Study it widely, question it closely,
reflect on it carefully, discern it clearly, and practice it faithfully.
—— The Doctrine of the Mean (pre-Qin classic); translation mine

---

This is the complete cheat sheet for Taiwan’s III Generative AI certification. It isn’t a textbook — it’s the fill-the-gaps card for the last hours before the exam: every term and study point that’s easy to miss or easy to confuse, each with a “see X, pick Y” hook.

Concept-level, no actual exam questions (those are copyrighted). If you’d rather first read how I prepared and passed two certs in two months, here’s the companion post: Two AI Certs in Two Months.

---

Part 1 — Full Key-Terms Reference (71 terms, five groups)

Each term has a ”🎯 how it’s used / how it’s tested.” Items marked [must-know / classic / easily-confused] are the highest-yield — learn to tell them apart.

Fundamentals & architecture

• RAG (Retrieval-Augmented Generation): retrieve from an external knowledge base first, then generate. 🎯 “how to solve stale knowledge or hallucination” → RAG, no retraining needed.
• Attention: lets the model weigh relationships between different positions in the input. 🎯 “why Transformers capture long-range dependencies.”
• Transformer: Encoder understands input, Decoder generates output. 🎯 GPT = Decoder-only, BERT = Encoder.
• RLHF (Reinforcement Learning from Human Feedback): fine-tune with human feedback to match human preferences. 🎯 why ChatGPT is more obedient/safe than raw GPT.
• Word2Vec: maps words to vectors, similar meanings close together. 🎯 static embedding (same word → same vector) ↔ contrast BERT’s contextual embedding.
• Embedding: turns text/images into vectors. 🎯 RAG / semantic search / recommendation all use it for similarity; it’s what a vector DB stores.
• Tokenizer / Token: splits text into tokens; a Chinese character ≈ 0.5–1 token. 🎯 API cost and context length are counted in tokens.
• CNN: traditional image-recognition architecture. 🎯 trap: “must images use CNN?” → no, Transformers can too and better.
• GAN: generator vs discriminator trained adversarially. 🎯 who proposed it → Ian Goodfellow; contrast with diffusion.
• Diffusion: generate images by progressively denoising random noise. 🎯 core of SD / Sora; “add noise then learn to denoise,” don’t confuse with GAN.
• Pre-training: learn general knowledge from massive unlabeled data. 🎯 three stages: pre-train (general) → fine-tune (domain/behavior) → RAG (external/live); most compute-heavy = pre-training.
• Fine-tuning: change model behavior or inject domain knowledge. 🎯 [most-tested distinction] tone/domain → fine-tuning; changing data → RAG.
• Overfitting: great on training, poor on test, no generalization. 🎯 fixes: Dropout, data augmentation, regularization.
• Temperature: controls output randomness/creativity. 🎯 raise for brainstorming, lower for precision; =0 always picks the top-probability word.
• Context Length: max tokens the model can process at once. 🎯 “forgets the original instruction late in a long chat” = exceeding it.
• Ground Truth: the correct answer/label. 🎯 the baseline for judging accuracy.
• Open Weights vs Open Source: open weights but restricted license ≠ strictly open source. 🎯 Llama counts as Open Weights.
• NLU / NLG / NLP: NLP splits into NLU (understanding) and NLG (generation). 🎯 a chatbot first NLU’s your meaning, then NLG’s a reply.
• Quantization: lower precision (32→8/4-bit) for smaller/faster. 🎯 “less memory, faster, slight accuracy loss” fits big models onto phones.
• Distillation: transfer a big model’s (Teacher) knowledge to a small one (Student). 🎯 see Teacher/Student → this.
• Hallucination: plausible-looking but incorrect/fabricated output. 🎯 “confidently wrong”; the reason for RAG / fact-checking.
• Zero/One/Few-shot: give zero / one / several examples. 🎯 if results are poor, upgrade Zero → Few-shot.
• Chain-of-Thought: have the model produce intermediate reasoning steps. 🎯 add “think step by step” for math/logic.
• MoE (Mixture of Experts): many experts, only some activated each time. 🎯 “huge params but controllable inference cost” (Grok-1, Mixtral, GPT-4).

