). For Tier 1–3 deployments, operators MUST:

analyze and document plausible bias pathways (data, labels, proxies, deployment operations), and
identify discrimination risk pathways, including indirect/proxy discrimination, under applicable law and policy.

AI-FJ-3 (Dataset Provenance and Representativeness). For Tier 1–3, operators MUST maintain dataset and label provenance records and evaluate representativeness relative to the intended deployment population, including:

known gaps and likely error distributions,
data collection and labeling procedures,
transformation and filtering steps, and
mitigation actions (e.g., additional sampling, reweighting, re-labeling).

AI-FJ-4 (Protected Attributes and Proxies). Operators MUST:

identify relevant protected or otherwise vulnerable groups for the context, subject to legal constraints, and
evaluate proxy variables that may reproduce protected-class effects (e.g., geography, language, device signals), documenting mitigations or justifications.

Where collection or use of protected attributes is restricted, operators MUST document alternative evaluation strategies and their limitations.

AI-FJ-5 (Pre-Deployment Fairness Evaluation). Before deployment in Tier 1–3 contexts, operators MUST perform a fairness evaluation that:

reports performance and error characteristics relevant to the context,
includes disaggregated analysis across relevant groups where legally and ethically permissible, and
includes a documented rationale for selected metrics and thresholds.

No single metric is sufficient for all contexts; metric selection MUST be justified.

AI-FJ-6 (Residual Risk Register). Tier 2–3 operators MUST maintain a fairness residual risk register that includes:

identified disparity risks and their expected severity,
mitigations implemented and residual risk after mitigation,
monitoring thresholds and escalation triggers, and
assigned owners and review cadence.

AI-FJ-7 (Deployment Monitoring). Tier 2–3 operators MUST implement post-deployment monitoring for:

distribution shift and data quality drift relevant to fairness,
changes in disparity indicators over time,
emergent proxy pathways caused by operational changes (policies, incentives, user behavior).

AI-FJ-8 (Actionability and Remediation). When monitoring indicates material fairness regression or unacceptable disparity, operators MUST:

initiate escalation per documented thresholds,
implement containment or rollback where feasible, and
document remediation actions and verification evidence.

AI-FJ-9 (Decision-Support Safeguards). When a system is deployed as decision-support for Tier 2–3 contexts, operators MUST implement safeguards against undue reliance, including:

reviewer guidance describing appropriate use and non-use cases,
presentation of uncertainty and limitations,
the ability for reviewers to override/ignore recommendations without penalty, and
periodic review of human-in-the-loop performance for disparate impacts.

AI-FJ-10 (Prohibited Practices). Operators MUST NOT:

claim fairness based solely on aggregate accuracy or a single fairness metric without justification,
use “fairness through unawareness” (excluding protected attributes) as a standalone fairness claim without evidence of proxy-risk handling,
suppress, delay, or selectively report known disparity findings for Tier 2–3 systems,
deploy systems for high-impact decisions without a documented contestation and remedy pathway (see also 02_ai_standards/accountability.md).

SHOULD

AI-FJ-11 (Equity Impact Assessment). Tier 2–3 operators SHOULD conduct an equity impact assessment that evaluates distribution of benefits and burdens, plausible harm pathways, and mitigations; if not performed, operators MUST document the rationale.

AI-FJ-12 (Independent Review). For Tier 3, operators SHOULD obtain independent review of fairness evaluations and monitoring design, including review of metric selection and thresholds.

AI-FJ-13 (Stakeholder Consultation). For Tier 2–3 systems, operators SHOULD incorporate structured consultation with affected parties or representatives into fairness objective selection and monitoring design.

MAY

AI-FJ-14 (Public Fairness Reporting). Tier 2–3 operators MAY publish redacted fairness reports, including metric rationales, monitoring approach, and remediation summaries, with redactions limited to privacy, safety, or legal constraints.

6. Risk-Tier Considerations

Tier 0: fairness evaluation is context-dependent and often exploratory; operators are encouraged to document datasets, run preliminary disparity checks, and avoid releasing artifacts that materially increase discrimination risk without mitigations.
Tier 1: requires documented context, provenance, and a pre-deployment evaluation (AI-FJ-1 to AI-FJ-5).
Tier 2: adds mandatory monitoring, residual risk tracking, and remediation discipline (AI-FJ-6 to AI-FJ-9), plus consultation expectations.
Tier 3: adds strong expectations for independent review, public-interest justification where dual-use or systemic harm is plausible, and stricter escalation thresholds consistent with ESCALATION_AND_PAUSE.md.

7. Compliance Evidence

Minimum evidence artifacts (as applicable):

fairness objective and context document (AI-FJ-1)
bias/discrimination pathway analysis (AI-FJ-2)
dataset provenance and representativeness report (AI-FJ-3)
proxy variable analysis and mitigations (AI-FJ-4)
pre-deployment fairness evaluation report with metric rationale (AI-FJ-5)
fairness residual risk register (AI-FJ-6)
monitoring dashboards, thresholds, and escalation logs (AI-FJ-7)
remediation records and verification tests (AI-FJ-8)
decision-support reviewer guidance and training materials (AI-FJ-9)
contestation/remedy process artifacts (AI-FJ-10)

Traceability Table (Requirement → Axiom → Evidence)

Requirement ID	Axiom(s)	Evidence Artifacts
AI-FJ-1	A3, A4	context doc, fairness objectives, scope statement
AI-FJ-3	A1, A3, A4	provenance records, representativeness report
AI-FJ-5	A3, A4	evaluation report, metric rationale, thresholds
AI-FJ-6	A1, A3, A4	residual risk register, owner assignments
AI-FJ-7	A1, A3, A7	monitoring dashboards, drift analyses
AI-FJ-8	A3, A4, A5	escalation logs, rollback/remediation evidence
AI-FJ-10	A3, A4	claim review artifacts, contestation workflow

8. Known Limitations

Fairness criteria are context-specific; universal metric mandates are not feasible across domains.
Legal constraints may limit collection or use of protected attributes; alternative strategies can reduce but not eliminate discrimination risk.
Some disparity drivers are structural and cannot be eliminated by model changes alone; this standard requires documentation and mitigation, not claims of total elimination.

9. Future Considerations

Methods for auditing fairness in generative and agentic systems where outputs are open-ended.
Privacy-preserving disparity measurement approaches suitable for restricted jurisdictions.
Standardized reporting formats for fairness residual risk registers to improve audit interoperability.

Appendix A (Non-normative): Rationale

Fairness is operationalized through context-specific constraints, evaluation, monitoring, and remedy pathways. Statistical measures are necessary but insufficient; contextual fairness ensures that evaluation reflects real-world use, stakes, and available recourse.

Appendix B (Non-normative): Failure Modes & Abuse Cases

selecting metrics that obscure harms (“metric gaming”)
relying on “fairness through unawareness” while proxies reproduce protected-class effects
shifting decision thresholds post-audit to reintroduce disparities
human-in-the-loop processes that amplify bias through uneven override behavior

Change Log

v0.2: Expanded fairness standard; renamed to 02_ai_standards/fairness_and_justice.md; added traceability, tiering, and prohibited practices.