Appendix B. Self-check answers and explanations

Chapter 1

1-1. Training use, retention, human review, location (storage and processing), and contractual status (DPA and processor status). Retention and human review are most often confused with training use. A no-training promise says nothing about retention periods or human review for abuse monitoring.

1-2. Terms of service define contracting parties, service scope and responsibilities. Usage policies define permitted uses, prohibitions and additional requirements. A DPA defines how a vendor processes customer data on customer instructions (purpose limitation, security, subprocessors, return and deletion).

1-3. Zero retention does not extend to every feature. OpenAI documentation excludes features requiring server storage, such as Conversations, files and vector stores, and explains that these objects remain until deleted. Even zero retention organizations must manage separate deletion procedures for material uploaded to them.

1-4. The OpenAI API region table supports storage but not in-region processing in South Korea, so inference occurs overseas and temporary storage outside the region is possible. Personal information therefore involves overseas entrusted processing, requiring a transfer basis first. If entrustment is necessary for a contract with the data subject, disclose the statutory details in the privacy policy or notify individually; otherwise another basis, such as separate consent, is needed. Non-US regions also require zero retention or Modified Abuse Monitoring approval.

1-5. Microsoft Copilot prompts and responses fall under the DPA, with Microsoft as processor. Generated web search queries sent to Bing do not, with Microsoft as controller. Rules should give web search a separate row and require admin policies disabling it for users handling data whose contracts prohibit processing outside DPA coverage.

1-6. Record the effective dates of both versions, changed requirements (such as high-risk areas or disclosure methods), and check date. Before commencement, ① compare changed requirements with the approved use list for internal impact, and ② assign owners and deadlines to revise affected approval conditions and employee guidance. Separately decide whether to retain company rules even if vendor policy relaxes.

1-7. Example: training default, training exceptions, default retention, admin retention settings, zero retention, storage and processing locations, DPA and usage policy version. Record a link, supporting sentence and check date in each cell. Move unconfirmed cells to vendor queries with reply deadlines, and do not enter data relying on those cells until replies arrive.

Chapter 2

2-1. Effective 22 January 2026; enacted as Act No. 20676 (promulgated 21 January 2025). The 20 January 2026 amendment (Act No. 21311) took effect on 22 January except provisions on AI-vulnerable groups (Article 3(5), the latter part of Article 35(1), etc.), which took effect six months after promulgation. The current Act's effective date is therefore shown as 21 July 2026.

2-2. Acts and Enforcement Decrees are binding; ministerial notices bind within statutory delegation; guidelines are nonbinding guidance. Ministerial notices were the uncertain layer as of the check date. No notices linked to the delegating provisions were visible at the National Law Information Center, so finalization must be checked directly.

2-3. The company is a utilization business operator providing a service through an external model. Its generative AI service requires Article 31(1) prior notice (terms, screen, etc.) and paragraph (2) output labeling (human-perceptible or machine-readable; the latter also requires at least one notice). If outputs are text, paragraph (3), concerning sounds, images or video difficult to distinguish from reality, ordinarily does not apply. High-impact status requires separate prior review.

2-4. The system must potentially cause significant impact or risk to life, physical safety or fundamental rights, and be used in an Article 2, subparagraph 4 area. Even within such an area, an auxiliary use without potential significant impact on rights or safety may not be high-impact.

2-5. If dissatisfied with the initial reply (within 30 days of request, with one possible 30-day extension), submit a reconfirmation request within 10 days of receiving it, attaching the previous reply and related materials. The minister seeks expert committee advice and replies within 30 days of the reconfirmation request (Enforcement Decree Article 25(3)–(5)).

2-6. Publish key risk management details, key explanation-plan details, the user protection plan, and the supervisor's name and contact details. Retain documentary evidence of measures for five years. Trade-secret portions may be excluded from publication, but trade secrets cannot justify withholding everything or omitting document retention.

2-7. Discovery, report or complaint → demands for document submission and investigation under Article 40(1) and (2) → suspension or corrective order under Article 40(3) if a violation is found → fine under Article 43(1), subparagraph 3 for noncompliance (Schedule 2: KRW 10 million first, KRW 20 million second, KRW 30 million third or later). Labeling failure alone is not grounds for an administrative fine. Government policy states that investigations during the grace period are highly exceptional.

2-8. There is no single answer. A good answer records all nine steps and statutory grounds per function, prior review and the confirmation-request decision for potentially statutory-area functions, and separates unresolved items into legal review, confirmation requests and pending notices.

Chapter 3

3-1. Use anonymous information if it can achieve the purpose; otherwise use pseudonymized information if it can. To justify originals, first explain why anonymous or pseudonymized information cannot meet the training purpose, specifying which information is needed and why.

3-2. Relationship to the original collection purpose; predictability from collection circumstances and processing practices; unfair infringement of data subject interests; safeguards such as pseudonymization and encryption. For ongoing additional use, disclose assessment criteria in the privacy policy, and have the CPO check compliance with them.