Prompt & agent

• System vs User Prompt: System sets the long-term role/rules; User is the immediate instruction. 🎯 Persona/rules → System; everyday questions → User.
• Prompt Tuning / Soft Prompts: train only a few soft-prompt vectors, main weights untouched. 🎯 low-cost customization vs full fine-tuning.
• Agent / AI Agent: plans autonomously, uses tools, executes multi-step. 🎯 contrast with a chat-only Chatbot.
• Tool Use / Function Calling: lets the model call external tools/APIs. 🎯 live prices, math, email; a core Agent capability.
• LangChain / LlamaIndex: frameworks for building LLM apps. 🎯 LangChain is general glue; LlamaIndex specializes in private data.
• ReAct: reason and act (call tools/search) in alternation. 🎯 the classic “Reasoning + Acting” agent pattern.
• Multi-agent: multiple role-based agents collaborating (one writes, one tests, one documents).
• Jailbreak: bypass safety filters to produce banned content. 🎯 “breaking safety limits with wordplay” (grandma exploit).
• Prompt Injection: hide malicious instructions in the input to hijack the model. 🎯 [easily-confused] Jailbreak (break safety) vs Injection (inject malicious instructions).
• Self-Consistency: run inference several times, take the majority vote.

Tools & applications

• Text-to-Video: text directly generates video. 🎯 e.g. Sora; contrast Image-to-Video (needs a still image first).
• TTS / STT / S2S: text→speech / speech→text / speech-to-speech translation. 🎯 dubbing TTS, subtitles STT, live cross-lingual call S2S.
• ControlNet: adds control conditions to steer image composition. 🎯 OpenPose controls pose, Canny controls edges.
• Inpainting / Outpainting: local repaint / extend the canvas outward. 🎯 remove a passerby with Inpainting; extend the frame with Outpainting.
• Upscaling: enlarge and fill in detail. 🎯 “enlarge and intelligently add detail,” old photos to HD.
• Style Transfer: apply one image’s style to another’s content. 🎯 photo into Van Gogh style; don’t confuse with Inpainting/Outpainting.
• LoRA: train very few params to learn a specific character/style. 🎯 small file, learns a character/style; a form of PEFT.
• Voice Cloning / Voice Banking: clone a voice / bank a voice for those who may lose it. 🎯 medically, ALS patients bank their voice pre-op.
• Stable Diffusion / Midjourney: mainstream image generators. 🎯 Midjourney uses —ar for aspect ratio; SD pairs with ControlNet/LoRA.
• Suno / ElevenLabs: Suno generates songs; ElevenLabs does voice synthesis/cloning. 🎯 for a full vocal song use Suno.
• AutoML: auto-selects models and tunes hyperparameters. 🎯 “lets non-experts build models.”
• Deepfake: synthesized/face-swapped/voice-cloned fakery. 🎯 ethics must-know: positive (de-aging), negative (fraud); needs detection + transparency labeling.
• The three chatbot elements: NLU (understanding), Dialog management, NLG (generation). 🎯 distinguish Intent / Entity / context tracking.

