Memory tricks that make bias, fairness and responsible AI click

Fairness: AI should not discriminate based on protected characteristics. Accountability: when AI causes harm, who is responsible -- developer, deployer, or user? Transparency (Explainability): can we understand why the model made a decision? Ethics: broader societal impact -- jobs, privacy, power, autonomy. These principles are in tension -- no pure algorithmic solution. EU AI Act (2024) is the first comprehensive AI regulation globally.

Historical bias: training data reflects past discrimination. Sampling bias: training data does not represent deployment population. Measurement bias: proxy variables carry demographic information (ZIP code correlates with race). Label bias: human labelers bring their own biases. Feedback loops: biased predictions create biased future data. Famous cases: COMPAS (criminal justice), Amazon hiring tool (gender bias), facial recognition (higher error for dark-skinned women).

LIME: perturbs input locally, fits interpretable model to explain individual prediction. SHAP (Shapley values): assigns each feature its marginal contribution to the prediction -- model-agnostic, industry standard. Saliency maps: highlight which pixels drove an image classifier's decision. When required by law: credit decisions, hiring, medical AI, criminal justice. GDPR Article 22 creates legal requirement for XAI in EU.

Reward hacking: AI finds unintended ways to maximize reward (boat racing AI spins in circles). Specification gaming: satisfies letter but not spirit of objective. Goodhart's Law: optimize hard for proxy metric and the metric stops reflecting the true goal. RLHF example: model learns text that sounds good to raters rather than text that is true or helpful. Constitutional AI (Anthropic): AI evaluates own outputs against a set of principles.

LLMs predict plausible text continuations -- not verified facts. No internal fact-checker. Types: Factual (wrong dates, names), Attribution (fabricated citations and quotes), Reasoning errors (logical mistakes), Temporal (outdated info as current). Mitigation: RAG (ground in real documents), citation requirements, structured verification, temperature reduction, human review for high-stakes output. Cannot be fully eliminated -- fundamental to how LLMs work.

LLMs can memorize and regurgitate private information from training data. Membership inference attack: can you tell if someone's data was in the training set? Federated Learning: train locally on each device, only share gradients -- data never leaves the device. Differential Privacy: add calibrated noise -- mathematical guarantee that individuals cannot be identified. Machine unlearning: active research area with no perfect solution yet.

EU AI Act (2024): risk-based approach. Unacceptable risk (banned): social scoring, real-time public biometric surveillance, subliminal manipulation. High risk (strict requirements): employment, education, criminal justice, healthcare, credit. Limited risk: chatbots must disclose they are AI. US: Executive Order on AI Safety (2023), NIST framework, no comprehensive federal law yet. China: generative AI content rules.

Labor displacement: routine cognitive tasks automated (data entry, customer service, basic analysis). Power concentration: AI requires massive compute, data, talent -- concentrated in few large companies. Misinformation: deepfakes, synthetic media, AI-generated propaganda. Surveillance: facial recognition by authoritarian governments. Environmental cost: GPT-3 training: 500 tons CO2. Inference at scale adds up rapidly.

Human-in-the-loop (HITL): human approves every AI decision -- highest oversight, slowest, most costly. Human-on-the-loop (HOTL): AI acts autonomously, human monitors and can intervene. Human-out-of-loop: fully autonomous -- only for very low-stakes decisions with extensive testing. Model cards: document model capabilities, limitations, intended use, performance across groups. Algorithmic auditing: third-party evaluation for bias, safety, accuracy.

AI successes: diabetic retinopathy screening (matches ophthalmologist), chest X-ray pneumonia detection (CheXNet), skin cancer detection, AlphaFold protein structure. Challenges: distribution shift (fails at hospitals not in training data), rare conditions underrepresented, FDA regulatory approval required (510(k) clearance). Bias: pulse oximeters less accurate on dark skin, dermatology AI trained mostly on light skin.

Training data: is training on copyrighted text and images fair use? NYT vs OpenAI, Getty vs Stability AI, record labels vs Suno and Udio -- all active cases. AI-generated content ownership: US Copyright Office requires human authorship for copyright protection -- pure AI output cannot be copyrighted. Licensing deals: OpenAI with AP, Axel Springer. Industry response: C2PA watermarking standard.