AI Ethics Guidelines Cheat Sheet

AI ethics and regulatory compliance — EU AI Act risk categories, bias detection methods, fairness metrics, transparency requirements, and responsible AI developme.

Last Updated: May 1, 2025

EU AI Act Risk Categories

Risk Level	Definition	Requirements	Examples
Unacceptable	Banned outright	Prohibited from EU market	Social scoring, real-time biometric surveillance, emotion recognition at work
High-Risk	Must meet strict requirements	CE marking, conformity assessment, human oversight, transparency, accuracy, cybersecurity	Medical devices, hiring systems, credit scoring, law enforcement AI
Limited Risk	Transparency obligations only	Must inform users they're interacting with AI	Chatbots, emotion recognition, deepfake generators
Minimal Risk	No specific regulation	Voluntary codes of conduct encouraged	Spam filters, AI-powered video games, inventory management
General Purpose AI (GPAI)	Additional rules for foundation models	Technical documentation, copyright compliance, energy reporting, systemic risk assessment for models >10^25 FLOPs	GPT-4, Claude, Gemini, Llama 3

Bias Detection Methods

Disparate Impact Analysis

Compare outcomes across demographic groups: ratio < 0.8 indicates potential discrimination

Equalized Odds

Measure if false positive / false negative rates are equal across groups

Demographic Parity

Check if positive prediction rates are similar across groups — P(ŷ=1|A=a) ≈ P(ŷ=1|A=b)

Intersectional Analysis

Test bias at intersections (e.g., race × gender) not just single attributes

Counterfactual Testing

Change only the protected attribute in inputs, check if model output changes

Subgroup AUC

Compute model performance metrics separately for each demographic subgroup

Calibration by Group

Check if predicted probabilities mean the same thing across groups — P(Y=1|score=s, A=a) = P(Y=1|score=s, A=b)

Word Embedding Association Test (WEAT)

Measure bias in word embeddings — geometric distance between concept and attribute vectors

Fairness Metrics & Interventions

Item	Description
`Pre-processing`	Balance training data: oversample underrepresented groups, reweight samples, or generate synthetic data
`In-processing`	Add fairness constraints to model training: adversarial debiasing, fairness regularization, constrained optimization
`Post-processing`	Adjust model outputs: threshold optimization per group, equal opportunity calibration
`Fairness-Aware Metrics`	Use Fairlearn (Microsoft), AI Fairness 360 (IBM), or What-If Tool (Google) for automated fairness evaluation
`Data Cards`	Document training data: source, collection method, demographic composition, known gaps — Google's Data Cards framework
`Model Cards`	Document model: intended use, limitations, evaluation results, ethical considerations — pioneered by Google / Hugging Face
`Human Review`	Establish diverse review panels to evaluate model behavior in context — technical metrics alone miss contextual fairness
`Redress Mechanism`	Provide a clear process for users to contest AI decisions — required by GDPR and EU AI Act for high-risk systems

Transparency Requirements

Item	Description
`Explainability`	Users have a right to meaningful explanation of AI decisions — SHAP, LIME, and attention visualization help
`Opt-Out Rights`	Users must be able to opt out of automated decision-making for consequential decisions (GDPR Article 22)
`AI Disclosure`	Users must be informed when interacting with AI vs human — chatbots, voice assistants, customer service
`Audit Trail`	Maintain records of model versions, training data provenance, and decision rationale for regulatory audits
`Energy Reporting`	EU AI Act: report model training energy consumption — encourages efficient model development
`Copyright Transparency`	Disclose copyrighted training data — EU AI Act requirement to enable opt-out and compensation mechanisms
`Third-Party Audits`	Independent audits for high-risk AI systems — similar to financial audits for public companies
`Open Source Ethics`	Open-source models aren't exempt from ethics — consider dual-use risks when releasing powerful models publicly

Pro Tip: Start with transparency. Even if you can't fix every bias immediately, documenting model limitations, training data composition, and intended use builds trust and is increasingly required by law.