AI Bias Assessment for Special Category Data

Overview

AI systems can amplify, perpetuate, or introduce bias against protected groups defined by GDPR Art. 9 special categories (race, ethnicity, political opinion, religion, trade union membership, genetic data, biometric data, health, sexual orientation) and by EU equality law (gender, age, disability). The AI Act Art. 10 requires data governance practices for training data that address bias, while Art. 5 prohibits AI-based social scoring. This skill provides the methodology for detecting, measuring, and mitigating bias in AI systems that process or infer special category data, with documentation requirements meeting both GDPR and AI Act obligations.

Art. 9 Special Categories and AI Bias

Direct Processing of Special Category Data

When AI systems directly process Art. 9 data:

Category	AI Bias Risk	Example
Racial or ethnic origin	Discrimination in hiring, credit, policing	CV screening penalising names associated with ethnic minorities
Political opinions	Political profiling, content suppression	News recommendation amplifying or suppressing political viewpoints
Religious beliefs	Service denial, discriminatory targeting	Insurance pricing varying by religious affiliation
Trade union membership	Employment discrimination	Performance scoring penalising union activity
Genetic data	Genetic discrimination in insurance/employment	Health insurance pricing based on genetic predisposition
Biometric data	Differential accuracy across demographics	Facial recognition with higher error rates for darker skin tones
Health data	Health-based discrimination	Hiring algorithms penalising disability or mental health history
Sexual orientation	Discrimination, outing	Content recommendation inadvertently revealing sexual orientation

Proxy Inference of Special Categories

AI models frequently infer Art. 9 data from non-sensitive features:

Proxy Feature	Inferred Category	Mechanism
Postcode/zip code	Race/ethnicity, income	Residential segregation patterns
First/last name	Race/ethnicity, religion	Name-ethnicity correlations
Browsing history	Political opinion, religion, health	Content consumption patterns
Purchase history	Health status, religion	Medication purchases, dietary products
Language patterns	National origin, education	Dialect, vocabulary, grammar patterns
Device/app usage	Age, income, disability	Accessibility features, device type

EDPB position: inferring Art. 9 data from non-sensitive inputs constitutes processing of special category data — the same protections apply.

Fairness Metrics

Group Fairness Metrics

Metric	Definition	When to Use
Demographic parity	P(positive outcome	group A) = P(positive outcome
Equalized odds	TPR and FPR equal across groups	When accuracy should be equal across groups
Equal opportunity	TPR equal across groups (relaxed equalized odds)	When true positive detection should be equal
Calibration	P(Y=1	score=s, group=A) = P(Y=1
Predictive parity	PPV equal across groups	When positive predictions should be equally reliable

Individual Fairness Metrics

Metric	Definition
Consistency	Similar individuals receive similar outcomes
Counterfactual fairness	Outcome would be the same if protected attribute were different
Causal fairness	No causal path from protected attribute to outcome

Metric Selection Guidance

Decision Context	Recommended Metric	Justification
Hiring/admissions	Equalized odds or equal opportunity	Equal detection of qualified candidates across groups
Credit scoring	Calibration	Score should mean the same probability regardless of group
Criminal risk	Equalized odds	Both FPR and TPR should be equal to avoid disproportionate impact
Healthcare	Equal opportunity + calibration	Equal detection of conditions; equal meaning of risk scores
Content moderation	Demographic parity	Content removal should not disproportionately affect groups

Note: Mathematical impossibility results show that demographic parity, equalized odds, and calibration cannot all be satisfied simultaneously when base rates differ across groups. Document the trade-off explicitly.

Bias Detection Methodology

Phase 1: Data Audit

1. Profile training data demographics against target population
2. Identify underrepresented groups that may have lower model performance
3. Check for historical bias in labelled data (e.g., biased hiring decisions used as ground truth)
4. Identify label noise differential across groups
5. Assess feature distributions across groups for proxy discrimination potential

Phase 2: Model Testing

1. Evaluate model performance disaggregated by protected groups
2. Compute selected fairness metrics for each protected group pair
3. Run counterfactual testing: change protected attributes, observe output changes
4. Test for intersectional bias (combinations of protected attributes)
5. Evaluate model performance on edge cases and adversarial examples per group

Phase 3: Output Analysis

1. Analyse decision distribution across groups
2. Identify threshold effects that disproportionately impact specific groups
3. Test for feedback loop amplification over time
4. Assess output explanation fairness (are explanations equally informative across groups?)