Data governance & MLOps

• MLOps / Feature Store: centrally manage ML features so training/inference stay consistent. 🎯 “avoid mismatch between training and serving features.”
• Data Drift vs Concept Drift [classic must-know]: input distribution changes vs the input→output relationship/definition changes. 🎯 new words = Data Drift; “the definition of spam itself changed” = Concept Drift.
• Data Lineage: trace data from source → transformation → destination. 🎯 trace “where did this number come from” + compliance audit.
• PII: personally identifiable information (name, ID, address). 🎯 de-identify or encrypt before processing.
• Anonymization vs Pseudonymization [must-know]: anonymization is irreversible vs pseudonymization is reversible. 🎯 fully unrecoverable = anonymization; can map back to the person = pseudonymization.
• Differential Privacy: add mathematical noise so individuals can’t be recovered. 🎯 “add noise, mathematical guarantee” — publish stats without leaking anyone.
• Homomorphic Encryption: compute directly on encrypted data. 🎯 “operate on ciphertext directly.”
• Federated Learning: train locally, only send parameters back. 🎯 mnemonic “data stays put, the model moves” (phone keyboard prediction, hospital collaboration).
• SMPC (Secure Multi-Party Computation): parties compute jointly without revealing inputs. 🎯 banks computing together without showing each other their data.
• XAI (Explainable AI): make decisions transparent and understandable. 🎯 medical/finance/legal need to explain “why this decision”; global vs local.
• Red Teaming: experts play attacker to actively find holes. 🎯 the point is the “active attacker perspective.”
• Data Poisoning: inject malicious data into training so the model learns badly. 🎯 defense: verify data provenance.
• Model Inversion Attack: reconstruct sensitive training data from outputs. 🎯 privacy threat, e.g. reconstruct faces from a face-recognition model.
• Zero Trust: never trust, always verify. 🎯 verify every access regardless of source.

Ethics & law

• LAWS (Lethal Autonomous Weapons Systems): weapons that autonomously select and attack targets. 🎯 the classic military-AI controversy: international bans, humanitarian law, “humans must retain control.”
• Copyright / Fair Use: whether training on copyrighted works is fair use. 🎯 core of the AI-training infringement dispute; also who owns AI-generated content.
• GDPR / Machine Unlearning: right to be forgotten; very hard to precisely delete once in weights. 🎯 “user asks to delete data already trained into the model.”
• EU AI Act: EU’s AI law, risk-tiered into four levels. 🎯 world’s first comprehensive AI law: unacceptable / high / limited / minimal.
• CCPA / PIPL / Taiwan PDPA: California / China / Taiwan data laws. 🎯 Taiwan = PDPA, needs consent + security measures.
• Opt-in / Opt-out / Consent: managing user consent for personal data use. 🎯 “record and manage consent status.”
• Accountability: being answerable for AI decisions. 🎯 one of the six ethics principles, “trace responsibility when things go wrong.”
• Transparency: operation and decision logic are open and understandable. 🎯 contrast Accountability (answerability) vs Transparency (openness).
• Digital Divide: those lacking network/devices/literacy benefit less from AI. 🎯 extends to “digital colonialism,” “cultural homogenization.”
• MHC (Meaningful Human Control): humans retain understanding/judgment/intervention. 🎯 stronger than Human-in-the-loop — humans must truly understand and be able to change it.

---

Part 2 — Chapter-by-Chapter Study Points

Area 1 — Fundamentals

1.1 Basic concepts of generative AI

• Scaling Laws: more compute/data/parameters → predictably better performance.
• Token: a Chinese character is roughly 0.5–1 token, and it’s also the billing unit.
• Temperature (controls randomness), Top-P (nucleus sampling) (sample from the smallest set of words whose cumulative probability reaches P); results vary because of probabilistic sampling.
• Hallucination: plausible-looking but incorrect/fabricated output.
• GAN (generator vs discriminator, Goodfellow 2014) vs diffusion models (add noise to destroy → learn to reverse and denoise).
• Term to know: vertical AI = trained and optimized for a specific industry.

1.2 Application domains

• Whole-chapter theme: from “automation” to “augmentation and collaboration” — boosting professionals, not wholesale replacement.
• Academic norms: AI may assist with polishing/ideation, may not ghost-write, and must be disclosed.
• Human-AI collaboration: fiction = human sets premise/outline, AI expands; translation = translator shifts to post-editing.
• Proof of Concept (PoC): test feasibility at small scale before deciding to scale up.
• High-frequency ethics: resume screening can discriminate; therapy AI most fears giving wrong medical advice / mishandling a crisis; policy drafting needs strict human review.
• Content moderation: AI understands context and subtle hate speech better than keyword filters.
• Enterprise RAG: solves keyword search’s failure to grasp meaning and synthesize across documents.
• Manufacturing: generative design (optimize weight/strength/cost), predictive maintenance, quality control generating rare defect samples.
• Automated journalism fits structured earnings/sports briefs, not investigative pieces or editorials.
• Term to know: Procedural Content = AI generates game art/levels to cut cost.

