Memory tricks that make bias, fairness and responsible AI click
From algorithmic bias to the alignment problem -- these memory tricks lock in the ethical frameworks, real-world cases, and regulatory landscape every AI student must understand.
Or continue to the sub-topics below for more specialized Study Rooms and Forums
AI Ethics and Society
Memory Tricks
Proven Mnemonics & Acronyms — fast to learn, hard to forget.
FATE Framework
FATE plus BFT -- FATE is the framework, BFT is the robot: Big Friendly Transparent
FAIRNESS AND ACCOUNTABILITY AND TRANSPARENCY AND ETHICS
Picture a Big Friendly Transparent robot -- that is what responsible AI looks like
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.
Fairness
No discrimination based on race, gender, age, disability
GARBAGE IN, GARBAGE OUT -- biased data produces biased models no matter how sophisticated
BIAS ENTERS THROUGH DATA AND LABELS AND FEATURES AND FEEDBACK LOOPS
COMPAS: falsely flagged Black defendants at twice the rate -- both bias claims were mathematically true
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).
Historical bias
Training data reflects past discrimination
Sampling bias
Training data does not represent deployment population
Feedback loop
Biased predictions lead to biased future training data
BLACK BOX vs GLASS BOX -- powerful but unexplainable vs interpretable but less accurate
LIME AND SHAP AND SALIENCY MAPS -- TOOLS TO EXPLAIN ANY MODEL
GDPR Article 22: right to explanation for automated decisions that significantly affect individuals
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.
LIME
Perturb locally, fit simple model -- explains individual predictions
SHAP
Shapley values -- feature contributions, model-agnostic, industry standard
Saliency maps
Highlight influential pixels in image classification
GDPR Article 22
Right to explanation for consequential automated decisions
AI Safety
ALIGNMENT PROBLEM -- ensuring AI pursues goals humans actually want, not proxy goals
REWARD HACKING AND SPECIFICATION GAMING AND GOODHART'S LAW
When a measure becomes a target it ceases to be a good measure -- Goodhart's Law
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.
Reward hacking
AI maximizes reward metric via unintended means
Specification gaming
Satisfies letter but not spirit of objective
Goodhart's Law
Optimize for measure and it stops being a good measure
Constitutional AI
AI evaluates own outputs against principles -- Anthropic's approach
Hallucination
LLMs generate PLAUSIBLE next tokens, not VERIFIED facts -- confident wrongness is a feature, not a bug
FACTUAL AND ATTRIBUTION AND REASONING ERRORS -- FUNDAMENTAL NOT A BUG
Lawyer cited fake ChatGPT-generated case citations to federal court and received sanctions
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.
Retrieve real documents, inject as context -- grounds the answer
Cannot be eliminated
Fundamental to token prediction -- only mitigated
Privacy and AI
DIFFERENTIAL PRIVACY -- add calibrated noise so individual data cannot be inferred from model outputs
MEMORIZATION AND MEMBERSHIP INFERENCE AND FEDERATED LEARNING
Machine unlearning: GDPR right to erasure creates legal obligation to remove training data
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.
Memorization
LLMs can regurgitate private training data verbatim
Membership inference
Determine if specific data was in the training set
Federated learning
Gradients only shared -- data stays on device (Apple, Google)
Differential privacy
Mathematical guarantee individual cannot be identified from output
AI Regulation
EU leads with rules -- US leads with principles -- China leads with national strategy
EU AI ACT (2024) IS THE FIRST COMPREHENSIVE AI REGULATION GLOBALLY
EU AI Act fines: up to 35 million euros or 7% of global annual revenue
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.
Unacceptable risk
Banned -- social scoring, real-time biometric surveillance in public
High risk
Must audit, register, document -- hiring, healthcare, criminal justice
Limited risk
Chatbots must disclose they are AI
Fines
Up to 35M EUR or 7% of global revenue for violations
Societal Impact
JOBS -- some lost, some changed, new ones created -- AI transforms labor, not just replaces it
DISPLACEMENT AND POWER CONCENTRATION AND MISINFORMATION AND SURVEILLANCE
GPT-3 training consumed approximately 500 tons of CO2 equivalent -- equivalent to 60 transatlantic flights
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.
