Chapter 3. Verifying Answers

Learning objectives

• Identify and mark parts in an AI answer that require verification (numbers, dates, names, places, amounts, and grounds).
• Request sources, directly locate those sources on official websites, and verify whether they match the answer and when the information was written.
• Select an appropriate place to verify depending on the type of question (official institution website, official representative phone, or specialist).
• Directly visit official websites to verify, instead of using links or numbers written in text messages or AI answers.
• For health, legal, and financial matters, use AI answers only for consultation preparation and consult with specialists.

Introduction

Yeong-sik Kim was seventy-three years old. As winter approached, he was worried about heating costs and asked an AI: "Is there heating support available for a senior in their 70s living alone in Garam City?" The AI answered kindly: "You are eligible to receive Garam City's 'Warm Winter Heating Support.' Anyone who is 65 years of age or older living in a single-person household can apply; the support amount is KRW 200,000, and the application deadline is November 30. Please inquire with the Garam City Hall Welfare Division (Phone: 031-○○○-○○○○)."

Yeong-sik Kim was pleased and was about to call right away, but paused. He recalled what he had heard in the senior welfare center AI class the previous month: "Underline numbers, dates, amounts, and phone numbers when they appear." In this answer, there were as many as seven places to underline.

The next day, Yeong-sik Kim visited the community service center. A similar support program did indeed exist. However, eligibility was not for "anyone"; there were income criteria, and the application deadline was December 15, not November 30. The method of support was also a voucher to purchase heating fuel, not cash. The phone number provided by the AI belonged to an entirely different department.

The AI's answer was useful, as it provided a clue that such support might exist. However, had he believed it as is, he might have called the wrong place, prematurely given up thinking he was ineligible, or missed the deadline. This chapter covers how to find places in AI answers that need verification and where to verify them. First, you will practice finding places to underline and learn how to examine sources and dates. Next, it covers how to verify using official institution websites and phone calls, and how to choose where to verify. Finally, it examines why you must consult specialists for health, legal, and financial matters.

Key concepts

Verification is a habit, not suspicion

When receiving change at a store, people count it once. This is not out of suspicion toward the cashier, but because anyone can make a mistake. A few seconds are enough to count it, and correcting an error on the spot is easy. Realizing it after returning home is too late.

Verifying an AI's answer is the same. It is not because you cannot trust the AI, but because, as seen in Chapter 1, an AI can be plausibly wrong. In its help documentation, OpenAI also recommends that you always verify citations, references, and mentions of external documents.

Fortunately, the parts that need verification do not make up the entire answer. There is no need to verify the warm tone of a thank-you message or the sequence in an itinerary. The parts that require verification are usually a few specific places within the answer. Cultivating an eye to spot those few places is the first step of this chapter.

Underlining places to verify

When you receive an AI answer, read it while imagining underlining the following eight items.

Be particularly careful with broadly assertive terms such as "anyone," "must," "unconditionally," or "without documents." Actual programs almost always have conditions and exceptions. The AI may have omitted conditions and stated things overly simply.

Try this: Underlining the answer from the Introduction

Underlining the answer Yeong-sik Kim received looks like this:

You can receive Garam City [Warm Winter Heating Support].
[Anyone] who is [65 years of age or older] in a [single-person household] can apply,
the support amount is [KRW 200,000], and the application deadline is [November 30].
Please inquire with the [Garam City Hall Welfare Division (Phone: 031-○○○-○○○○)].

The bracketed parts are the places to verify. The remaining phrases ("You can receive", "Please inquire") are connecting words, so there is no need to verify them. Looking at it this way makes the task of verification much lighter.

Tips for re-examining numbers, names, and places

When re-examining underlined parts, keeping the following common mistakes in mind will help you spot them quickly.

Even a single different character in a name or place can make it a completely different location. In particular, for information that includes regional names, be sure to verify whether it pertains to your neighborhood or is similar information from another city.

Asking for sources

A source answers "where did this information come from?" If there is a source, you can visit it directly to verify. If the AI did not state a source, you can ask for one.

