Chapter 4. Personal Data and Safe Usage

Learning objectives

• Explain the meaning of personal information, and identify cases where individually harmless pieces of information combine to identify a person.
• Explain why resident registration numbers, account and card numbers, passwords, and authentication codes should not be entered into AI.
• Distinguish between information that can and cannot be entered into AI using a three-color traffic light system.
• Rewrite the same question safely by removing, replacing, or broadening personal information.
• Protect conversations using chat history, training settings, temporary chats, and logging out on shared devices.
• State what to check before entering other people's information, photos, or company and school materials.

Introduction

Young-hee Choi wanted to file an insurance claim after receiving treatment for her knee. Because the paperwork was complicated, she decided to get help from an AI. She took photos of her medical bill receipts and diagnosis certificate with her mobile phone. The photos clearly showed her name, resident registration number, home address, disease name, and part of the payment card number. She was about to upload the three photos to the AI chat window, type in her password for the insurance company website, and ask, "Please write an insurance claim for me using this."

Her son, who was watching from beside her, stopped her. "Mom, wait a minute. What you upload here does not just stay inside your phone; it goes to the AI company's computers. And it gets stored there." Young-hee Choi was surprised. "I thought I was just using it by myself on my phone."

Many people feel that an AI chat window is like a diary. It feels like a space only they see. However, the text written and photos uploaded in the chat window are sent to the company operating the service and stored there. Depending on the settings, they may be used to improve the service, or in some cases, reviewed by humans.

This does not mean that one should avoid using AI altogether. Even Young-hee Choi could have received sufficient guidance on the list of required documents and the procedure for filing an insurance claim if she had asked while masking her name and numbers. This chapter covers those methods. First, it looks at what personal information is and where text entered in the chat window goes. Next, it covers how to distinguish between information that must never be entered and information that can be entered using a traffic light system. Then, it sequentially addresses how to ask questions after removing personal information, how to handle photos and other people's information, chat history and training settings, and how to log out on shared devices.

Key concepts

What is personal information

Personal information refers to information relating to a living individual that makes it possible to identify that individual. Article 2, Subparagraph 1 of the Personal Information Protection Act (PIPA) classifies personal information as follows:

• Information that can identify a specific individual through name, resident registration number, images, etc. (Item (a))
• Information that, even if it cannot identify a specific individual by itself, can easily be combined with other information to identify the individual (Item (b))
• Information resulting from pseudonymizing the information in Item (a) or (b) so that the individual cannot be identified without additional information to restore it to its original state—that is, pseudonymous information (Item (c))

The second item is crucial. Even without using a name, someone's identity can be determined when multiple pieces of information are combined. Consider an example:

• "A man in his 70s" → It is impossible to know who he is.
• "A man in his 70s, Building 102, Garam Apartments" → The scope narrows.
• "A man in his 70s, Building 102, Garam Apartments, retired elementary school principal, had knee surgery last month" → Neighbors will immediately know who he is.

Although there is not a single name, the third sentence effectively points to that specific person. Therefore, removing personal information does not end with merely deleting a name. One must also consider pieces of information that, when combined, make it possible to identify who the person is.

Where text written in the chat window goes

AI chatbots generally do not run solely inside one's phone. The text and photos sent travel over the internet to the service company's computers (referred to as servers), where the AI generates answers and sends them back. Conversations exchanged in this way are usually stored as chat history, allowing them to be reopened and viewed later.

How stored conversations are used depends on the company and the settings. As of the verification date, the official guidance from the companies is as follows:

In the Gemini Apps Privacy Hub, Google clearly states that if this setting is on, "Please do not enter confidential data or any data you wouldn't want a reviewer to see or Google to use to improve its services."

Things never to enter ①: Passwords and authentication codes

A password is the key to an account. An authentication code is a multi-digit number sent via text message by a bank or website to verify "Is this really you?" The rule for both is never to disclose them to anyone. The same applies to family members, people claiming to be bank employees, and AI.

There is never a case where AI needs a password or authentication code to answer a question. Do not make requests such as "I will give you my password for the insurance company website, so please log in and file the claim." Conversations containing passwords are saved, and other people may be able to see those records later.

Things never to enter ②: Resident registration numbers

A resident registration number is a number assigned uniquely to each person in South Korea, and it rarely changes throughout a lifetime. If a password is leaked, it can simply be changed, but a resident registration number cannot be changed so easily.