Bias Mitigation Strategies

Pre-processing (Data-Level)

Strategy	Description	Trade-off
Resampling	Over-sample underrepresented groups, under-sample overrepresented	May reduce data diversity or introduce duplicates
Reweighting	Assign higher weights to underrepresented group samples	Computationally simple; may not address structural bias
Relabelling	Correct historically biased labels	Requires domain expertise; may be subjective
Fair representation learning	Learn latent representation that removes protected attribute information	May lose legitimate correlations

In-processing (Algorithm-Level)

Strategy	Description	Trade-off
Adversarial debiasing	Train adversary to predict protected attribute from model; penalise success	Accuracy-fairness trade-off; requires protected attribute data
Fairness constraints	Add fairness metric as training constraint	May reduce overall accuracy; constraint satisfaction varies
Regularisation	Add fairness-related regularisation term to loss function	Balances accuracy and fairness; requires tuning
Causal modelling	Use causal graph to block discriminatory paths	Requires causal knowledge; complex to implement

Post-processing (Output-Level)

Strategy	Description	Trade-off
Threshold adjustment	Different decision thresholds per group to equalise metrics	May be perceived as unfair; legally complex
Score calibration	Calibrate scores per group	Requires sufficient group data; may reduce discrimination
Reject option	Abstain from decision for borderline cases across groups	Reduces coverage; requires human fallback

AI Act Art. 10 Data Governance

Art. 10 requires for high-risk AI training data:

Requirement	Implementation
Relevant data	Training data must be relevant to the intended purpose
Sufficiently representative	Data must represent the population the system will be deployed on
Free of errors	Data quality assessment and cleaning processes
Complete	Sufficient coverage of deployment scenarios
Appropriate statistical properties	Distribution analysis including demographic representation
Bias examination	Examine training data for possible biases, especially related to Art. 10(2)(f)

Art. 10(5): Processing of special category data for bias detection is permitted for high-risk AI if:

• Strictly necessary for bias monitoring, detection, and correction
• Subject to appropriate safeguards (pseudonymisation, access controls)
• Data is not used for other purposes
• Deleted after bias assessment unless retention is required for compliance documentation

Documentation Requirements

Bias Assessment Report

Section	Content
System description	Model, purpose, affected groups
Protected attributes assessed	Art. 9 categories + equality law characteristics
Fairness metrics selected	With justification for selection
Data audit results	Training data demographics, representation gaps
Model testing results	Per-group performance, fairness metrics, counterfactual results
Bias findings	Identified disparities with severity assessment
Mitigation measures	Applied strategies with effectiveness evidence
Residual bias	Remaining disparities after mitigation
Trade-off documentation	Accuracy-fairness trade-offs, metric impossibility acknowledgement
Ongoing monitoring plan	Post-deployment fairness monitoring

Enforcement Precedents

• Dutch Tax Authority (SyRI, 2020): Court struck down algorithmic fraud detection for discriminatory profiling — system disproportionately targeted residents of low-income, immigrant-background neighbourhoods.
• Italian DPA v. Deliveroo (2021): Algorithmic worker management found to discriminate based on protected characteristics — Art. 22 and equality law violations.
• Austrian DPA v. AMS Algorithm (2020): Austrian employment service algorithm that scored job seekers lower based on gender, age, disability, and citizenship — DPA found processing unlawful.
• AEPD guidance (2021): Spanish DPA issued guidance on algorithmic discrimination, emphasising DPIA requirement for AI systems processing protected characteristics.
• CNIL (2024): AI bias assessment framework published requiring fairness testing for high-impact AI systems.

Integration Points

• ai-dpia: Bias assessment feeds into DPIA risk assessment for discrimination harms
• ai-automated-decisions: Biased automated decisions trigger Art. 22 and equality law violations
• ai-training-lawfulness: Art. 10(5) processing of special category data for bias detection requires documentation
• ai-transparency-reqs: Bias findings should be disclosed in model cards and transparency documentation
• ai-act-high-risk-docs: Art. 10 data governance documentation is part of conformity assessment