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(wrong dates, names), Attribution (fabricated citations and quotes), Reasoning errors (logical mistakes), Temporal (outdated info as current).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.

AI bias originates in training data that reflects historical discrimination. A hiring algorithm trained on past decisions inherits past biases against women or minorities. A facial recognition system trained mostly on lighter skin performs worse on darker skin. The model doesn't "intend" to discriminate — it simply learned patterns from biased data. Fixing bias requires: diverse training data, fairness metrics during evaluation, and ongoing auditing after deployment.

Deep learning models can have billions of parameters. When a model denies your loan application or flags you for fraud, it cannot explain why in human terms. This is the black box problem — high performance but zero transparency. Explainable AI (XAI) methods like LIME and SHAP try to approximate explanations. High-stakes decisions (medicine, law, hiring, credit) demand explainability. The tradeoff: the most powerful models (deep neural networks) are the least interpretable.

Privacy in AI means: Data Minimization (collect only what you absolutely need — not everything you can), Anonymization (remove or mask identifying information from training data), and Consent (people should know and agree to how their data is used). GDPR in Europe and CCPA in California codify these rights. Federated learning (train on device, never upload raw data) is a technical privacy solution. Privacy by design means building these protections in from day one, not bolting them on later.

Deepfakes use GANs or diffusion models to generate photorealistic fake videos, images, or audio of real people. The risks: political misinformation (fake speeches by world leaders), non-consensual intimate imagery, financial fraud (fake CEO voice calls), and erosion of trust in all media. Detection tools exist but lag behind generation quality. The societal challenge: when anyone can fake anything convincingly, how do we establish truth? Digital watermarking and content provenance are emerging solutions.

Alignment is ensuring AI systems pursue goals humans actually intend, not unintended shortcuts. Classic example: tell an AI to maximize paperclip production — a misaligned superintelligent AI converts all matter on Earth into paperclips (technically achieved the goal). Reward hacking: AI finds unexpected ways to maximize its reward that violate the spirit of the task. Corrigibility: can we correct or shut down a powerful AI that doesn't want to be corrected? These are the core challenges of long-term AI safety research.

Disparate impact: a facially neutral algorithm produces outcomes that disproportionately harm a protected group. Example: an algorithm uses ZIP code as a feature. ZIP code correlates strongly with race (due to historical redlining). The algorithm "learns" a racial proxy without ever seeing race explicitly. Other proxies: arrest record (policing bias), educational institution (socioeconomic status), name (ethnicity). Auditing for disparate impact requires testing model outputs across demographic groups, not just looking at the input features.

Three main XAI techniques: SHAP (SHapley Additive exPlanations) — assigns each feature a value showing its contribution to the prediction for a specific instance. Based on game theory. Model-agnostic, gold standard for tabular data. LIME (Local Interpretable Model-agnostic Explanations) — approximates the black-box model locally with a simple interpretable model. Attention weights — for transformer models, visualize which tokens attended to which. Required by EU AI Act for high-risk decisions.

The EU AI Act (2024) classifies AI systems by risk level: Unacceptable risk — banned entirely (social scoring by governments, real-time facial recognition in public spaces, subliminal manipulation). High risk — strict requirements (medical devices, hiring, credit, education, law enforcement, critical infrastructure). Limited risk — transparency obligations (chatbots must disclose they are AI, deepfakes must be labeled). Minimal risk — no regulation (spam filters, AI games). Applies to any AI used in the EU regardless of where built.

GDPR (EU) and CCPA (California) establish: Consent — data subjects must agree to how their data is used, with right to withdraw. Purpose limitation — data collected for one purpose cannot be reused for another without new consent. Data minimization — collect only what is necessary, no more. For AI specifically: right to explanation for automated decisions (GDPR Article 22), right to opt out of automated processing, right to erasure ("right to be forgotten" — complicated when data is baked into model weights).