[Follow-up question]
What is the basis for the heating support information you just mentioned?
Please also provide an official website address where I can verify it.
If you do not know, please tell me you do not know.

If it is an AI with search capability, it will attach an actual website link. If it is an AI without search capability, it may say "I cannot verify this" or sometimes fabricate a plausible address. As seen in Chapter 1, OpenAI also notes that an AI can fabricate non-existent sources. Therefore, simply receiving a source is not enough; you must visit it directly and verify.

Visiting sources: Four checks

The source provided by the AI is used to find out which document from which institution to look at. Rather than clicking the link directly, type the name of the institution into a search engine or navigate to the official website saved in your bookmarks, and then look for the same document within it (see "Navigating to links directly" in this chapter). Then, check the following four items:

1. Can it be found on the official website? If you cannot find the same document or announcement on that institution's website, that source cannot be used. The AI may have fabricated the source.
2. Who wrote it? Check whether it is an official website of a city hall, government ministry, public institution, hospital, or bank, or a personal blog or promotional post.
3. Does the content match the AI's answer? Compare whether the numbers, dates, and conditions mentioned by the AI are actually written in that document as stated. Even if the document is authentic, the AI may have altered the details.
4. When was it written? Check the posting date or modification date. If it is last year's announcement, this year may be different.

Only after passing all four checks can you say it has been "verified." If it fails even one, verify it again elsewhere.

Checking dates

Even for the same content, it may be correct or incorrect depending on when the information is from. As seen in Chapter 1, an AI may not know about changes that occurred after its training cutoff. Articles on websites can also remain outdated as they are.

When checking dates, look at places like these:

• Announcements and articles: The posting date or modification date below the title or at the very end
• Support programs: The application period and program year ("2026 ○○ Support Program")
• Laws and policies: The enforcement date. For example, the National Law Information Center displays the currently applicable date below the law's title, such as "[Enforced Sept. 11, 2026]."
• Rates and timetables: Notices such as "Effective from ○ Month ○ Day"

Verifying on official institution websites

The most reliable place to verify is the institution directly responsible for that information. If it is heating support, it is the city hall or district office running that program; if it is bus times, it is that bus company or terminal; if it is hospital clinic hours, it is that hospital. Internet sites operated by such institutions are called official websites (nurijip). Nurijip is the native Korean word meaning "website."

Examples of places to visit by question type are as follows:

Even with official websites, keep one thing in mind: general information may differ from your specific case. Disease descriptions on the National Health Information Portal provide general information; a judgment suited to your body is made by a doctor. Explanations on Easy-to-Understand Practical Law Information are also general guidance; judgments regarding your specific case are made by a specialist.

Navigating to links directly

There are also precautions to take on your way to verify information. Instead of clicking links received via text message or messenger, or links written in AI answers, make a habit of visiting official websites directly through search or bookmarks.

The prevention guidelines of the Financial Supervisory Service Voice Phishing Keeper emphasize the same point. Because phishing sites do not use normal addresses, it advises that you verify whether the website addresses of financial institutions and public agencies received via text message or email are correct through internet searches or other means.

1. Type the name of the institution into the search box (e.g., "Garam City Hall").
2. In the results, skip those marked as advertisements.
3. Enter the official website of the institution.
4. Save websites of frequently used institutions in bookmarks and access them from there in the future.

Verifying by phone

If looking at a website is difficult or the content is confusing, inquiring by phone is the most reliable method. The number to call at this time is the official representative number listed on the official website. Do not call the number written in the AI's answer or received via text message directly. As in Yeong-sik Kim's case, the AI might provide the number of a different department, and fraudulent contacts seen in Chapter 5 intentionally provide fake numbers.

When verifying by phone, it is helpful to write down what to ask in advance. An AI is very helpful with this preparation.

[Prompt to send]
I plan to call the City Hall Welfare Division to inquire about the heating support program.
Please create five short questions about what I should ask to find out if I, as a senior in my 70s in a single-person household, am eligible. Do not include my name or address in the question list.

[Example AI answer]
1. Is there a heating support program this year? What is the name of the program?
2. What are the eligibility criteria (age, household, income)?
3. What is the application period and deadline?
4. Where and with what documents do I apply?
5. In what form is the support received (cash, vouchers, etc.)?

