There is no single fairness metric that satisfies all stakeholders for any non-trivial AI system. This is not a limitation of current metrics; it is a mathematical impossibility proved formally in 2016-2017: when group prevalence rates differ, demographic parity, equalized odds, and predictive parity cannot all be satisfied simultaneously by a non-trivial classifier. Yet every Annex III high-risk system in 2026 must operate against some explicit fairness commitment. The decision of which metric to use is therefore not a technical decision. It is a policy decision that allocates harm differently across groups. This lesson walks the four most-cited fairness metrics (demographic parity, equalized odds, predictive parity, four-fifths rule) on a single hiring-screen dataset, names the impossibility tradeoff, maps to EU AI Act Article 10 data governance expectations, and ships the fairness-metric-selection memo the AI Governance Committee approves and the External Auditor reads against ISO 42001 A.7.

Why This Lesson Is Uncomfortable - And Why Skipping It Costs the Program

Most AI Governance Committee meetings in 2025-2026 treated fairness as a technical problem to be solved by selecting the right metric and configuring the right vendor product. The result was usually one of two postures: (a) "we use disparate impact / four-fifths rule because EEOC and NYC LL 144 reference it, so we're covered," or (b) "we measure all four major metrics and the vendor's product tunes against them." Both postures are operationally fine and both fail the deep audit-defensibility test.

The deep test: an auditor or plaintiff's lawyer asks, "Show me the analysis that demonstrates your fairness-metric choice is appropriate for the harm pattern this system produces, given that other metrics would have produced different outcomes for different groups." The technical answer ("we tune against demographic parity") is insufficient. The legal-defense answer ("EEOC says four-fifths rule") is insufficient. The audit-defensible answer is a policy-grade rationale memo that explains why this metric was chosen, what tradeoffs it makes, what tradeoffs it accepts, and what alternative metrics would have produced.

For an Annex III §4 employment system, Annex III §5(b) creditworthiness system, or Annex III §5(c) life/health insurance pricing system, the fairness-metric-selection memo is the load-bearing artifact in the FRIA. For a generative-AI deployment subject to NIST AI 600-1 Risk 12 (harmful bias and homogenization), the same memo is the operational substrate. For an organization seeking ISO 42001 Stage 2 certification, the memo is part of the A.7 data-governance evidence.

The Four Metrics - Defined, With Math, Applied to One Hiring Dataset

Build a worked example. The hypothetical: an HR resume-ranker AI that scores 10,000 candidates for a single high-volume sales role. Demographic breakdown: 5,000 male candidates, 4,000 female candidates, 1,000 non-binary candidates. The AI returns a binary "shortlist / no shortlist" recommendation. After deployment, of the 5,000 male candidates, 1,000 were shortlisted (20% selection rate). Of the 4,000 female candidates, 600 were shortlisted (15% selection rate). Of the 1,000 non-binary candidates, 100 were shortlisted (10% selection rate). The base rate of "actually qualified for the role", measured by post-hire success, is 25% male, 25% female, 25% non-binary (the role does not have group-specific qualification requirements).

Metric 1 - Demographic Parity (Statistical Parity, Group Fairness)

Demographic parity requires that the selection rate be equal across groups, regardless of qualification. Formally: P(predicted_positive | group=A) = P(predicted_positive | group=B). For our dataset, the male selection rate is 20%, female 15%, non-binary 10%. Demographic parity FAILS. The disparity is 20% vs. 15% vs. 10%, substantial across the three groups.

To remediate against demographic parity, the deployer would adjust selection thresholds per group: lower the female threshold to surface 1,000 (20%) female candidates, lower the non-binary threshold to surface 200 (20%) non-binary candidates. The fix increases representation but ignores qualification: some of the additional female and non-binary candidates surfaced may be less qualified than the male candidates already shortlisted at the higher threshold.

Metric 2 - Equalized Odds (Conditional on True Outcome)

Equalized odds requires that the true positive rate AND the false positive rate be equal across groups, conditional on the actual qualification. Formally: P(predicted_positive | group=A, qualified=true) = P(predicted_positive | group=B, qualified=true) AND P(predicted_positive | group=A, qualified=false) = P(predicted_positive | group=B, qualified=false). For our dataset, assume the post-hire-success data shows: of 1,000 male shortlisted, 800 succeeded (80% true positive rate); of 4,000 male not shortlisted, 200 would have succeeded (5% false negative rate). Of 600 female shortlisted, 540 succeeded (90% TPR, actually higher than males). Of 3,400 female not shortlisted, 460 would have succeeded (13.5% false negative rate). The equalized-odds analysis shows that females are MORE qualified per shortlist position than males. I.e., the AI is using a stricter standard for female candidates. Equalized odds FAILS in the direction of disadvantaging female and non-binary candidates.

