Arah · Security and AI
AI tools feel like a private conversation, but every sentence you type travels to someone else's server. This guide shows which data you should never send, which data is acceptable under conditions, and how to mask it before you press Enter.
Introduction
Rina works in human resources at a distribution company with 40 employees. Near the end of the year she has to prepare a summary of performance reviews and a proposal for salary increases. Her spreadsheet holds names, national identity numbers (NIK), birth dates, salaries, performance scores, and disciplinary notes. At nine in the evening, tired, she copies the whole table into an AI assistant and writes, “Summarize this and give me a proposal.” Two minutes later a tidy answer appears. Rina feels efficient. She does not stop to think that the 40 people whose data she just sent know nothing about it.
Rina is a fictional character, but her habit is real. In 2023, several Samsung engineers were reported to have entered confidential source code into ChatGPT to help fix a database, and other employees pasted a full meeting recording to get minutes written. On 1 May 2023 the company banned the use of generative AI tools for work. In August 2025, a number of ChatGPT conversations shared through links were found in Google search results. They included full names, email addresses, CVs, and personal confessions. The provider later removed the feature that made conversations discoverable.
Many people face this problem. Employees speed up their work. Freelancers edit client documents. Small business owners ask for advice on financial reports. Developers paste code to find bugs. Students upload research data about respondents. Parents ask about their child's lab results. All of them type into the same box, with the same feeling that this is just a quick chat.
This decision matters for three reasons. First, data that has been sent cannot be reliably taken back. Second, some of the data is not yours, so the harm falls on people who had no say in the decision. Third, Indonesia has Law No. 27 of 2022 on Personal Data Protection (the PDP Law), which sets duties for anyone who processes personal data. It also carries criminal penalties for unlawfully disclosing and using personal data.
This article sets out a way of thinking you can use every time you are about to send something to an AI tool. You will learn the terms that people often mix up, a three-zone framework for sorting data, how to choose the right channel, and how to mask data without defeating your purpose. At the end you will find a sensitive-data scanner that runs in your browser, a decision checker, and a worksheet. After reading, you should be able to decide, before pressing Enter, whether a piece of data may be sent, through which channel, and in what form.
This article gives general information and does not replace legal or cybersecurity advice from a professional who knows your situation. AI providers change their policies from time to time, so check the settings and terms of your own account. All figures and sample data in this article are illustrations and fake data.
Short answer
No single list of prohibitions fits everyone, because the answer depends on the type of data, the channel you use, and the rules that bind you. One basic rule holds up. Do not send data that would harm you or someone else if provider staff read it, if it were stored for a long time, if it leaked, or if it appeared in the wrong place.
You apply that rule through five steps, in order.
- Classify the data. Mark each type of data you want to send as red, yellow, or green.
- Choose the channel. A personal account, a business account under contract, API access, and a local model give different levels of protection.
- Reduce and mask. Send only the part you need, and replace identities with codes.
- Set up the account. Check model training, history, memory, share links, and connectors to other apps.
- Prepare a plan for when it has already happened. Decide in advance what you will revoke, whom you will tell, and what you will record.
Three kinds of data do not belong, as they are, in a personal account on a public AI service. These are credentials such as passwords and API keys, specific personal data that can still be linked to a person, and secrets bound by a legal duty or a confidentiality agreement. Beyond these, AI can still safely help with a lot of work if you mask the data first.
This method gives you more measured decisions. It does not give you guarantees. The scanner and the worksheet help reduce carelessness, but they cannot read context and cannot guarantee that your data is safe on the provider's side.
Definitions and basic concepts
Several terms below are often used interchangeably even though they mean different things. The differences decide which channel you may use.