This list of questions does not contain facts, so it does not need verification. It is a good example of using an AI to prepare for verification.

Cross-checking across two or more sources

For important information, comparing across two or more sources provides greater peace of mind. This is called cross-verification. For example, if an announcement on the city hall website matches the explanation of a community service center employee, it is reliable.

There are tips for cross-verification as well. Quality is more important than quantity. A single official announcement is more reliable than ten blogs copying the same content. This is because the blogs may have all copied the same erroneous article. Therefore, it is best to set the order as follows:

1. Check the official guidance from the institution responsible for the information first.
2. If the official guidance is confusing, ask the institution by phone.
3. Use posts on blogs or online communities only as clues that "such things exist."

Place the AI's answer in the same position as item 3. While very useful as a clue, it is not a basis for verification.

Choosing where to verify

Summarizing the content so far into a single table yields the following. Where to go varies depending on what needs to be verified.

Consult doctors and pharmacists for health issues

Questions about health are among the most frequent questions asked of an AI. An AI is very helpful for explaining difficult medical terms in plain language and organizing what to ask at a hospital. However, do not leave the following to an AI:

• Determining what illness corresponds to one's symptoms
• Deciding whether to discontinue, increase, or take medications together with other drugs
• Deciding whether or not to visit a hospital

The AI has not examined you, does not know all the medications you take or your underlying conditions, and cannot be held responsible when it is wrong. General explanations from an AI may not fit your specific case.

The safe way to use an AI for health issues is to use it to prepare for a consultation.

Consult specialists for legal issues

Issues such as contracts, inheritance, neighborhood disputes, and jeonse deposits are also easy to ask an AI about. An AI is useful for explaining legal terms, describing general procedures, and compiling a checklist of materials to bring to a consultation.

However, in legal matters, the outcome varies depending on specific circumstances. A single line in a contract or a one-day difference in dates can alter the result. Moreover, as seen in Chapter 1, an AI may not know the details after a law has been amended, may conflate foreign systems, or may fabricate non-existent statutory provisions.

Therefore, follow this sequence:

1. Ask the AI to explain terminology and generate a list of questions.
2. Directly verify statutory articles on the National Law Information Center and plain-language explanations on Easy-to-Understand Practical Law Information.
3. Before making irreversible decisions such as signing, transferring money, or filing a lawsuit, consult with specialists such as attorneys or legal consultation organizations.

Address financial issues through official counters

The same applies to money-related matters such as deposits, loans, insurance, investments, and pensions. An AI is useful for clearly explaining terms like "early cancellation" and "compound interest" and organizing questions to ask when visiting a bank. However, decisions to select products and transfer money should be made only after listening to direct explanations at the official counter of the relevant financial company.

In particular, be suspicious from the start in situations such as the following:

• When an AI answer or someone's message mentions "guaranteed high returns" or "principal protection"
• When you are urged to buy a specific investment product or cryptocurrency right now
• When you are told to conduct transactions using an unfamiliar app or link instead of your bank app

Plausible investment descriptions created by AI or investment advertisements impersonating celebrities' faces are covered again in Chapter 5.

Going deeper

When you cannot verify

There are times when verification is not possible no matter how hard you search. There is no information on the official website, and phone calls do not go through. In situations like this, there are three paths:

1. Do not use it. Do not pay money, submit documents, or pass along information to others based on unverified details. This is the safest path.
2. Share it with a "not verified" note. If you must pass it along, explicitly state the fact that it has not been verified, such as: "This was provided by an AI, but I have not verified it yet. Please check for yourselves."
3. Wait. Check again the next day when the organization opens. Unless it is truly urgent, the downside of waiting a day is usually not large.

Conversely, the most dangerous moment is when you think, "I don't have time to verify it, so let's just trust it for now." As we will see in Chapter 5, fraud creates this very feeling that "there is no time."