To remediate against equalized odds, the deployer adjusts thresholds per group to make TPR and FPR match across groups, which requires lowering the female and non-binary thresholds. The fix improves both fairness and overall hiring quality (because the system was rejecting qualified female and non-binary candidates the model was already capable of identifying).

Metric 3 - Predictive Parity (Calibration, Conditional on Prediction)

Predictive parity requires that the probability of being actually qualified, conditional on being predicted-positive, be equal across groups. Formally: P(qualified=true | predicted_positive, group=A) = P(qualified=true | predicted_positive, group=B). For our dataset: male shortlisted have 80% post-hire success; female shortlisted have 90%; non-binary shortlisted have, say, 85%. The predictive parity disparity is 80% / 90% / 85%, modest but present. Predictive parity is roughly satisfied with some disparity.

Predictive parity is the metric most lenders, insurers, and traditional risk-scoring deployers default to. The intuition: if our scoring tool predicts "qualified" and that prediction has the same accuracy across groups, the tool is fair. The problem is that predictive parity ignores the base rates of who is being scored. A tool that surfaces 100 female candidates and 1,000 male candidates can have identical predictive accuracy on both groups but still produce highly disparate selection rates.

Metric 4 - The Four-Fifths Rule (Disparate Impact)

The four-fifths rule, codified in EEOC's Uniform Guidelines on Employee Selection Procedures (1978) and reaffirmed in 2023-2024 EEOC AI guidance, requires that the selection rate for any group be at least 80% of the selection rate of the highest-selecting group. Formally: selection_rate(group=A) / selection_rate(group_with_highest_selection) ≥ 0.80. For our dataset: male selection rate 20% is highest; female 15% / 20% = 0.75 (FAILS four-fifths); non-binary 10% / 20% = 0.50 (FAILS four-fifths substantially). The four-fifths rule FAILS for both female and non-binary candidates.

The four-fifths rule is the most-cited metric in U.S. employment AI law (EEOC Title VII disparate impact framework; NYC LL 144 bias audit). It is roughly equivalent to a relaxed demographic parity (allowing some disparity within a 20% band). Its critics note that it is arbitrary (why 80% and not 90% or 70%?); its defenders note that it provides a clear bright-line test for litigation purposes.

The Impossibility Theorem - Why You Cannot Have All Three

The 2016-2017 impossibility result, proved formally by Kleinberg, Mullainathan, and Raghavan (2017) and by Chouldechova (2017), states: when group base rates differ, demographic parity, equalized odds, and predictive parity cannot all be satisfied simultaneously by a non-trivial classifier. The math is straightforward: each metric imposes a different conditional independence constraint, and the constraints are incompatible when base rates differ.

The implication for deployer-side fairness work: the choice of which metric to use is a policy decision, not a technical decision. The choice allocates harm differently across groups:

• Demographic parity allocates harm to over-represented groups (the model may reject qualified majority candidates to make room for under-represented candidates who may be less qualified by the model's measure).
• Equalized odds allocates harm to false-negative groups (the model maintains accuracy across groups but may not address representation disparity in the underlying applicant pool).
• Predictive parity allocates harm to under-selected groups (the model may be accurate per prediction but fail to surface enough candidates from disadvantaged groups).
• The four-fifths rule allocates harm in a bounded way (disparities of less than 20% are tolerated) but does not address smaller, persistent disparities or contextual differences.

The audit-defensible posture is to choose a metric, document the rationale, name the tradeoffs, name the alternative metrics that would have produced different outcomes, and integrate with the broader Article 27 FRIA + GDPR Article 22 + EEOC Title VII + state-law overlay. The technical "we tune all four" answer is operationally good but legally insufficient, the auditor still asks which metric is the policy-load-bearing choice.