South Korean law also treats resident registration numbers with great gravity. Article 24-2 of the Personal Information Protection Act (PIPA) strictly limits the circumstances under which a personal information controller (an institution, company, organization, etc., that processes personal information in the course of its operations) may process resident registration numbers to cases specifically required or permitted by statutes or presidential decrees, or where manifestly necessary for the urgent benefit of life, body, or property. Even when processed, they must be stored securely through encryption (Paragraph 2), and during the step of registering as a member on an internet website, an alternative method of registration without a resident registration number must be provided (Paragraph 3).

While this law defines obligations for companies and institutions rather than directly regulating what an individual writes in a chat window, the message is clear: it is information of such gravity that even institutions cannot handle it carelessly. There is no reason to volunteer and enter such information into an AI chat window. Even when asking an AI how to fill out documents, it is sufficient to leave the resident registration number field blank, such as "○○○○○○-○○○○○○○".

Things never to enter ③: Account and card numbers, and ID photos

Do not enter the following into the chat window either:

• Account numbers and card numbers, security codes on the back of cards, expiration dates
• Photos of identity documents (resident registration card, driver's license, passport)
• Copies of bankbooks, photos of payment cards
• Photos of banking app screens (balances, transaction histories)

Such information is directly connected to money. If leaked, one could suffer identity theft or financial loss. If you want to ask an AI to organize household accounts or analyze expenses, remove all numbers and ask by listing only categories and amounts, such as "food expenses KRW ○○,000, transportation expenses KRW ○○,000".

Distinguishing with traffic lights

Dividing information into three traffic light categories—information that may and may not be entered—makes judgment quicker.

Notice that "disease names and medical records" are in Yellow. Health information is specifically treated as sensitive information even under the Personal Information Protection Act (Article 23). However, this does not mean one cannot discuss health at all. Stating only the circumstance, such as "in their 70s with a bad knee," is generally acceptable. If original diagnosis certificates, hospital names, treatment dates, and personal names are included together, it shifts from Yellow toward Red.

Asking questions after removing personal information

There are four ways to safely modify a question that contains Yellow information:

Try it this way: Rewriting an insurance claim question by masking

[Before masking — Do not send]
I am Young-hee Choi (born in 1961, resident registration number 61XXXX-2XXXXXX),
and I had knee cartilage surgery at Garam Orthopedic Clinic on October 3.
I will upload a photo of the receipt. I paid with ○○ Card 1234-...,
and my insurance company ID is young61 and password is ○○○○. Please file the claim for me.

[After masking — Send like this]
A woman in her 60s had knee surgery and wants to file an actual-expense health insurance claim.
Please provide a numbered list of the documents usually required and the general procedure for filing a claim.
For parts that may differ depending on the insurance company, please mark them as "check with the insurer."

In the masked question, the name, date of birth, resident registration number, hospital name, date, card number, ID, and password have all been omitted. Even so, there is no problem in asking what Young-hee Choi wants to know—namely, the required documents and procedure. The last line, as learned in Chapter 3, is a request to indicate parts that may vary across insurance companies. Young-hee Choi files the actual claim herself through the insurance company's official app or service desk.

When uploading photos and documents

Asking questions to AI by uploading photos is very convenient. However, photos contain far more information than text. Scan the photo once before uploading it.

There are three things to do before uploading:

1. Does it have to be a photo? If you only need three amounts from a receipt, writing down only the numbers in text is sufficient.
2. Have you kept only the necessary parts? Crop out unnecessary parts using the photo editing features on your mobile phone.
3. Have you covered the areas that need masking? If names or numbers are inevitably captured, cover them by drawing over them with the pen in the editing tool.

Information and photos of other people

Something to be just as careful about as your own information is other people's information. Confiding stories about family or friends to AI is more common than one might think.

• Taking a full screenshot of a family group chat and asking, "Please summarize this"
• Describing a friend's situation in detail, saying, "My friend ○○ got divorced; please write a comforting text message"
• Uploading a grandchild's facial photo and asking, "Turn this into a drawing"
• Pasting a roster of club members (names, phone numbers) and asking, "Organize this in alphabetical order"

The owner of such information is that person, not yourself. That person does not know that their story is being stored on the AI company's servers and has never consented to it. Therefore, adhere to the following principles:

1. Change names and describe only circumstances. Saying "A friend is going through a difficult time" is more than enough to write a comforting message.
2. Do not upload group chat screenshots. If a summary is needed, transcribe only the necessary parts yourself and delete names.
3. Do not upload photos of another person's face without their permission (or parental permission for a child). Be especially cautious with children's photos.
4. Do not paste rosters as they are. If you only want to ask how to organize them, ask using a dummy list with names changed to "Member 1," "Kim ○○," etc. Organize the actual roster yourself using phone notes or spreadsheet programs instead of AI.