1.3 Technical terminology

• Parameters (learned by the model) vs hyperparameters (set by humans, e.g. learning rate, batch size).
• Backpropagation (compute gradients and update weights during training) vs Inference (predict on new data after deployment).
• Overfitting (great on training, poor on test — no generalization) vs Underfitting (too simple); Bias-Variance Tradeoff (high bias = underfitting, high variance = overfitting).
• Precision vs Recall trade off; F1 is their harmonic mean (trap: not the arithmetic mean).
• Decoding: Greedy (repetitive), Beam Search (width=1 equals Greedy), Top-K, Temperature=0 most deterministic.
• Architecture: Encoder-Decoder → Seq2Seq (translation); Decoder-only (GPT continuation); Self-Attention (words within a sentence); Cross-Attention (across two sequences).
• RNN hard to parallelize + long-range forgetting; LSTM gating solves vanishing gradients; GRU is a simplified LSTM.
• Word2Vec (static, same word same vector) vs BERT (contextual, same word changes with context, like “bank”).
• Pretraining tasks: GPT = CLM (causal/unidirectional); BERT = MLM (masked/bidirectional).
• Terms: Ground Truth, Dropout (randomly disable neurons to fight overfitting), Data Augmentation (rotate/crop/add noise), Quantization (32→8/4-bit, smaller and faster), LLM-as-a-Judge, Ablation Study (remove a component to see its effect).

1.4 Technical principles

• SD ecosystem: Automatic1111 (most complete WebUI), ComfyUI (node-based), Fooocus (minimal, high quality), Civitai (model/LoRA community).
• Parameters: CFG Scale (how closely the image follows the prompt), Sampler (detail/convergence), Checkpoint (full model, sets style), LoRA (small file, learns a specific character/style).
• ControlNet preprocessors: Canny (edges), OpenPose (pose).
• RAG tools: LangChain (glue framework), LlamaIndex (connect private data to an LLM).
• Local LLM: Ollama, LM Studio, GGUF (a “format,” not software), GPT4All.
• Search-type: Perplexity.ai (cites sources, fights hallucination), Poe (aggregates many models).
• Inference optimization: TensorRT (NVIDIA acceleration), ONNX (cross-framework format), QAT (quantize during training) vs PTQ (quantize after training).
• Edge: Edge AI is not “unlimited compute”; NPU for low-power inference; TOPS measures AI chip performance.
• Terms: MoE (Mixture of Experts, only part activated each time), IP-Adapter (use an image as a prompt), Delimiter (separates instructions from data, blocks injection), Pipeline (Hugging Face wraps pre/inference/post into a one-line API).

Area 2 — Prompting

2.1 Prompt optimization

• A good prompt states: goal / audience / tone / length and other concrete constraints.
• Reasoning guidance: CoT (think step by step), Least-to-Most (break into sub-questions), Step-Back (abstract first), Tree of Thoughts (multi-path + backtrack), Self-Consistency (majority vote).
• Long-conversation failures: Prompt Decay (forgets the original instruction), Instruction Creep (instructions pile up messily), Lost in the Middle (ignores the middle).
• Security: Prompt Injection (broad) ⊃ Jailbreak (a subclass); Prompt Leaking (trick out the System Prompt); defenses = Sandwich Defense, Instruction Defense.
• Delimiters (```, ###, XML) separate instructions from data to block injection; Claude is good at XML.
• Temperature high = creative; Top-P low = conservative/deterministic.
• Best way to reduce hallucination: require “answer only from the provided data; say you don’t know if you don’t.”
• Image parameters: —stylize (artistic), —chaos (variety), :: (weighting), Seed (lock composition), Negative Prompt (exclude elements).
• Bias: English prompts tend to be higher quality; without a specified culture, output skews Western (Wedding defaults to a white dress).
• Terms: Meta-Prompting (ask AI to write your prompt), Grandma Exploit, Flipped Interaction (the model asks you questions instead).