EU AI Act Article 10 Data Governance - The Regulatory Overlay

EU AI Act Article 10 lays out data-governance obligations for high-risk AI systems. The relevant sub-articles for fairness work:

• Article 10(1): Training, validation, and testing data sets shall be subject to data-management and data-governance practices appropriate for the intended purpose, including methods of data preparation, examination of data for biases that may affect health or safety of natural persons or lead to discrimination, identification of relevant data gaps or shortcomings.
• Article 10(2): Data sets shall meet quality criteria including relevance, representativeness, freedom from errors, completeness.
• Article 10(3), Data sets shall have appropriate statistical properties regarding the persons or groups on whom the system is intended to be used.
• Article 10(5): Special categories of personal data (race, ethnicity, political opinions, religion, sexual orientation, etc.) may be processed strictly when necessary for the purpose of bias monitoring, detection, and correction in high-risk systems, with appropriate safeguards.

Article 10 is the regulatory anchor for the fairness-metric-selection memo. The bias-examination obligation under Article 10(1) requires the deployer to examine data for biases; the documentation must demonstrate the specific examination methodology, including which fairness metrics were applied and which group attributes were analyzed. Article 10(5) is the carve-out that permits collecting protected-attribute data specifically for bias-monitoring purposes, without it, GDPR Article 9 special-category-data restrictions would block the very analysis Article 10(1) requires.

The Fairness-Metric-Selection Memo - The L1 Artifact

The L1 artifact for this lesson is a fairness-metric-selection memo: a per-system document that names the AI system, the population it scores, the protected attributes analyzed, the fairness metrics computed, the chosen primary metric with rationale, the named tradeoffs and the rejected alternatives, the integration with Article 27 FRIA + Article 10 data governance + EEOC / NYC LL 144 / Texas TRAIGA overlays, and the refresh cadence. A representative memo for the hiring-screen example:

• System: Acme Talent Acquisition AI (high-volume sales role screening).
• Population: 10,000 candidates / quarter; demographic breakdown roughly 50% male / 40% female / 10% non-binary (representative of the candidate pool, not the broader workforce).
• Protected attributes analyzed: sex (M/F/NB), race/ethnicity (5 categories), age (3 bands), intersectional (sex × race).
• Fairness metrics computed: demographic parity (current state: 20% / 15% / 10% selection rate), equalized odds (TPR analysis shows female candidates more qualified per shortlist position), predictive parity (80% / 90% / 85% post-hire success), four-fifths rule (FAILS for female 0.75 and non-binary 0.50).
• Chosen primary metric: Four-fifths rule for compliance threshold (EEOC + NYC LL 144 alignment); equalized odds for operational quality target (because the data shows the model is under-shortlisting qualified female / non-binary candidates).
• Rationale: Four-fifths is the bright-line legal-defense baseline. Equalized odds is the operational improvement target because the data demonstrates the model is rejecting qualified female / non-binary candidates the model itself can identify. Demographic parity rejected as primary target because the data does not support equal base rates of "qualified" across groups (the post-hire-success data shows base rates are roughly equal but the model's behavior is to over-select males at a TPR-disparate rate). Predictive parity rejected as primary target because it does not address the selection-rate disparity that drives candidate-pipeline harm.
• Tradeoffs: The chosen combination increases shortlist size for female and non-binary candidates (with corresponding marginal-quality concerns the model itself can address through threshold adjustment); reduces shortlist size for male candidates at the margin; requires ongoing monitoring as the candidate pool shifts.
• Rejected alternatives: Pure demographic parity (would surface less-qualified candidates from under-represented groups, harming both candidates and employer). Pure predictive parity (would ignore the selection-rate disparity that drives broader pipeline harm).
• Article 27 FRIA integration: The fairness-metric-selection memo is appended to the FRIA Section 4 (specific risks of harm) and Section 5 (description of human oversight measures).
• Article 10 data-governance integration: The memo references the Article 10(1) bias examination methodology and the Article 10(5) special-category data carve-out for protected-attribute analysis (with GDPR Article 9 safeguards documented).
• U.S. state-law overlays: NYC LL 144 four-fifths-rule analysis (independent auditor); Texas TRAIGA intentional-discrimination review (intent-based standard satisfied because the model is not designed to discriminate); EEOC Title VII disparate-impact analysis (four-fifths rule failure flags adverse impact; business-necessity defense documented; less-discriminatory-alternative analysis documented).
• Refresh cadence: Quarterly recomputation; triggered updates on candidate-pool shifts, model updates, regulatory developments.
• Auditor reviewer: Annual independent bias audit per NYC LL 144 + Stage 2 ISO 42001 audit per A.7 + EU AI Act Article 27(3) notification to national supervisory authority.