Chat history: What remains and what to delete

Most AI apps display past conversations as a list. This is chat history. Chat history is convenient because one can look up yesterday's answer again. However, if there is personal information in that history, someone sharing the phone or another person accessing the account can also view it.

Chat history can be deleted. However, deleting it does not mean it disappears from everywhere immediately. As of the verification date, the guidance from the companies is as follows:

• OpenAI: The help article "Deleting and archiving chats in ChatGPT" explains that deleted conversations disappear immediately from the account screen and are scheduled for permanent deletion from OpenAI systems within 30 days. However, it notes exceptions where data has already been de-identified (processed so that the individual cannot be identified) and disassociated from the account, or where it must be retained for security or legal reasons.
• Google: It states that while "Gemini Apps" chats can be manually deleted at any time, chats (and related data) that have already been reviewed by human reviewers are not deleted even when the activity is deleted.

Therefore, order matters. Not entering it in the first place comes first, and deleting it is a secondary remedy.

Training settings: Whether to use my conversations for AI improvement

AI companies may use user conversations to make AI better (referred to as model improvement or training). They provide settings so users can decide this. In this book, these are called training settings. As of the verification date, the official guidance is as follows:

There are exceptions in the Gemini app. Google states that even if Keep activity is turned off, if a user submits feedback, that feedback and related conversations are collected and used to maintain and improve Gemini apps, and the feedback is reviewed by human reviewers. Furthermore, even with the setting off, it notes that chats continue to be used in various ways, including with the assistance of human reviewers, to answer users and protect Google, users, and the public.

Setting names and locations may change as apps are updated. Therefore, rather than memorizing exact menus, the important habit is to check whether the AI you use has training settings and how they are currently configured. If unsure, search for "data," "privacy," or "training" in the service's help center.

Temporary chats

Several services provide separate conversations that are not kept in history. As of the verification date, the official guidance is as follows:

• ChatGPT's Temporary Chat: Does not appear in chat history, does not create or update memories, and is not used to improve OpenAI models. However, it may be kept for up to 30 days for safety purposes. If a temporary chat is saved, it becomes a regular conversation and follows the account's settings.
• Claude's Incognito chat: Is not used to improve Claude, even if Model Improvement is turned on in Privacy Settings.
• Gemini app's Temporary chat: Is kept in the account for 72 hours and is not used to train Google AI models.

Temporary chats are suitable for quick inquiries and tasks where you prefer no record to remain. However, there is something to remember: temporary chats also pass through company servers and are retained, even if briefly. A temporary chat does not mean that Red information is allowed.

Before entering company and school materials

Although this course does not cover AI for business use, there are many "materials that do not belong to me" in daily life as well:

• Customer rosters or internal documents from workplaces (including part-time and hourly jobs)
• Member information and meeting materials from volunteer organizations, clubs, or apartment tenants' associations
• Newsletters from a grandchild's school, classmate lists, teacher contact information

Check two things before entering such materials into AI:

1. Does the organization have AI usage rules? A company, school, or organization may have established rules such as "Do not enter materials into external AI." If unsure, ask the person in charge first.
2. Does it contain other people's information? If it contains member rosters, customer information, or classmates' names, it is Yellow or Red. Remove it before entering, or do not enter it at all.

General notices that contain no other people's information, such as school newsletters, are great materials to ask for summaries, such as "Please summarize only what guardians need to do from this notice." However, it is better to remove school, class, and teacher names before entering them.

Logging out on shared devices

Be especially careful when using AI on computers at libraries, welfare centers, and community service centers, or on tablets shared with family. If you leave the seat while still logged in, the next person can see your chat history as it is. Chat history may contain concerns you confided, family stories, or health matters.

Try it this way: Using public computers safely

Before and during use:

1. When logging in, do not check boxes such as "Stay logged in" or "Automatic login" if present.
2. When the browser (internet window) asks "Save password?", click "Never" (or "Do not save").
3. If possible, use temporary chat features, and do not enter personal information from the start.

Three things to do before leaving:

4. Click Log out in the account menu. Merely closing the window might not log you out.
5. Reopen the AI website to check if the login screen appears. If the login screen appears, you are logged out.
6. Verify that your password or automatic login was not saved in the browser. If accidentally saved, delete it in browser settings.

If it is a family tablet, it is advisable to use separate accounts for yourself and your family members.

Whom and how the law protects

The Personal Information Protection Act (PIPA) primarily imposes obligations on personal information controllers—that is, institutions, companies, organizations, or individuals that process personal information for the operation of personal information files in the course of their work (Article 2, Subparagraph 5). Companies operating AI services must also comply with this act when handling users' information.