Labor displacement
Routine cognitive tasks increasingly automated
Power concentration
Compute + data + talent concentrated in few large companies
Deepfakes
AI-generated synthetic media -- video, audio, images
Environmental cost
GPT-3 training: ~500 tons CO2 -- inference at scale adds up
Responsible AI
HUMAN in the LOOP -- keep humans accountable for high-stakes AI decisions
HUMAN-IN-LOOP AND HUMAN-ON-LOOP AND HUMAN-OUT-OF-LOOP -- THREE AUTOMATION LEVELS
Model cards: standardized documentation of capabilities, limitations, and performance across demographic groups
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.
HITL
Human approves every decision -- criminal sentencing, medical diagnosis
HOTL
AI acts, human monitors and can intervene
Model cards
Document capabilities, limitations, and demographic performance
Algorithmic auditing
Third-party evaluation for bias, safety, accuracy
AI in Healthcare
DOCTOR plus AI = better outcomes -- but AI errors in medicine can kill
FDA CLEARS 500+ AI MEDICAL DEVICES -- AI AUGMENTS PHYSICIANS NOT REPLACES THEM
AI plus physician consistently outperforms either alone -- augmentation, not replacement
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.
Fails at hospitals different from training hospitals
FDA clearance
Required -- 510(k) for AI medical devices
AI + physician
Consistently outperforms either alone -- augment, not replace
Copyright and AI
TRAINING on copyrighted data plus GENERATING similar content = active legal battleground
WHO OWNS AI-GENERATED CONTENT -- COURTS ARE STILL DECIDING IN 2025
NYT vs OpenAI and Getty vs Stability AI and major record labels vs Suno and Udio -- all active
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.
Training data
Fair use? NYT vs OpenAI, Getty vs Stability AI still pending
AI-generated content
US Copyright Office: requires human authorship -- AI output uncopyrightable
Licensing deals
OpenAI with AP, Axel Springer -- growing trend
C2PA watermarking
Content authenticity standard -- identifies AI-generated content
🎯 Exam Favorite
ALGORITHMIC BIAS = GARBAGE IN, PREJUDICE OUT — biased training data bakes discrimination into the model
DATA REFLECTS HISTORY · MODEL AMPLIFIES IT
Where AI bias comes from — and why it's hard to fix
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.
🧠 Vivid Story
BLACK BOX = You see the INPUT and OUTPUT but the middle is a MYSTERY — even to its creators
EXPLAINABILITY vs PERFORMANCE TRADEOFF
The black box problem — why explainability matters
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.
🔑 Key Distinction
PRIVACY BY DESIGN = Build the LOCK before building the HOUSE — not after
DATA MINIMIZATION · ANONYMIZATION · CONSENT
AI and privacy — the three core principles
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.
💡 Concept Anchor
DEEPFAKE = AI-generated media so realistic it can PUT WORDS IN ANYONE'S MOUTH
SYNTHETIC MEDIA · MISINFORMATION · TRUST EROSION
Deepfakes — the AI ethics challenge of synthetic media
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.
📅 Quick Reference
AI ALIGNMENT = Teaching the robot to WANT what we ACTUALLY want — not what we literally said
AI alignment — the hardest long-term safety problem
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.
⭐ Most Important
DISPARATE IMPACT = The algorithm treats groups differently even without intending to
PROXIES CARRY BIAS — ZIP CODE IS RACE IN DISGUISE
How algorithmic bias hides in seemingly neutral features
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.
Direct discrimination
Explicit use of protected attribute — illegal and detectable
Proxy discrimination
Neutral feature (ZIP, name) that correlates with protected attribute
Audit method
Test model outputs separately for each demographic group — measure gaps
🐍 Code
# Fairness audit — check outcomes by group import pandas as pd results = df.groupby("demographic_group").agg( approval_rate=("prediction", "mean"), count=("prediction", "count") ).round(3) print(results) # Flag if groups differ significantly
🎯 Exam Favorite
EXPLAINABLE AI = SHAP values show WHICH features pushed the prediction UP or DOWN
SHAP · LIME · ATTENTION WEIGHTS — three XAI methods
Explainable AI methods — making black boxes talk
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.
