Appendix G: Quick Reference Cards

Rapid Lookup Reference for AI Ethics Practitioners

Reference Card 1: Fairness Metrics Cheat Sheet

The Six Core Fairness Metrics

Metric 1: Demographic Parity (Statistical Parity / Group Fairness)

Definition: The proportion of individuals receiving a positive outcome is equal across demographic groups.

In Plain Language: The approval rate should be the same for Group A and Group B — regardless of other characteristics.

When to Use: - When the goal is proportional representation in outcomes - Anti-discrimination law contexts where equal selection rates are required - When base rates of the underlying characteristic are similar across groups - When there is reason to believe historical data encodes discrimination that should not be perpetuated

Key Limitation: - Does not account for legitimate differences in qualifications or risk between groups - Can require approving less-qualified applicants from one group to match approval rates - Mathematically: satisfying demographic parity is incompatible with predictive parity when base rates differ

Example: A hiring algorithm that screens 1,000 male applicants and 1,000 female applicants should, under demographic parity, select the same percentage of each group. If it selects 30% of men but only 15% of women, it violates demographic parity.

Metric 2: Equal Opportunity

Definition: The true positive rate (recall, sensitivity) is equal across demographic groups — among those who deserve a positive outcome, the same proportion of each group receives it.

In Plain Language: Among qualified applicants, the acceptance rate should be the same for all groups.

When to Use: - When the cost of false negatives (missing qualified candidates) falls unequally on groups - Employment and lending contexts where deserving individuals are the focus - When you want to ensure that merit is rewarded equally across groups

Key Limitation: - Only addresses false negatives (people who should be accepted but are rejected); does not address false positives - Requires reliable ground-truth labels for who "deserves" a positive outcome, which may themselves be biased

Example: Among loan applicants who would repay their loans if approved, the same proportion of Black and white applicants should receive loan approval. If 80% of creditworthy white applicants are approved but only 55% of creditworthy Black applicants are approved, equal opportunity is violated.

Metric 3: Equalized Odds

Definition: Both the true positive rate AND the false positive rate are equal across demographic groups.

In Plain Language: Among people who deserve a positive outcome, equal proportions get one; among people who do not deserve a positive outcome, equal proportions still get one incorrectly.

When to Use: - When both types of errors (false positives and false negatives) have significant consequences - Criminal justice contexts where both wrongful conviction (false positive) and wrongful acquittal (false negative) matter - The most stringent standard for error-rate-based fairness

Key Limitation: - Impossible to satisfy simultaneously with calibration when base rates differ (the Chouldechova impossibility result) - Requires ground truth about who "deserves" a positive outcome

Example: If COMPAS assigns a "high risk" score to a defendant, equalized odds requires that (1) among defendants who will reoffend, equal proportions of Black and white defendants get the "high risk" score; AND (2) among defendants who will not reoffend, equal proportions of Black and white defendants are falsely labeled "high risk."

Metric 4: Predictive Parity (Calibration / Test Fairness)

Definition: Among individuals who receive the same predicted score or decision, the actual outcome rate is the same across demographic groups.

In Plain Language: When the model says someone has a 40% risk of X, that 40% prediction should be equally accurate for all demographic groups.

When to Use: - When the score is used to communicate a probability (e.g., recidivism risk, credit risk) - When decision-makers rely on the score's literal meaning, not just its ranking - The standard emphasized by Northpointe in the COMPAS controversy

Key Limitation: - Can be satisfied while still producing racially disparate false positive rates (as in the COMPAS case) - Impossible to simultaneously satisfy with equalized odds when base rates differ across groups

Example: If a recidivism algorithm assigns a score of 7/10 to Black and white defendants, and among those scored 7/10 exactly 70% of both groups reoffend, the algorithm is calibrated even if it assigns the 7/10 score to very different proportions of Black and white defendants.

Metric 5: Individual Fairness

Definition: Similar individuals should receive similar predictions or outcomes — regardless of their demographic group.

In Plain Language: Two people who are alike in all relevant ways should be treated alike.

When to Use: - When fairness is best conceptualized at the individual level - When group-level metrics are insufficient to capture inequities - Theoretical counterpoint to group-based metrics

Key Limitation: - Requires defining a meaningful notion of "similar" individuals — a hard problem that can hide contested value judgments - Does not guarantee equal treatment across demographic groups at the population level - Often difficult to operationalize in practice

Example: Two loan applicants with the same credit score, income, and debt-to-income ratio should receive the same lending decision regardless of their race or gender.

Metric 6: Counterfactual Fairness

Definition: An individual's outcome would be the same in the counterfactual world where they belonged to a different demographic group, with all causally non-relevant characteristics unchanged.

In Plain Language: Changing only your race (or gender, etc.) should not change the model's decision about you.

When to Use: - Causal inference frameworks; when you want to isolate the direct effect of protected characteristics - Theoretically rigorous basis for assessing direct discrimination

Key Limitation: - Requires a causal model of how protected characteristics relate to other variables — which is itself contested - Difficult to operationalize; often theoretical rather than practical - Does not address disparate impact arising from variables causally downstream of race (e.g., neighborhood, education)

Example: If you applied for a loan and were rejected, you were treated counterfactually fairly if you would have received the same decision had you been of a different race with all non-race characteristics unchanged.

Reference Card 2: Ethical Frameworks at a Glance

Reference Card 3: EU AI Act Risk Tiers

The Four-Tier Risk Classification

TIER 1: UNACCEPTABLE RISK — PROHIBITED

These AI applications are prohibited entirely. No exceptions.

Key requirement: Absolute prohibition. No mitigation or safeguard can make these uses compliant.

TIER 2: HIGH RISK

High-risk AI systems are permitted subject to extensive mandatory requirements.

Categories of High-Risk AI: - Biometric identification and categorization - Critical infrastructure (energy, water, transport) - Education and vocational training - Employment and worker management (including hiring tools) - Essential private and public services (credit scoring, benefits eligibility) - Law enforcement (risk assessment, polygraphs, crime analytics) - Migration and border control - Administration of justice and democratic processes

Key Requirements: - Conformity assessment before deployment - Registration in EU AI Act database - Risk management system throughout lifecycle - Training data governance requirements - Technical documentation and logging - Transparency and provision of information to users - Human oversight measures - Robustness, accuracy, and cybersecurity requirements

Who Must Comply: Providers (who place AI on the market) and deployers (who use it in their business)

TIER 3: LIMITED RISK (TRANSPARENCY OBLIGATIONS)

These AI systems must meet transparency requirements but face no substantive restrictions.

Applies to: - Chatbots and conversational AI (must disclose AI identity) - Deepfakes (must label as AI-generated) - Emotion recognition systems (must notify users) - AI that generates or manipulates images, audio, video (must label)

Key Requirement: Users must be informed they are interacting with or viewing AI-generated content. No functional restrictions on design or use.

TIER 4: MINIMAL RISK

The vast majority of AI applications. No specific requirements under the AI Act.

Examples: Spam filters, AI-enabled video games, most recommendation systems, AI in manufacturing processes

Note: General-purpose AI models (including large language models) have their own distinct requirements under Title VIII of the Act, separate from this four-tier classification.

Reference Card 4: AI Bias Types (Suresh & Guttag Taxonomy)

Reference Card 5: Key Laws and Regulations

Reference Card 6: XAI Methods Comparison

These reference cards provide quick-lookup definitions. For deeper treatment of any concept, consult the relevant chapter of the textbook. For regulatory details, always verify current requirements with qualified legal counsel, as AI law is developing rapidly.