2.2 Data governance

• Data Drift (input distribution changes) vs Concept Drift (input-output relationship changes) [classic mix-up].
• Access control: RBAC (role-based) vs ABAC (attribute-based, finer-grained).
• Data architecture: Data Lake (unstructured) / Warehouse (ACID) / Lakehouse (combined) / Fabric (unified access) / Mesh (decentralized, data as a product).
• GDPR right-to-be-forgotten technical challenge = Machine Unlearning (very hard to precisely delete once trained into weights).
• Terms: Golden Record (the single cleaned, correct version / SSOT), Dark Data (collected but never analyzed), Data Poisoning, Shadow AI (employees using unapproved AI tools), HNSW (approximate nearest-neighbor vector algorithm).

2.3 Recent developments

• Linear-complexity architectures (solving Transformer’s quadratic cost): Mamba (SSM), RWKV, RetNet, Griffin, Jamba (hybrid).
• Inference acceleration: Flash Attention (IO-aware), PagedAttention (KV cache paging, vLLM), Speculative Decoding (small model drafts, large model verifies).
• Quantization extreme: BitNet b1.58 (weights of -1/0/1, ≈1.58 bit).
• RAG variants: Graph RAG (knowledge graph), Self-RAG (self-reflective), CRAG (assesses relevance, triggers web search), Agentic RAG.
• MoE: Grok-1 (314B but ~25% used per pass), DBRX, Mixtral; large total params, controllable inference cost.
• Post-training (RLHF/RLAIF) decides safety/alignment, key to reasoning models like o1.
• AI Agent vs Chatbot: Agent has planning/tools/loops; e.g. Devin, SWE-agent.
• Science breakthroughs: AlphaFold 3 (protein/DNA/RNA), GraphCast (weather), GNoME (new crystals), AlphaGeometry (geometry proofs).
• Natively multimodal (Gemini 1.5, end-to-end) beats stitching models together.
• Terms: Compound AI Systems (multiple models/tools combined), Model Merging (merge fine-tuned weights), Sovereign AI (nations build their own AI infrastructure), SynthID (invisible watermark), Embodied AI (AI in physical robots).

2.4 Self-directed learning

• Methods: Feynman technique (explain it to a novice to test real understanding), Project-Based, Building in Public, Peer Learning.
• Dev tools: Cursor (AI editor that understands the whole project), VS Code+Copilot, Replit (in-browser), Colab (free GPU).
• Cloud certs: Google PMLE / Azure AI Engineer / AWS ML-Specialty.
• Data preprocessing often takes 80%+ of an AI project’s time.
• Math mapping: linear algebra → matrices/tensors; calculus → gradient descent/backprop; probability/statistics → predicting the next token.
• First step of self-learning = get hands-on (sign up for ChatGPT/Midjourney and try it yourself).
• Terms: roadmap.sh, Pomodoro (25+5), Growth Mindset vs Abundance Mindset, AutoGPT / BabyAGI.

Area 3 — Applied skills (eight generation tools)

3.1 Text generation

• Core positioning = “intelligent copilot”: boost efficiency, replace repetitive low-value writing, not human creation.
• Task names (loves testing “pick the right function”): Summarization / Paraphrasing / Expansion / Translation / Sentiment Analysis.
• Customer-service rule: AI tries first, then hands off seamlessly to a human (with a summary) for professional/medical/legal/sensitive matters.
• Professional docs (legal/medical/financial): AI is only a draft/aid, human review required.
• Consistency control: Glossary/Style Guide, brand tone-of-voice guide, persona prompt.
• Fairness: job descriptions avoid gender/age/race bias; news uses factual statements.
• Multimodal bridging: text AI often acts as the “input layer” (video scripts, image prompts, narration) but doesn’t make images/audio itself.
• Personalized recommendations: not just right/wrong, but explanations + avoiding filter bubbles.
• Tools: Grammarly (grammar + tone), Persona-based Content, Style Guide / Glossary.