Before sharing in group chats

There are many times when you want to share information that looks useful in family group chats or school alumni group chats: "Didn't you know this? I asked an AI, and it said this." However, posts in group chats get forwarded to other chat rooms, and forwarded again from those rooms. Even if the person who sent it first later realizes it was wrong, they cannot visit every chat room to correct it.

Therefore, set a slightly higher standard before passing information on to others.

• If it contains numbers, dates, amounts, or eligibility requirements, verify it through official notices before sharing.
• Include the name of the official website where you verified it. A single line such as "Based on the Garam City Hall website announcement" allows the recipient to verify it as well.
• Be especially cautious with health information (such as what is good for which illness). It can lead the recipient to stop or change their medications.
• If you discover that it was wrong, immediately post a correction in the same chat room.

The usefulness and limits of having AI check itself

Asking an AI, "Point out parts of your previous answer that might be wrong," can sometimes produce useful answers, such as "Operating hours may have changed, so please check." This helps in identifying where to verify.

However, you must not regard this as verification. As seen in Chapter 1, asking the AI again is like asking the same person twice. The AI can miss mistakes in its own answers, or conversely, claim that correct parts are wrong. If you ask in a suspicious tone, it may even change a correct answer.

In summary, AI can help with finding where to verify, but verification itself is done outside the AI. Official organizations, original source materials, and the people responsible for the work serve as the basis for verification.

At work

Young-sik Kim's heating subsidy application

Situation. Young-sik Kim, whom we met in the Introduction, underlined seven parts in the AI's answer and set out to verify them.

Thought process. Young-sik Kim recalled the table for choosing where to verify. For the eligible candidates and application period of a support program, the "official announcement from the responsible agency" was the top priority. He decided not to use text message links or phone numbers provided by the AI.

Discarded options.

• Calling the number provided by the AI immediately. He discarded this because the AI might have given an incorrect number. In fact, it was the number for a different department.
• Taking only the paperwork and going, since it said "anyone can apply." He discarded this because going without verifying the requirements could lead to a wasted trip.
• Asking the AI again, "Can really anyone receive it?" He discarded this because asking the AI again was not verification.
• Looking for other people's experiences on online forums. While this could provide clues, he did not treat it as a top priority because it could have been from the previous year.

Decision. Young-sik Kim searched for the Garam City Hall website, accessed it directly, and looked up "heating expenses." There was an announcement for the current year, but the income criteria were written in difficult terms. So he pasted the announcement into the AI and asked, "Please explain the eligibility section of this announcement in simple terms so that a senior citizen in their 70s can understand." After reading the simple explanation, it became clear what questions he needed to ask. The next day, he went to the community service center, checked with a staff member whether he met the income criteria, and completed his application before the December 15 deadline.

In this scenario, AI was used twice. First, it provided a clue, and later, it simplified the official announcement. In both instances, establishing the facts was done by the official organization.

Cold medicine for Mi-seon Oh's father

Situation. Mi-seon Oh's father was taking blood pressure medication. Having caught a cold, her father asked whether he could take cold medicine bought from a pharmacy together with it. Because it was 10:00 PM, their regular pharmacy was closed. Mi-seon Oh asked the AI, and the AI answered, "Generally, they can be taken together, but some cold medicine ingredients may affect blood pressure, so please exercise caution."

Thought process. Mi-seon Oh could not tell whether this answer meant "it is fine to take" or "it is not." The words "generally" and "some ingredients" did not reveal where her father's case fell. Because it concerned the body and medications, it fell under the second question ("Does it concern the body and health?") among the six checklist questions in Chapter 1, "Recognizing situations where AI answers should not be trusted blindly."

Discarded options.

• Believing that "generally, they can be taken together" and having him take it. She discarded this because her father could fall into the "some" category.
• Uploading a photo of the medicine box to the AI to assess the ingredients. She discarded this because the AI might misread the photo, and it would leave the assessment itself to the AI.
• Searching for similar experiences on online forums. She discarded this because other people's experiences could differ from her father's case.

