生成AIパスポート StudyGenerative AI Passport Exam

Learning goals

• Distinguish rule-based systems from machine learning, and the types of learning (supervised, unsupervised, reinforcement) with examples
• Explain how neural networks learn by adjusting weights, and overfitting and its countermeasures
• Explain CNN, RNN, LSTM, VAE and GAN, and the difference between discriminative and generative AI
• Explain the features of the three AI booms and the path that led to generative AI

Key points

• Rule-based systems run on rules written by people and are weak against unexpected input. Machine learning learns rules (a model) from data.
• Supervised learning does classification and regression on labeled data; unsupervised learning does clustering and dimensionality reduction without labels; reinforcement learning learns a policy from actions and rewards. Semi-supervised learning combines a little labeled data with a lot of unlabeled data.
• No free lunch theorem: no single algorithm is always best for every problem.
• A neural network consists of input, hidden and output layers, and adjusts weights with backpropagation and gradient descent so that the loss function value becomes smaller.
• Overfitting means fitting the training data too closely so performance drops on unseen data. Countermeasures include regularization, dropout, early stopping and more data.
• CNNs capture image features with convolution; RNNs handle sequence data but suffer from vanishing gradients; LSTMs keep long dependencies with gates. Attention and Transformers later became mainstream.
• A VAE compresses input into a latent vector with an encoder and generates with a decoder; a GAN generates by pitting a generator against a discriminator.
• The first boom was search and inference, the second knowledge representation and expert systems, the third machine learning, deep learning and big data. Generative AI based on Transformers, LLMs and diffusion models extends from it.

Workplace uses

• General AffairsSwitch inquiry routing from keyword rules (rule-based) to a model trained on past routing results (supervised learning), making it easier to handle new phrasings.
• SalesCluster customer purchase data without labels and change proposals for each group of similar customers.

Exam objectives (from the exam guide)

• AIの基本的な仕組み（ルールベースと機械学習）
• 機械学習（教師あり・教師なし・強化・半教師あり学習、ノーフリーランチ定理）
• ニューラルネットワーク（学習の仕組み、過学習と対策、転移学習、CNN・RNN・LSTM・VAE・GAN、識別系AIと生成系AI）
• AIの歴史（ダートマス会議から第三次AIブーム、生成AI、シンギュラリティ）

Learning resources

• Generative AI Passport exam syllabus (applies from the February 2027 exam, PDF) ↗
• 生成AIパスポート exam guide / study guide ↗

Learning goals

• Explain the Transformer's self-attention and where diffusion models, foundation models and multimodal models fit
• Explain the three stages of pre-training, fine-tuning and RLHF, plus scaling laws and emergent abilities
• Explain the link between next-token prediction and hallucination, and what Temperature, Top-p and the context window mean
• Explain the differences between cloud, on-premises and open-source (open-weight) LLMs and how to choose, plus cautions for generative AI in each modality

Key points

• The Transformer computes relationships between all words in a sentence at once with self-attention (Query, Key, Value), and is strong at parallel processing and long-range dependencies.
• Diffusion models learn the process of adding noise and generate images from noise by reverse diffusion. Latent diffusion models work in a compressed latent space to reduce computation.
• A foundation model is pre-trained on large amounts of data and can adapt to diverse tasks. Multimodal models handle multiple modalities such as text, images and audio.
• LLM training proceeds as pre-training (next-token prediction) → fine-tuning (instruction tuning) → RLHF (building a reward model from human ratings and applying reinforcement learning).
• Scaling laws: increasing parameters, data and compute improves performance in a predictable way. Abilities that suddenly appear as scale grows are called emergent abilities.
• Generation is repeated probabilistic next-token prediction. Because it is not looking up facts, plausible errors (hallucinations) occur. Giving external information as grounding with RAG reduces them.
• Raising Temperature makes output more diverse; lowering it makes output more stable. Top-p limits candidates to the most probable ones up to cumulative probability p. The context window is the maximum number of tokens handled at once.
• On-premises and local LLMs avoid sending data outside the company (data sovereignty) but require operations and securing performance. Check license terms even for open-weight models.
• Reasoning models spend more compute at inference time before answering and think step by step with chain of thought, improving accuracy in math, coding and similar tasks.
• With image, audio and video generative AI, be careful about portrait rights, consent and deepfakes. Provenance information such as C2PA and digital watermarks are used to show that content is generated.

Workplace uses

• PRWhen brainstorming taglines, set Temperature higher for varied ideas; when summarizing terms, set it lower for stable output.
• ITFor departments handling confidential data, consider an on-premises LLM run in-house instead of a cloud LLM, and check the license and operational burden.