The memo is the artifact the AI Governance Committee approves, the External Auditor reads, the General Counsel cites in any litigation defense, and the AI Officer references in regulator engagement. It is the policy-load-bearing artifact for the fairness work.

Six Common Fairness-Metric Mistakes

Mistake 1 - Pretending a Single Metric Solves the Problem

The impossibility theorem means no single metric satisfies all stakeholders. Treating the metric choice as a technical decision (e.g., "the vendor product tunes against demographic parity") rather than a policy decision (e.g., "we chose four-fifths + equalized odds because the harm analysis shows...") leaves the program legally and operationally weak.

Mistake 2 - Skipping the Base-Rate Analysis

The four metrics produce different outcomes when base rates differ. If your data does not establish base rates (e.g., post-hire success rates), you cannot meaningfully apply equalized odds or predictive parity. The base-rate analysis is foundational; without it, only demographic parity and four-fifths can be computed.

Mistake 3 - Treating the Metric Choice as Static

The fairness-metric-selection memo is living. Quarterly refresh plus triggered updates on candidate-pool shifts (e.g., a major workforce-demographic change), model updates, regulatory developments (e.g., new EEOC guidance, new NYC LL 144 enforcement posture), and litigation precedents.

Mistake 4 - Accepting Vendor Defaults Without Analysis

Many HR-tech vendors ship products tuned against one fairness metric (often demographic parity or four-fifths). Accepting the default without analyzing whether it is the right choice for the deployer's specific population and harm pattern leaves the deployer with a vendor-driven, not deployer-driven, fairness posture. The vendor's default is a starting point; the deployer's analysis is the policy decision.

Mistake 5 - Skipping Intersectional Analysis

Single-attribute fairness analysis (e.g., sex alone, race alone) can mask intersectional disparities (e.g., the model performs well on white women and black men but poorly on black women). Intersectional analysis where data permits is the audit-defensible practice. NYC LL 144 explicitly references intersectional analysis as best practice.

Mistake 6 - Failing to Document Rejected Alternatives

The fairness-metric-selection memo's most-cited section by auditors and litigators is the "rejected alternatives" section. Without explaining why demographic parity / equalized odds / predictive parity / four-fifths were considered and which were rejected and why, the chosen metric appears arbitrary. The documentation of rejected alternatives is what demonstrates the policy reasoning.

Key Takeaways

• The impossibility theorem (Kleinberg-Mullainathan-Raghavan 2017; Chouldechova 2017) is real and mathematically proved. When group base rates differ, demographic parity, equalized odds, and predictive parity cannot all be satisfied simultaneously. The choice of metric is a policy decision, not a technical decision.
• Four primary metrics with distinct meanings. Demographic parity (equal selection rates); equalized odds (equal TPR and FPR conditional on actual qualification); predictive parity (equal accuracy conditional on prediction); four-fifths rule (selection rate ratio ≥ 0.80).
• Each metric allocates harm differently. Demographic parity → over-represented groups; equalized odds → false-negative groups; predictive parity → under-selected groups; four-fifths → bounded disparity tolerance.
• EU AI Act Article 10 is the regulatory anchor. Article 10(1) bias examination; Article 10(3) statistical properties; Article 10(5) special-category data carve-out for bias monitoring.
• The fairness-metric-selection memo is the L1 artifact. Per-system document: chosen metric, rationale, tradeoffs, rejected alternatives, integration with Article 27 FRIA + Article 10 + U.S. state-law overlays, refresh cadence, audit reviewer.
• The "rejected alternatives" section is the most-cited by auditors and litigators. Without explaining why other metrics were considered and rejected, the chosen metric appears arbitrary.
• NYC LL 144 four-fifths rule is the U.S. bright-line litigation standard. EEOC Title VII disparate-impact framework also references the four-fifths rule. Audit-defensible programs use it as compliance threshold plus equalized odds as operational quality target where data permits.
• Intersectional analysis is best practice. Single-attribute analysis can mask intersectional disparities. NYC LL 144 references intersectional analysis explicitly.
• The fairness-metric-selection memo integrates with the broader L3 conformity package. Annex IV §2(d) data requirements, Article 27 FRIA, Article 10 bias examination, ISO 42001 A.7 data-governance evidence, NIST AI RMF Measure 2 fairness evaluation, OWASP / ATLAS red-team coverage where bias-driven attacks are in scope.
• The memo is a living document. Quarterly refresh + triggered updates on candidate-pool shifts, model updates, regulatory developments, litigation precedents.