3.2 Image generation

• Three editing techniques (loves the distinction): Inpainting (remove/repair local), Outpainting (extend the canvas), Upscaling (enlarge and fill detail).
• ControlNet/OpenPose for precise composition/pose control.
• Career impact = replaces some repetitive work, still needs human creative direction/taste/finishing.
• E-commerce: composite product photos into scene backgrounds / virtual models.
• Copyright & ethics: complex ownership, bias from training data, must label AI-generated and be traceable.
• Ethics line: no porn/violence/hate, no non-consensual Deepfake.
• Prompt keywords “photorealistic”/“8k” = more realistic detail (not size change / speed-up).
• Tools: Stable Diffusion, Virtual Staging, virtual try-on/makeup, Remini (photo restoration).

3.3 Presentation generation

• Core principle “Less is More”: minimal text, more visuals.
• Script/outline needs: topic, target audience, talk length, core arguments.
• Speaker Notes = each slide’s expanded explanation / data source / speaking prompts.
• Three stages: before (content review) / during (delivery analysis) / after (summary report).
• Storytelling: Problem-Solution, STAR frameworks.
• Visual consistency, distinguish: visual coherence / language style / brand norms.
• Fact-checking: needs reliable data + human review, not just AI auto-correction.
• Alt Text = auto-generated image description for screen readers = accessibility.
• Data functions in three layers: live updates / data sync (Excel/Sheets) / data governance.
• Interactivity: Q&A hosting, polls, quizzes, gamification (points/leaderboard/badges).
• Virtual Presenter, version control (revert to any version), chart selection (recommends by data type + goal).

3.4 Audio generation

• The three voice conversions (loves testing): TTS (text→speech), STT (speech→text), S2S (speech-to-speech translation, converts directly to speech).
• TTS naturalness = simulating pitch/intonation/stress/rhythm.
• Vocaloid-type = turning text into a “singing voice” of a specific timbre.
• Music generation: rhythm / harmony / arrangement / composition.
• Dubbing / ADR (automated dialogue replacement) = synthesized speech matching lip-sync/emotion.
• Audio editing: Adobe Podcast (Enhance Speech) removes noise/echo, lifts to studio quality.
• Three security items (easily confused): voice biometrics (identity verification) / AI voice forensics (detect if AI-generated/tampered) / audio watermark (provenance).
• AI voice fraud = cloning a relative’s voice to scam (the most concrete ethical risk).
• Two-way emotion: voice emotion recognition vs emotion expression generation — don’t mix.
• Foley Sound / 3D spatial audio.

3.5 Video generation

• Covers the full pre- to post-production pipeline.
• Text-to-Video vs Image-to-Video (the latter needs one still image + an action prompt).
• Facial animation (from voice/emotion/mocap), voice cloning, De-aging, visual style transfer.
• Generative Fill (remove objects/extend background), video background removal, Upscaling.
• Deepfake = the biggest ethical controversy → needs detection + transparency labeling.
• Ethics: no porn/violence/hate/non-consensual Deepfake.
• For professionals = faster and cheaper, still needs human creative direction; promotes the “democratization of filmmaking.”
• Tools: Pika Labs (Motion Brush), HeyGen (Lip-syncing).

3.6 Code generation & websites

• AI Pair Programming (a virtual programmer collaborating in real time).
• Code Refactoring (optimize structure/readability without changing external behavior).
• Auto-generate unit tests; natural language to SQL (lets non-technical users query databases).
• IaC (Terraform/CloudFormation, manage infrastructure as code), containerization (Dockerfile/K8s), CI/CD (GitHub Actions).
• Web integration: payments (Stripe/PayPal), CRM (Salesforce), ERP (SAP).

3.7 Data analysis

• Core = auto-detect patterns/anomalies/trends.
• AutoML (auto modeling, lower barrier), BI (natural-language charts/reports, NL-to-SQL).
• Feature engineering, data cleaning, anomaly detection (fraud/failure).
• Customer churn prediction vs customer lifetime value (LTV) prediction.
• Data visualization (auto chart choice) / data storytelling, predictive maintenance, supply-chain optimization.