3-3. Legitimacy of AI development purpose, necessity of processing public personal information, and specific balancing of interests. The guidance is PIPC interpretation and recommendation, not legislation or a binding standard. Identify the memo as an assessment with reference to the guidance.

3-4. Genetic information obtained through genetic testing and similar means; criminal records; race and ethnicity; and information concerning physical, physiological or behavioral characteristics generated through technical means to identify a specific individual. Characteristic information becomes sensitive when both identification purpose and technical generation are present.

3-5. Entrustment performs the entrusting party's work, and processors cannot use beyond its scope (Article 26(5)). Vendor use for its own model improvement exceeds that scope and is substantively recipient-purpose processing, similar to third-party provision, requiring consent or another basis. That is why no-training commitments and DPAs are checked.

3-6. A concise, meaningful explanation covering the outcome, major personal information types used, key criteria such as their influence, and decision procedures including processing of major information. Provide it in writing or similar form within 30 days of request. For justifiable reasons, notify the requester and extend up to twice by 30 days each (Enforcement Decree Article 44-3). For decisions without significant effects on rights or obligations, notice of the disclosed relationship between major information types and the decision, considerations and processing procedures may substitute.

3-7. Example: collection (contract performance, period of use); search index (same purpose, original retention period, permission filters); external model calls (entrustment; for overseas regions, confirm transfer grounds and disclose statutory details or obtain consent); conversation logs (quality management, fixed period); destruction (originals, indexes and logs together; retain records required by other laws separately). A good answer marks statutory grounds and points where sensitive information may enter.

Chapter 4

4-1. At least five of: work and owner; tool, plan, contracting party and route; input data level; connected systems and external actions; legal review results; status and review date. The most easily missed use is AI added through updates to existing SaaS.

4-2. Example: Policy: "Process customer personal information only in approved tools." Standard: "Approved tools must provide no training use, fixed retention and disabled feedback." Procedure: "Submit a new-tool application for security review, with an approval decision communicated within the set period." Guidance: "Remove customer names and contact details before pasting."

4-3. Outputs going directly to customers without human review; outputs informing decisions concerning individuals such as recruitment or assessment, or AI directly taking hard-to-reverse actions such as sending or payment. Even at the same data level, greater use impact or action permissions raise the category.

4-4. Requester, purpose, data level, period (expiry date), alternatives considered, compensating controls, approver and methods for deprovisioning accounts and reclaiming data at expiry. Repeated exceptions signal that rules or approved tools do not fit reality; consider revising rules or adding tools.

4-5. Settings saved as intended in an admin screen may apply differently in practice. Gemini app group settings override OU settings, for example, so an OU change may not affect group members, and propagation takes time. Only checking with an ordinary account demonstrates actual implementation.

4-6. Hard-to-reverse or externally consequential actions such as sending, payments and refunds, deletion or permission changes. Prompt injection or model errors can bypass instructions, so independent systems must enforce amount, volume and target limits for them to be real controls.

4-7. There is no single answer. A good table includes allowed, conditional and prohibited categories, with every conditional cell marked with an implementation method (retention settings, feedback disabling, OU restrictions, web search policy, etc.) or "Relies on human behavior".

Chapter 5

5-1. Example: "A policy-information chatbot confidently describes nonexistent coverage, causing a customer who relies on it to miss a claims opportunity, with only partial recovery possible." The affected person, condition, harm and recoverability must be apparent.

5-2. Valid and reliable; safe; secure and resilient; accountable and transparent; explainable and interpretable; privacy-enhanced; fair with harmful bias managed. "Valid and reliable" is the necessary condition; "accountable and transparent" relates to all others.

5-3. Examples: Confabulation, Data Privacy, Harmful Bias or Homogenization, Human-AI Configuration, Information Integrity (others: CBRN Information or Capabilities; Dangerous, Violent, or Hateful Content; Environmental Impacts; Information Security; Intellectual Property; Obscene, Degrading, and/or Abusive Content; Value Chain and Component Integration). Review all twelve as a checklist for each case and write concrete harm scenarios for applicable categories to avoid omissions.

5-4. Mitigate, transfer, avoid and accept. Transfer shifts some cost or liability to another party without reducing harm itself. Major human-impact risks therefore also require considering mitigation or avoidance.

5-5. Five-year retention of documentary evidence for high-impact responsibilities under AI Basic Act Enforcement Decree Article 27(2); preparation and retention of pseudonymized information processing records for at least three years from destruction under PIPA Article 28-4(3).

5-6. ISO/IEC 42001 certification demonstrates that an AI management system is established and operated, and that an independent certification body's assessment was passed (within certified scope). It does not demonstrate a particular product's accuracy or fairness, or the adopting organization's use-case risks or legal compliance.

5-7. There is no single answer. A good answer connects scenarios to controls row by row, distinguishes legal and contractual requirements from discretionary controls, raises acceptance authority with risk level, and lists unmeasured aspects as uncertainty.