Appendix B. Self-check answers and explanations

Because many Self-check questions ask for answers in your own words, there is no single correct answer. Below are sample model answers and the key thoughts that must be included in your answer.

Chapter 1

1-1. Example: "Just as a mobile phone keyboard suggests 'what' or 'together' after 'this evening', generative AI builds answers by piecing together, one fragment at a time, natural words to follow a question based on patterns learned from vast amounts of text. Therefore, while sentences sound natural, you must separately verify whether their contents are factual." The key points are that AI generates rather than retrieves, and that its standard is naturalness, not factuality.

1-2. Tasks suitable for search or official notices: (a) Library closed days, (c) Intercity bus fares. Both are current facts that are subject to change, so they must be verified through the organization's official notices. Tasks suitable for generative AI: (b) Greeting text message, (d) Notice summary, (e) Candidate names for a gathering. These are language and writing tasks where you can directly read and judge whether the result is good. However, dates and amounts included in the summary of (d) should be cross-checked against the original text.

1-3. Example: (1) Precise numbers such as times, amounts, and phone numbers (2) Names of people, stores, organizations, and systems (3) Grounds such as "According to the announcement by ○○" (4) Assertive words such as "definitely", "anyone", and "100%". Any three of these are sufficient. The more specific they are, the more useful they may appear, but these are areas where errors cause significant harm.

1-4. What can be known: That this AI consistently provides the same answer to this question—namely, that there are no warning signs of conflicting answers. What cannot be known: That the date is a fact. Because the same AI can repeat the same mistake, it must be verified through official notices.

1-5. Example (Kimchi-making): Delegate to AI — Announcement message for the family group chat on kimchi-making day, side dish menu ideas to accompany boiled pork. AI assists and I verify — Seasoning ingredient list tailored to the number of cabbages (verify quantities using the original recipe or past experience), kimchi-making schedule. Do directly myself — Ordering and paying for salted cabbages, dietary management for family members who must restrict salty food (consult a doctor if necessary). It is sufficient if tasks involving financial transactions or health are placed in the "Do myself" category.

1-6. Example: (1) Operating hours and closed days (2) Fees and prices (3) Eligibility and application periods for subsidies or government programs (4) Event dates (5) Responsible department and phone number. Any three of these are sufficient. All of these change frequently and may have altered since the time the AI was trained.

Chapter 2

2-1. What (desired deliverable, e.g., a thank-you text message), Why (situation and conditions, e.g., in return for moving-in rice cakes), To whom (target reader and tone, e.g., elderly neighbor next door, polite), How much (length, e.g., three lines), In what format (format, e.g., short lines for a text message, no emojis). It is sufficient if the names of the five fields are paired with examples from your everyday life.

2-2. Example: "I want to make kimchi stew for dinner for a two-person household (a couple in their 60s). At home, we have aged kimchi and pork front leg meat, and my husband needs to reduce salty food. Please provide the steps with numbers in seven steps or fewer, and give ingredient amounts for two servings." It is sufficient if the deliverable (cooking steps), situation (ingredients and conditions), audience (couple in their 60s), length/quantity (seven steps, two servings), and format (numbered steps) are included, and no names or addresses are present.

2-3. (1) Pinpoint specifically what is lacking ("It is too long and the second line is too stiff"). (2) Ask to keep the good parts ("Leave the first line as is"). (3) Do not change too many things at once; ask for one or two key items starting with the most important. Providing the reason or target audience ("Because an eight-year-old will read it directly") yields more accurate revisions.

2-4. Example: (a) "In short lines without tables or special symbols, so it can be pasted directly into a group chat." (b) "As a list grouped by supermarket aisle (produce, meat, dairy), one line per item." (c) "In short sentences suitable for reading aloud, without difficult Sino-Korean words." Match the format to where it will be used.

