Sources of Bias in the ML Pipeline: 1. Historical bias Data reflects past discrimination (e.g., fewer women in tech hiring data) 2. Representation bias Training data doesn’t represent all groups (e.g., facial recognition trained mostly on light-skinned faces) 3. Measurement bias Proxy variables encode protected attributes (e.g., zip code correlates with race) 4. Aggregation bias One model for diverse populations (e.g., medical AI trained on one ethnicity) 5. Evaluation bias Benchmarks don’t test for fairness (e.g., accuracy is high overall but low for minority subgroups) 6. Deployment bias System used outside intended context

COMPAS (Criminal Justice, 2016): Predicted recidivism risk for sentencing ProPublica found: Black defendants falsely flagged as high-risk at 2x the rate of white defendants. White defendants falsely flagged as low-risk at higher rates. Used in courts across the US Amazon Hiring Tool (2014–2018): Trained on 10 years of resumes (mostly male) Penalized resumes with “women’s” (e.g., “women’s chess club captain”) Downgraded all-women’s college graduates Amazon scrapped the project entirely Healthcare Algorithm (2019): Used health spending as proxy for health need Black patients spend less due to systemic barriers → algorithm rated them as healthier than equally sick white patients Affected ~200 million patients (Obermeyer)

Facial Recognition (Buolamwini, 2018): “Gender Shades” study tested 3 commercial systems (Microsoft, IBM, Face++) Error rates: Light-skinned males: 0.8% Dark-skinned females: 34.7% 43x higher error for dark-skinned women Google Photos (2015): Auto-tagged Black users as “gorillas” Google’s fix: removed the “gorilla” label Still not properly fixed years later LLM Bias (Ongoing): GPT models associate certain professions with specific genders/races “The doctor... he” vs “The nurse... she” Stereotypes embedded in training data

Demographic Parity: P(Ŷ=1 | A=0) = P(Ŷ=1 | A=1) Equal selection rates across groups “Hire 50% from each group” Equalized Odds: P(Ŷ=1 | Y=1, A=0) = P(Ŷ=1 | Y=1, A=1) P(Ŷ=1 | Y=0, A=0) = P(Ŷ=1 | Y=0, A=1) Equal TPR and FPR across groups “Same accuracy for everyone” Calibration: P(Y=1 | Ŷ=p, A=0) = P(Y=1 | Ŷ=p, A=1) = p When model says 70% risk, it’s 70% for all “Scores mean the same thing” Individual Fairness: Similar individuals get similar predictions “Treat like cases alike”

Explainability Methods: LIME (Local Interpretable Model Explanations) Perturb input, observe output changes Fit simple model locally to explain one prediction SHAP (SHapley Additive exPlanations) Game-theoretic feature attribution Each feature gets a contribution score Mathematically grounded (Shapley values) Attention Visualization Show which tokens/pixels the model attends to Caution: attention ≠ explanation Concept-Based Explanations TCAV: test with human-understandable concepts “This image was classified as doctor because of the stethoscope concept”

EU AI Act (Aug 2024, fully effective 2027): Risk-based classification: UNACCEPTABLE RISK (banned): Social scoring by governments Real-time facial recognition (most cases) Manipulative subliminal techniques Exploitation of vulnerable groups HIGH RISK (strict requirements): Hiring & employment decisions Credit scoring & insurance Criminal justice & law enforcement Education & access to services Critical infrastructure LIMITED RISK (transparency): Chatbots must disclose they are AI Deepfakes must be labeled Emotion recognition must be disclosed MINIMAL RISK (no requirements): Spam filters, video games, etc.

NIST AI Risk Management Framework: Govern → Map → Measure → Manage Voluntary, US-based guidance GPAI (General-Purpose AI) Rules: Providers must publish training data summary Implement copyright compliance policies Systemic risk models (>10²³ FLOPs): Cybersecurity protections Incident reporting Adversarial testing Global Landscape: EU: Comprehensive regulation US: Executive orders + sector-specific UK: Pro-innovation, light-touch China: Algorithm regulation + content rules G7: Hiroshima AI Process (voluntary)

Current Safety Techniques: RLHF / DPO: Align with human preferences Constitutional AI: Self-critique against rules Red teaming: Adversarial testing by humans Guardrails: Input/output filters Sandboxing: Limit agent capabilities Open Problems: Scalable oversight: How do humans supervise AI smarter than them? Deceptive alignment: Could AI appear aligned during training but pursue different goals when deployed? Goal stability: Do goals remain stable as capabilities grow? Emergent capabilities: New abilities appear unpredictably at scale

Labor Market Impact: At risk: Data entry, translation, basic coding, customer service, content moderation, paralegal work, bookkeeping Augmented: Doctors, lawyers, engineers, designers, researchers — AI as copilot New roles: Prompt engineers, AI trainers, alignment researchers, AI auditors Copyright & IP: Trained on copyrighted data without consent NYT v. OpenAI (2023): landmark lawsuit Getty v. Stability AI: image copyright EU AI Act: must disclose training data No consensus on AI-generated content ownership Environmental Cost: GPT-4 training: ~50 GWh estimated One ChatGPT query: ~10x a Google search Data centers: 2–3% of global electricity Water usage for cooling: billions of liters

Responsible AI Principles: 1. Fairness: Test for disparate impact across all subgroups 2. Transparency: Explain decisions to affected individuals 3. Privacy: Minimize data collection, differential privacy 4. Safety: Red team, monitor, have kill switches 5. Accountability: Human oversight for high-stakes decisions 6. Inclusivity: Diverse teams, diverse data, diverse evaluation Practical Checklist: □ Audit training data for representation □ Disaggregate metrics by subgroup □ Use fairness toolkits (Fairlearn, AIF360) □ Document model cards & datasheets □ Establish human-in-the-loop for high stakes □ Monitor for drift and emerging bias □ Create incident response plans