Furthermore, this act grants rights to data subjects, who are the owners of the information. Article 4 provides that data subjects have the right to be informed about the processing of their personal information, the right to confirm whether it is being processed and to request access, the right to request suspension of processing, correction, deletion, and destruction, and the right to prompt and fair redress for damages.

The habits in this chapter (not entering, masking, deleting, and checking settings) serve as an additional safety belt worn over the rights protected by law. The existence of the law does not mean one does not need to be careful, nor does being careful relieve companies of their legal obligations.

Going deeper

Even if each piece is fine on its own, when combined

Earlier, we saw that when "in their 70s, Garam Apartments, retired principal" come together, who the person is becomes revealed. This does not happen only within a single question.

• Monday conversation: "I am in my 70s and my knees are bad."
• Wednesday conversation: "I live in Garam-dong; please tell me about orthopaedic clinics nearby."
• Friday conversation: "What kind of volunteer work can I do after retiring as a principal?"

Each individual conversation is close to green. However, in the chat history of the same account, these three remain together. Furthermore, some services have a feature that remembers (memory) conversation contents and uses them in subsequent conversations. The reason OpenAI's help center explicitly explains that temporary chats "do not create or update memories" is precisely because of this feature. While the memory feature is convenient, it also means that pieces you mentioned in separate conversations are gathered into one place.

Therefore, it is good to review the contents of your chat history and memory feature from time to time. Some apps show what the AI remembers and allow you to delete it. Look up "memory" in the help documentation of the app you use.

Why "Turning off Training" alone is not enough

Turning off the Training setting provides peace of mind. However, the Training setting is only one of multiple layers of doors. Dividing what the Training setting changes and what it does not change looks like this:

Therefore, while the Training setting is a setting worth checking, it is not a magic button that makes it acceptable to enter red information. The most reliable protection is still not entering it from the start.

If you have already entered it

If you entered red information by mistake, do not panic, and take the following steps in order:

1. Delete that conversation. While not a complete solution, it reduces the possibility that others will see it in the chat history.
2. If you entered a password, change it immediately. If you use the same password elsewhere, change it in all of those places as well.
3. If you shared a verification code with someone, or if you suspect your account or card number has been used for fraud, contact the customer service center of the relevant financial institution immediately.
4. If your personal information has leaked through phishing, follow the response guidelines of the Financial Supervisory Service (FSS) Voice Phishing Keeper. These guidelines explain steps such as registering with the FSS Personal Information Exposure Accident Prevention System (pd.fss.or.kr) in the event of personal data leaks, and checking whether mobile phones have been opened in your name through the Korea Association for ICT Promotion (KAIT) identity theft prevention service (www.msafer.or.kr).

At work

Choi Yeong-hee's insurance claim preparation

Situation. Choi Yeong-hee from the Introduction listened to her son and did not upload the three photos. Instead, she rethought what and how to ask.

Thought process. What Choi Yeong-hee wanted to know was "what is needed to claim insurance benefits and in what order it should be done." She divided the information needed for this question into traffic lights. Name, resident registration number, card number, and insurance company ID and password were red. Hospital names, surgery dates, and medical certificates were yellow. "Woman in her 60s," "knee surgery," and "indemnity health insurance" were green.

Discarded options.

• Uploading the photos while only turning off the Training setting. She discarded this because even if Training was turned off, the photos would still go to the server and remain in the history.
• Uploading the photos in a temporary chat. She discarded this because temporary chats also pass through the server and are retained for a certain period, so this did not justify entering red information.
• Resolving it alone without using AI. She discarded this because AI was a great help in organizing the document checklist and procedure.

Decision. Choi Yeong-hee asked questions using only green information (the prompt after masking in "Asking questions after removing personal information" in this chapter). The AI organized the list of documents generally required and the claiming procedure, adding a note to "check with the insurance company" for parts that varied by insurer. Taking that checklist, Choi Yeong-hee filed the claim directly in the insurance company's official app. She uploaded the receipt photos only to the insurance company's official app, not to the AI.

Park Jeong-ho's welfare center computer

Situation. Park Jeong-ho was preparing the host's greeting for his daughter's wedding next month. Because he did not have a computer at home, he borrowed a computer in the digital classroom of a community welfare center. He refined the greeting while sharing memories with his daughter, stories about his son-in-law's family, and even the story of his wife passing away a few years ago with the AI. He printed out a greeting he liked and went home with a joyful heart.