SHAP
Feature contribution per prediction — gold standard for tabular data
LIME
Local approximation with interpretable model — works on any model
Attention
Visualize which tokens influenced which — useful for transformers
🐍 Code
import shap explainer = shap.TreeExplainer(model) # or shap.Explainer(model) shap_values = explainer.shap_values(X_test) shap.summary_plot(shap_values, X_test) # global feature importance shap.waterfall_plot(shap_values[0]) # single prediction explanation
🔑 Key Distinction
EU AI ACT tiers by RISK — Unacceptable · High · Limited · Minimal
FOUR RISK TIERS · BAN TO TRANSPARENCY
The EU AI Act — the world's first comprehensive AI law
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.
Unacceptable
Banned — social scoring, real-time biometric surveillance, manipulation
High risk
Strict rules — medical, hiring, credit, law enforcement, education
Limited risk
Transparency only — disclose AI is involved
Minimal risk
No regulation — spam filter, video games, product recommendations
🐍 Code
# EU AI Act compliance checklist (high-risk systems): # 1. Risk management system documented # 2. Training data quality and bias documentation # 3. Human oversight mechanisms in place # 4. Accuracy, robustness, cybersecurity measures # 5. Transparency and user information provided
💡 Concept Anchor
CONSENT · PURPOSE LIMITATION · DATA MINIMIZATION — the three pillars of AI privacy law
GDPR + CCPA + AI = PRIVACY BY DEFAULT
Privacy regulations every AI practitioner must know
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).
Consent
Must be freely given, specific, informed, and withdrawable
Purpose limitation
Cannot repurpose data without new consent
Data minimization
Collect only what you need — not everything you can
Right to erasure
Tricky for ML — removing a person from model weights is an open problem
🐍 Code
# Federated learning — privacy-preserving ML # Train locally, share only gradients (not raw data) # Model learns from distributed devices without data ever leaving # Used by: Google (Gboard), Apple (Siri), hospitals (medical AI) # from flwr import client, server # Flower federated learning framework
0
Correct
0
Missed
0
Remaining
What does this mean / stand for?
0
Correct
0
Wrong
0
Remaining
🔗 Related Sub-Subjects
🧠 AI Basics
The foundational AI concepts and capabilities that raise ethical questions.
Q: What are the three main sources of bias in AI systems?
A: (1) Training data bias — historical discrimination baked into labeled data (hiring records, loan decisions, criminal justice data). (2) Label bias — human annotators bring their own biases when labeling data. (3) Feedback loop bias — biased model outputs influence real-world outcomes which become future training data, amplifying the original bias. Example: biased recidivism model → more arrests in certain communities → more "evidence" of criminality → model becomes more biased.
Q: What is the difference between fairness and accuracy in AI?
A: These can conflict: a model maximizing overall accuracy may do so by performing poorly on minority groups with less training data. Different fairness definitions include: demographic parity (equal positive prediction rates across groups), equalized odds (equal TPR and FPR across groups), individual fairness (similar individuals treated similarly). These fairness criteria are mathematically incompatible — you cannot simultaneously satisfy all of them. The choice of which to optimize is an ethical and policy decision, not a purely technical one.
Q: What is AI hallucination and why is it a safety concern?
A: Hallucination is when an LLM confidently generates factually incorrect information that sounds plausible. It occurs because LLMs predict the most probable next token — not the most accurate. Safety concerns: medical advice with false drug interactions, legal research citing nonexistent cases, scientific summaries with fabricated citations. Mitigation: RAG (grounding in retrieved facts), citations with source linking, confidence calibration, human review for high-stakes outputs.
Q: What is the alignment problem in AI safety?
A: The alignment problem is ensuring AI systems pursue goals that humans actually intend, not proxy goals that satisfy the letter but not the spirit of instructions. Classic example: "maximize user engagement" → algorithm learns outrage and misinformation maximize clicks. Reward hacking: AI finds unexpected ways to maximize reward that violate the intent. As AI systems become more capable, misalignment risks become more severe. Current approaches: RLHF, constitutional AI, interpretability research, and red-teaming.
Q: What is differential privacy and how does it protect individuals in ML training?
A: Differential privacy (DP) provides mathematical guarantees that an individual's data cannot be inferred from a trained model. It works by adding carefully calibrated random noise to model training (DP-SGD) or to query results. The privacy budget (epsilon) controls the tradeoff: smaller epsilon = stronger privacy = lower accuracy. Used by Apple (iOS data collection), Google (Chrome), and the US Census. Allows training on sensitive data (medical, financial) with provable privacy guarantees.