Decision. After confirming that her father's symptoms were not urgent, Mi-seon Oh had him hold off on taking the cold medicine that night. Instead, she asked the AI, "Organize the questions I should ask the pharmacist tomorrow. I want to check whether my father in his 70s who takes blood pressure medication can take cold medicine together with it," to create a list of questions. The next morning, she took her father's blood pressure medication envelope and the cold medicine box to the pharmacy and consulted with the pharmacist. The AI prepared the consultation, and the pharmacist made the assessment.

Common mistakes

1. "A source link is attached, so it is verified information." You must check four things: whether the document can be found on the official website, whether it is an official institution, whether the content is identical, and when the text was written. A source is the starting point of verification.
2. "Calling the phone number provided by the AI will verify it." The AI can state the number incorrectly. Call the main telephone number listed on the official website.
3. "The same content is on multiple blogs, so it is correct." They may have copied the same incorrect post. A single official notice is more reliable than ten blogs.
4. "It is on the official website, so it is still correct today." Outdated announcements may remain online. Check the posting date, project year, and enforcement date.
5. "I asked the AI to point out errors, so the check is finished." The AI's review only helps find what to verify; it is not verification.
6. "An official institution's general explanation applies directly to me as well." For health and legal information, general explanations and one's specific case may differ. Your case must be assessed by a professional.
7. "There is no time to verify, so let us just believe it and act." For matters that are difficult to undo—such as money, health, and documents—you must never skip verification. The feeling that "there is no time" is also a common tactic in fraud.

Summary at a glance

• Verification is a habit. Just like counting change, look at only a few key spots rather than the entire answer.
• Eight items to underline: numbers, dates and times, names of people, names of stores, institutions, or systems, places and addresses, telephone numbers and website addresses, qualifications and rules, and grounds.
• Broad, assertive words ("anyone", "without fail", "without documents") are a signal that conditions have been omitted.
• Four source checks: Can it be found on the official website, who published it, is the content the same, and when was it written? Do not click the link directly; visit the official website yourself to find it.
• Check the dates: posting date, project year, and enforcement date. Do not make this year's decisions based on last year's information.
• The responsible institution is top priority: the administering institution for support programs, the National Law Information Center for laws, and doctors and pharmacists for health.
• Find links and numbers yourself: Visit the official website via search or bookmarks instead of using links or numbers in text messages or AI answers.
• Quality over quantity: A single official notice is better than ten blogs.
• Health, law, and money: AI is for preparing for consultations; decisions belong to professionals and official channels. For urgent symptoms, dial 119 immediately.
• When unverified: Do not use it, label it "unverified," or wait.

Check the official docs

• Ministry of Government Legislation National Law Information Center: Original text and enforcement dates of statutory provisions. Check the "[Enforced ...]" mark below the law title.
• Ministry of Government Legislation Easy-to-Understand Everyday Law Information: A website that explains everyday legal issues in plain language.
• Korea Disease Control and Prevention Agency (KDCA) National Health Information Portal: General information on diseases and health.
• National Fire Agency — 119 Emergency Reporting Guide: What to report to 119 when there is an emergency patient.
• Financial Supervisory Service (FSS) Voice Phishing Guard — Prevention Tips: Guidance on verifying website addresses received via text message or email through search.
• OpenAI — Does ChatGPT tell the truth?: Recommendation to always verify citations, materials, and references to external documents, and to visit links directly.

Self-check

3-1. Write down at least five of the eight items you should underline when reading an AI response.

3-2. An AI provided a source link. In one line, write down how you should navigate to this source to say you have "verified" it, and write down the four things you need to check.

3-3. Write down the top-priority place to verify each of the following pieces of information: (a) This week's flu vaccination hours at the local public health center (b) From when a certain law enters into force (c) Whether two medications you are taking can be taken together

3-4. Write down two reasons why it is better to directly look up the official website and main telephone number instead of using a link received by text message or a number provided in an AI answer.

3-5. The same content appears on five blogs, but the official website does not have that content. Write down how this information should be handled.

3-6. When using AI for health issues, write down two "tasks suitable to delegate to AI" and two "questions to ask a doctor or pharmacist."

3-7. You want to share information about a subsidy program in a family group chat. Write down what to do before sharing it, and write down one sentence that is good to include when sharing it.

Hands-on tasks