Problem. That evening, he received a phone call from a neighbor who had used the same computer at the welfare center. "Sir, the AI screen was open under your account. I could see the whole chat history, so I quickly closed the window." Park Jeong-ho had not logged out.

Thought process. Park Jeong-ho realized two things. First, closing the window and logging out were different. Second, in writing the greeting, there was no need to write in detail about his wife's story or the son-in-law's family circumstances. A single line mentioning the circumstance that "I lost my wife a few years ago" was sufficient.

Discarded options.

• Letting it pass as fine since the neighbor closed the window. He discarded this because if the next person opened it again, the logged-in state could continue.
• Deleting the AI account altogether. He discarded this because it was an excessive reaction, and he intended to continue using AI in the future.

Decision. The next morning, Park Jeong-ho went to the welfare center, logged out of that computer, and confirmed that the login screen appeared. At home, he changed his account password on his mobile phone and deleted the conversation containing detailed circumstances of his son-in-law's family. After that, he formed the habit of minimizing personal stories from the start on shared computers and logging out every time he left his seat.

Common mistakes

1. "I use it alone on my mobile phone, so only I see it." Conversations go to the company's servers and are stored. Depending on settings, they may be used for service improvement or reviewed by humans.
2. "If I just remove the name, it is not personal information." When age, neighborhood, occupation, and events are put together, who the person is can be determined. Look at the combined pieces together.
3. "Since I turned off the Training setting, I can enter anything." The Training setting only changes the use for Model Improvement. Conversations going to the server, remaining in history, safety reviews, and exposure on shared devices may remain unchanged.
4. "Temporary chats do not remain anywhere." Temporary chats also pass through the server and, depending on the service, are stored for a certain period.
5. "If I delete a conversation, it disappears everywhere immediately." Companies indicate that conversations may still be retained for a certain period after deletion, or that conversations that have already been reviewed may remain.
6. "Family stories are my own stories, so I can use them as I please." Family members are also other people. Change names, describe only the circumstances, and obtain permission for face photos.
7. "Since I closed the window, I am logged out." Even if you close the window, your login session may be maintained. On shared devices, be sure to click log out and verify.
8. "A resident registration number is needed when filling out documents, so I must also provide it to the AI." When asking an AI how to fill out a document, leave the number field blank. Use actual numbers only at official counters.

Summary at a glance

Check the official docs

• National Law Information Center — Article 2 of the Personal Information Protection Act: Definitions of personal information, data subject, and personal information controller.
• National Law Information Center — Article 24-2 of the Personal Information Protection Act: Limitation on the processing of resident registration numbers, encrypted storage, and alternative means at the membership registration stage.
• National Law Information Center — Article 4 of the Personal Information Protection Act: Rights of data subjects.
• OpenAI — Data controls in ChatGPT: The "Improve the model for everyone" setting, Temporary Chat.
• OpenAI — Deleting and archiving chats in ChatGPT: How to delete conversations, when deleted chats are permanently deleted, and exceptions.
• Google — Gemini Apps Privacy Hub: Retention of activity history, human review, 72-hour retention, and notice not to enter confidential data.
• Anthropic — Is my data used for model training?: The Model Improvement setting, Incognito chat, and exceptions for feedback and safety reviews.
• Financial Supervisory Service Voice Phishing Keeper — How to respond to voice phishing damage: Actions to take when personal information is leaked.

Self-check

4-1. Create an example of a sentence that allows someone to be identified even without a single name in it, and write down what to broaden or remove from that sentence to make it safe.

4-2. Classify the following information into red, yellow, and green: (a) a verification code received by text from a bank (b) "in their 60s, has a bad back" (c) a photo of a granddaughter's face (d) the contents of a general notice received from a community service center (e) a photo of the back of a card (f) a friend's name and phone number.

4-3. Rewrite the question below using the appropriate method among the four methods in this chapter (removing, replacing, broadening, and cutting): "I am Soon-ja Park (010-○○○○-○○○○), and my son Min-soo Kim, who lives at 15 Garam-dong, is moving next week. Write a moving congratulations text message."

4-4. Write down one thing that changes and two things that do not change when you turn off the training setting.

4-5. Write down one situation where it is good to use a temporary chat, and one thing you should not do even in a temporary chat.

4-6. Close the book and write down steps 4 through 6 of "Using Shared Computers Safely" in this chapter from memory. Then, add one line explaining why merely closing the browser window is not enough.

4-7. Choose one AI service that you use (or intend to use), and write down how to check whether it has a training setting, where it is located, and how it is currently set. If there is no service you use yet, write down what you would search for to find it.

Hands-on tasks