2-5. When starting a new conversation is better (two cases): Any two among: when the topic completely changes, when the conversation gets very long and earlier conditions are dropped, when the model repeats the same mistake despite being asked to correct it, when asking about someone else's business, or when you shared a lot of personal stories in the previous conversation. What must be done: Because how much the new conversation knows about previous conversations depends on the app and settings (such as memory features), re-state the necessary conditions.

2-6. As seen in Chapter 1, AI fills in unknown blanks with plausible words, and it is particularly prone to fabricating dates or store names. Asking it from the beginning to leave them blank gives the AI no opportunity to invent details, and since you fill in the verified facts yourself, the risk of incorrect dates slipping in is reduced. Another advantage is that areas requiring verification are visible at a glance.

2-7. Example: "I am planning to travel abroad for the first time this year. I would like your help preparing. Please ask me five questions first to find out what you need to know to create a preparation checklist tailored to me. Once I answer, make the checklist after that." It is sufficient if the request to ask questions first and the sequence of providing the deliverable after you answer are included.

Chapter 3

3-1. At least five among: numbers, dates and times, people's names, names of stores/organizations/programs, locations and addresses, phone numbers and website addresses, eligibility and rules, and cited sources. Also beware of broadly sweeping words like "anyone" or "definitely".

3-2. How to locate: Do not click the link directly; instead, navigate directly to the official website of the organization cited in the source via a search or bookmarks, and find the same document. What to check: (1) Can the document be found on the official website? (2) Who authored it (is it an official institution, or a personal blog or advertisement)? (3) Does the content of the document match the numbers, dates, and conditions in the AI's answer? (4) When was it written (posting date, revision date, project year)? All four criteria must be satisfied before it can be considered verified.

3-3. (a) The official website or primary phone number of the public health center (or the municipal/county/district office administering that public health center). (b) The effective date indication on the Ministry of Government Legislation's National Law Information Center. (c) A doctor or pharmacist. General health information can be viewed on the National Health Information Portal, but interactions between medications you take must be judged by a professional.

3-4. (1) Links received via text message may be fake websites (phishing sites), and the Financial Supervisory Service also advises verifying received addresses through web search. (2) AI can state numbers and addresses incorrectly or fabricate them. Navigating directly to the official website and calling the main phone number listed there avoids both risks.

3-5. Do not trust it merely based on the numbers, as five different blogs may have copied the same erroneous post. Treat them only as clues, and verify by searching the official website of the responsible organization or calling its main telephone number. If verification is not possible, do not use the information, pass it on marked as "unverified", or wait until it is verified.

3-6. Tasks suitable to delegate to AI (two items): Any two among creating a list of questions to ask during a doctor's appointment, looking up the general definitions of words on a prescription, creating a symptom tracking log, or rewriting hospital instructions in simpler language. Tasks to ask a doctor or pharmacist (two items): Any two among the cause and treatment of your symptoms, whether medications can be taken together, whether to reduce or discontinue medication, or whether exercise and dietary restrictions are appropriate for you.

3-7. What to do before sharing: Verify numbers and conditions such as amounts, eligibility, and application periods against the official announcement from the responsible agency. A line to include: Note where and when you verified it, for example, "Based on the announcement on the Garam City Hall website (posted on [Month] [Day])", so the recipient can also verify it.

Chapter 4

4-1. Example: "A retired principal in his 70s living in Building 102 of Garam Apartments had knee surgery last month; write a get-well text message." There is no name, but a neighbor would know who it is. Omit the apartment name and building number (omission), broaden "retired principal" to "longtime elderly neighbor" (broadening), and broaden "knee surgery last month" to "recent surgery". The core point is not the name itself, but reducing the pieces of information that, when combined, reveal identity.

4-2. Red: (a) Verification code, (e) Photo of the back of a credit card. Yellow: (c) Photo of granddaughter's face (uploading requires parental consent, and avoid uploading if possible), (f) Friend's name and phone number (omit or replace). Green: (b) "In their 60s, poor back", (d) Content of a general notice. However, even green items can become yellow when combined with other information.