Exam objectives (from the exam guide)

• 生成AIを支える技術革新（Transformer以降、拡散モデル、基盤モデル・マルチモーダル）
• 生成AIの学習プロセス（事前学習・ファインチューニング・RLHF、スケーリング則）
• 生成AIの生成（推論）プロセス（トークン化、埋め込み、次トークン予測、ハルシネーション、Temperature・Top-p、コンテキストウィンドウ）
• LLMの基礎（LMとLLM、オンプレミス／オープンソースLLM）
• ChatGPTと主要な生成AIモデルの動向（推論モデル、Gemini・Claude・Copilot）
• 各モダリティの生成AI（画像・音楽・音声・動画、C2PA・電子透かし）

Learning resources

• Generative AI Passport exam syllabus (applies from the February 2027 exam, PDF) ↗
• 生成AIパスポート exam guide / study guide ↗

Learning goals

• Explain how AI agents differ from conventional conversational generative AI, and how workflow and autonomous agents differ
• Explain how RAG works (embeddings, vector search, chunking, reranking) and its advanced forms
• Explain the loop of task decomposition, tool execution, memory, observation and replanning, and the roles of MCP, A2A and Agent Skills
• Explain agent-specific risks (excessive permissions, cascading errors, prompt injection) and the ideas of least privilege, sandboxes, human approval and evaluation

Key points

• An AI agent plans toward a goal, acts using tools, observes results and decides its next action. Conversational generative AI basically responds once per instruction.
• Workflow agents follow steps defined by people and are predictable. Autonomous agents let the LLM decide the steps themselves; they are flexible but harder to control.
• RAG searches for documents related to the question and adds their content to the prompt for answering. Documents are split into chunks, turned into embeddings (vectors) and searched semantically in a vector database.
• Advanced RAG: hybrid search combined with keyword search, reranking, and query rewriting (Advanced RAG); Agentic RAG where an agent repeats searches; Graph RAG using knowledge graphs.
• With Function Calling and Tool Use, an LLM calls external functions and APIs. Short-term memory (conversation context) and long-term memory (stored past information) are used as appropriate, and the craft of putting the needed information into context is called context engineering.
• MCP (Model Context Protocol) is a common standard connecting AI apps with external tools and data sources. A2A is a protocol for agents to work with each other. Agent Skills package procedures and materials to give to agents.
• Coding agents investigate, modify and test code and create pull requests. Harness engineering means preparing the environment agents run in (tools, permissions, review mechanisms).
• Risks: excessive permissions, errors cascading and growing, and following instructions embedded in external documents or the web (prompt injection). Countermeasures are the principle of least privilege, sandboxes, allowlists, human approval before important actions (Human-in-the-Loop), defense in depth, evaluation (Eval) and continuous monitoring.

Workplace uses

• General AffairsUse RAG for a chat that answers internal policy questions, showing the relevant policy sections as the basis. Filter by metadata such as effective date so old policies are not mixed in.
• AccountingEven when invoice processing is delegated to an agent, payment is executed only after staff approval, and the agent is given only view and draft permissions.

Exam objectives (from the exam guide)

• AIエージェントとは（定義、ワークフロー型と自律型、自律性のレベル）
• RAG（仕組み、発展形、ユースケース）
• AIエージェントの仕組み（タスク分解と計画、ツール実行、メモリ、コンテキストエンジニアリング、フィードバックループ）
• MCP・Agent Skills と外部連携（MCP、A2A、マルチエージェント）
• AIエージェントのツールと活用事例（コーディングエージェント、ブラウザ操作、ディープリサーチ）
• AIエージェントのリスク・安全性・評価（最小権限、プロンプトインジェクション、Human-in-the-Loop、Eval）

Learning resources

• Generative AI Passport exam syllabus (applies from the February 2027 exam, PDF) ↗
• 生成AIパスポート exam guide / study guide ↗

Learning goals

• Recognizing that AI is an assistant and final responsibility lies with people, decide whether to use, use carefully or not use it according to purpose and impact
• Explain risks such as hallucination, information leaks, rights infringement, deepfakes and shadow AI, and how to prevent them
• Decide what information may be entered, based on Japan's Act on the Protection of Personal Information (personal information, individual identification codes, special care-required personal information, provision to third parties) and the concept of trade secrets
• Explain the key points of copyright (reliance and similarity, AI training and copyright, copyrightability of AI output), the AI Guidelines for Business, internal guidelines, and the AI Act and international rules