3.8 Chatbots

• Intent Recognition (“book a flight”) / Entity Extraction (“tomorrow,” “Taipei”).
• Rule-based (needs all paths defined manually, inflexible) vs AI-powered.
• Proactive Assistance, Emotional Intelligence (empathetic responses).
• Customer service: multi-channel, self-service, seamless handoff with summary, multilingual.
• Crisis handling: detect suicidal ideation → immediately provide professional resources and referral.

Area 4 — Ethics & law

4.1 Social and individual impact

• Whole-chapter theme “double-edged sword”: every domain brings efficiency and new risk at once.
• Augmentation vs substitution.
• Automation Bias: over-trusting AI, blindly following even when it’s wrong.
• Deepfakes / disinformation threaten elections, news credibility, social trust.
• Echo Chamber / Filter Bubble → social polarization.
• Human-AI emotional attachment, cognitive decline.
• Digital divide, three variants: digital divide (lack of devices/literacy) / digital colonialism (loss of data sovereignty) / cultural homogenization (marginalizing minorities).
• Labor: Upskilling/Reskilling, UBI (universal basic income), Ghost Work (the low-paid labeling/moderation labor behind AI).
• Environmental cost: training/inference energy use → higher carbon emissions (a problem, not a green benefit).

4.2 Ethical principles

• Non-maleficence (do no harm, passive) vs Beneficence (do good, active).
• Human control gradient: Human-in-the-loop → Meaningful Human Control (MHC) (retain understanding/judgment/intervention).
• XAI two layers: global explainability (whole model) vs local explainability (single prediction).
• EU AI Act: the world’s first comprehensive AI law, risk-tiered.
• NIST AI RMF: voluntary, four stages — govern/map/measure/manage.
• Bias: detection (diversity analysis + fairness testing + counterfactual) vs mitigation (pre-processing + training constraints + post-processing).
• Alignment, Double-use Problem, Red Teaming.
• Terms: MHC, global vs local explainability, regulatory sandbox (evaluate ethical impact in a controlled environment).

4.3 Laws and regulations

• US AI Bill of Rights: non-binding guiding principles (not law).
• Liability: when AI errs, it’s not “the AI is responsible” — it falls on designer/manufacturer/deployer/user.
• Copyright: disputes over ownership of AI-generated content; whether training on copyrighted works infringes; whether AI can be an inventor.
• Data sovereignty / localization: data must be stored in a specific geography.
• Military: LAWS international bans, humanitarian law, chain-of-command responsibility.
• Digital ethics codes have no legal force (a code ≠ a law); AI legal personhood is still debated, unresolved.
• Terms: AI Bill of Rights, PIPL (China’s data law) / Taiwan’s data law, Data Localization, LAWS.

4.4 Privacy and data security

• Privacy-Enhancing Technologies (PETs) family [the most-tested distinction of the chapter]:
  • Differential privacy (add mathematical noise, can’t identify an individual)
  • Homomorphic encryption (compute directly on encrypted data)
  • Federated learning (train locally, only send parameters — data stays put, the model moves)
  • Secure Multi-Party Computation (SMPC) (parties compute jointly without leaking inputs)
  • Data masking (replace with correctly-formatted fake data for test environments)
• Access control: RBAC (role) / ABAC (attribute, finer) / ACL (list) / least privilege.
• Zero Trust architecture: “never trust, always verify.”
• Encryption’s three states: at Rest / in Transit / KMS key management.
• Threats: Data Poisoning, Model Inversion Attack (reconstruct training data such as faces), Adversarial Robustness (resisting adversarial samples, a defensive goal).
• AI security applications: SIEM (unified monitoring), IDS (intrusion detection), SOC (security operations center).
• Isolation: TEE (Trusted Execution Environment, hardware isolation), Secure Data Enclaves; MFA, separation of duties.

---

How to use it: the day before, run through “Part 1 — terms” and the distinction pairs (Data Drift vs Concept Drift, TTS/STT/S2S, Precision vs Recall, Non-maleficence vs Beneficence, the PETs family) twice; skim the rest for recognition. Good luck out there — for the method, see the companion post: Two AI Certs in Two Months.

Raw Markdown