4-3. Example: "My son is moving to a new home next week. Write a warm, three-line housewarming congratulatory message from a mother." I omitted my name and phone number (omission), replaced the son's name with "son" (replacement), and omitted the address. Names, numbers, and addresses are not needed to write a congratulatory text message.

4-4. What changes: Your conversations stop being used for AI Model Improvement (Training) (based on company guidance; exceptions may apply). What does not change (two items): Any two among conversations going to company servers, remaining in conversation history, being subject to safety reviews, being viewable by others on shared devices, or conversations submitted via feedback buttons being handled separately.

4-5. Good situations to use: Brief one-off questions, things you do not want saved in history (e.g., brief use on a shared device, briefly sharing personal worries). Things you must not do: Entering Red information such as resident registration numbers, passwords, verification codes, and account or card numbers just because it is a temporary chat. Even temporary chats pass through servers and may be retained for a certain period.

4-6. (1) Click Log out in the account menu. (2) Reopen the AI website to check whether the login screen appears. (3) Check that your password or automatic login has not been saved in the browser, and delete it if saved (it is even better if you declined "Stay signed in" and password saving from the start). The key point is that simply closing the window might not log you out.

4-7. Example: "Look for items containing words such as 'Data', 'Privacy', 'Training', or 'Model Improvement' in the app's settings menu. If you cannot find them, search for those same words in the service's official help center. If the setting exists, note its name and current status (on/off)." The table in Chapter 4 of this book summarizes the official guidance of three services as of the verification date, so the key is to verify directly yourself, keeping in mind that the apps may have changed.

Chapter 5

5-1. (1) Creates a sense of urgency. (2) Demands secrecy. (3) Demands money or financial information (account number, card number, verification code). (4) Prompts clicking a link or installing an app. (5) Arrives from an unfamiliar number or messaging app.

5-2. (1) Stop — stop everything: money, numbers, links, and apps. (2) Hang up. (3) Call back directly using a saved number (family) or an official representative number (institution). (4) Ask the verification question agreed upon among family members. (5) Notify other family members as well. (6) If it appears to be fraud, report it; if money has been sent, immediately contact the financial institution and the police.

5-3. As the Financial Supervisory Service advises, caller ID numbers can be manipulated, so it is impossible to know whether a call is genuinely from a bank based on the number displayed on the screen alone. Verifying on that call amounts to verifying with the fraudster, so you must hang up and call back directly using the representative number on the back of your card or on the official website.

5-4. Examples: (1) "When is Mom's birthday?" — This can be found through documents or social media. (2) "What is our home address?" — The scammer may already know personal information. (3) "What is the name of the grandchild's school?" — This can be revealed through social media photos or posts. A good verification question cannot be found on the internet, is easy for family members to remember, can be answered briefly, and has a meaning that outsiders will not understand even if overheard.

5-5. What can be known: Only the fact that an AI-generated label is not attached to this video. What cannot be known: That this video is genuine. The labeling obligation under Article 31 of the AI Basic Act applies to business operators, and fraudsters or individuals who remove the label and circulate the video may not label it, so judgment must be made based on the source and verification procedures.

5-6. First: Immediately call the customer service center of the bank from which money was sent or the bank holding the receiving account to report the damage and request a suspension of payment on the account (connection is also possible via 112 or 1332). After that (two of the following): reporting the damage to 112 and obtaining a Certificate of Incident/Accident Facts or similar documentation; applying in writing for damage relief at a branch of that bank within 3 business days from the date the payment suspension was requested by phone; deleting malicious apps or resetting the mobile phone if a malicious app was installed; or registering with the Personal Information Exposure Accident Prevention System (Privacy Exposure Accident Prevention System).

5-7. Before sending: Delete personal information, such as names, addresses, contact details, and customer numbers, from the notice. After receiving: Check the dates, times, and amounts included in the summary against the original text once more. AI can alter or omit numbers while summarizing.