Key points

• Do not take generative AI output at face value; check facts, figures and sources against primary information. The greater the purpose and impact, the higher the level of verification; consult experts when needed and decide whether to adopt, revise or discard.
• Shadow AI means using AI tools not approved by the company for work. It leads to information leaks and lack of traceability. Provide approved tools and set rules for what information may be entered.
• Personal information is information that can identify a living individual. Individual identification codes (My Number, passport numbers, facial recognition data, etc.) are personal information on their own. Obtaining special care-required personal information (medical history, criminal records, beliefs, etc.) in principle requires the person's consent.
• To be protected as a trade secret (Unfair Competition Prevention Act), three requirements are needed: secrecy management, usefulness and non-public status. Entering confidential information into external AI may undermine secrecy management.
• Copyright infringement is judged by whether there is similarity to and reliance on an existing work. The same applies to AI output. Whether AI output is a copyrighted work is judged by whether there is human creative intent and creative contribution.
• The AI Guidelines for Business set out common guiding principles such as human-centricity, safety, fairness, privacy protection, ensuring security, transparency, accountability, education and literacy, ensuring fair competition and innovation.
• Internal guidelines include purpose of use, approved tools, information that may be entered, output review, usage logs and incident response, and are reviewed continuously as agile governance.
• Japan's AI Act (Act on the Promotion of Research, Development and Utilization of AI-Related Technologies) is a promotion law balancing innovation and risk response, establishing the AI Basic Plan and the AI Strategy Headquarters. The EU AI Act regulates with a risk-based approach. The G7 Hiroshima AI Process set out international guiding principles and an international code of conduct for advanced AI systems.

Workplace uses

• HRFor a proposal to use generative AI in document screening for hiring, check how applicants' personal information is handled and the risk of evaluation bias, and have a person make the final decision.
• SalesDo not enter customer names or quote amounts into unapproved tools on personal accounts. Use company-approved tools configured not to use input data for training.

Exam objectives (from the exam guide)

• 生成AI利用者の基本マインドセット
• 生成AIの代表的リスク
• シャドーAIと企業コンプライアンス
• AI出力の評価と検証
• 個人情報・プライバシー・秘密保持
• 著作権・知的財産・生成物の権利
• AI倫理とAI事業者ガイドライン
• AIガバナンスと社内ガイドライン
• AI法・国際ルール・AI安全保障

Learning resources

• Generative AI Passport exam syllabus (applies from the February 2027 exam, PDF) ↗
• 生成AIパスポート exam guide / study guide ↗

Learning goals

• Use the basic prompt structure (instruction, context, input, output format) and choose among Zero-Shot, Few-Shot and Chain-of-Thought
• Use generative AI for emails, meeting minutes, summaries, translation and slides, and check output for fact accuracy and proofreading
• Explain how to use generative AI in planning, marketing, creative work and data analysis, and cautions about rights, personal information and checking primary sources
• Explain how to automate work with coding agents and browser automation, and the keys to success in each industry

Key points

• Basic prompt elements: instruction (what to do), context (background, purpose, target reader), input data and output format (table, bullet points, length). Setting a role (e.g., as an experienced editor) helps align the perspective of the output.
• Zero-Shot gives instructions without examples. Few-Shot shows a few examples to align format and criteria (In-Context Learning). Chain-of-Thought has the model think step by step to improve reasoning accuracy.
• Controlling specificity: vague instructions tend to produce average output. Convey purpose, constraints and evaluation criteria concretely, and improve interactively.
• Document creation: strong at summarizing, proofreading, changing tone, translating and extracting key points from minutes. Always check proper nouns, numbers and dates against the original.
• Planning and marketing: use for brainstorming, sounding out ideas, taglines and survey design. Back up market and competitor research with primary information.
• Creative work: when generating images, video and audio, check license terms, whether commercial use is allowed, similarity to existing works or people, and portrait rights.
• Data analysis: helps with formatting, classifying and aggregating tables, and extracting information from large numbers of documents. Double-check calculations and be careful about entering data containing personal information. Deep research investigates across multiple sources.
• Development and automation: use coding agents to investigate, generate and review code, and browser agents to automate web procedures. Narrow permissions and have people check important actions.

Workplace uses

• SalesFrom meeting notes, show past good minutes as Few-Shot examples to create minutes in the same format and a list of next actions.
• MarketingGenerate 20 landing page tagline ideas with generative AI, check for trademarks and similarity to competitors' copy, and have a person choose and edit.

Exam objectives (from the exam guide)

• プロンプトエンジニアリング（基本構成要素、Zero-Shot・Few-Shot・Chain-of-Thought、役割設定・出力形式の指定）
• 文書作成・コミュニケーション業務での活用
• 企画・マーケティング業務での活用
• クリエイティブ制作での活用
• データ分析・リサーチ業務での活用
• 開発・業務自動化での活用、業種別の活用

Learning resources

• Generative AI Passport exam syllabus (applies from the February 2027 exam, PDF) ↗
• 生成AIパスポート exam guide / study guide ↗