Personal data: general and specific
The PDP Law divides personal data into two groups in Article 4. This split is a useful starting point, because specific data calls for more care.
| Group | What it includes | Examples in everyday work |
|---|---|---|
| Specific personal data | Health data, biometric data, genetic data, criminal records, children's data, personal financial data, and other data as provided by law and regulation | Lab results, fingerprints or face photos used for verification, student data, payslips, bank statements, disciplinary records of a legal nature |
| General personal data | Full name, gender, nationality, religion, marital status, and personal data that is combined to identify a person | Customer name lists, email addresses and phone numbers, a combination of job title, division, and age that points to one person |
The last item in the general data group deserves a second reading. Data that looks harmless, such as job title, division, and age, can become personal data when the combination matches only one person. Removing the name alone does not necessarily make data anonymous.
Credentials and trade secrets
Credentials are anything that proves you have the right to access something. They include passwords, PINs, OTP codes, API keys, access tokens, private keys, database connection strings, and crypto wallet recovery phrases. Credentials have no negotiable level of sensitivity. Once sent, assume the credential is exposed and needs to be replaced.
Trade secrets, or confidential business information, are information that has value because competitors do not know it. Examples are a company's source code, product formulas, customer lists, special pricing, product plans, unreleased financial reports, and documents covered by a non-disclosure agreement (NDA). Trade secrets are not automatically personal data, but they can carry contractual duties that are just as heavy.
Data subject and data controller
Under the PDP Law, the data subject is the person whose data is processed. The data controller is the party that decides the purpose and the means of processing. Rina does not own the data of those 40 employees. The company controls it, and the employees are the data subjects. When Rina sends the table to an AI service, she acts on behalf of the company. Whether the AI provider acts as a processor or in some other role depends on the contract and the terms of service in use, so your organization's legal or data protection team needs to assess this.
Stored, read, trained on, shared
The word “privacy” is often used for five different things. Separating them shows you which setting actually answers your concern.
| Stage | What it means | What you should check |
|---|---|---|
| Sent | Text, files, or images move to the provider's servers | The country and company behind the service, and the terms of service |
| Stored (retention) | The conversation is kept for a set period, including after you delete it from view | How long it is kept and what happens when you delete it |
| Read | Provider staff, contractors, or partners can access the content for security, abuse prevention, or service development | Who is allowed to read it and under what circumstances |
| Trained on | The content of conversations is used to improve future models | Whether training is on by default and whether you can turn it off |
| Shared | Links, team workspaces, or connectors make the content visible to other people or systems | The list of share links, team members, and connected apps |
Turning off model training answers only one of those five stages. The data is still sent, can still be stored, and can still be read under certain conditions. So “I turned off training” does not automatically mean your data is safe to send.
Personal account, business account, API, and local model
The channel you use also shapes how your data is protected. The terms below are quoted from the official help pages and announcements of two providers at the time of writing, and they may change.
- Personal (consumer) account. OpenAI states that for individual services such as ChatGPT, content may be used to train models unless the user opts out. Anthropic updated its consumer terms in 2025 so that users of the Free, Pro, and Max plans choose whether their conversations may be used for training. If they allow it, data is kept for up to five years. If they do not allow it, the retention period is 30 days.
- Business account and API. OpenAI states that ChatGPT Team, ChatGPT Enterprise, and the API Platform are not used to train models. Anthropic states that the updated consumer terms do not apply to the Team and Enterprise plans, to government and education services, or to API use.
- Local model. A model that runs on your own device or server does not send data to a third party. Its security becomes your responsibility again, including the device, plugins, and who can access that machine.
Pseudonymization, anonymization, and masking
Pseudonymization replaces identities with codes, for example “K07” for one employee, while you keep the mapping table separately. The data can still be linked back to the person if the mapping falls into the wrong hands. Anonymization removes any reasonable possibility of linking the data to an individual, and its standard is far higher. In everyday practice, the realistic option is masking. It means removing direct identifiers, generalizing values (age becomes a range, salary becomes a band), and making sure no combination of attributes points to one person.
The risk that comes with masking is re-identification, meaning people can be recognized again from the attributes that remain. A common rule of thumb is to make sure every combination of attributes is shared by at least five people. This rule is a heuristic and not a legal standard, but it helps you catch cases such as the only manager in a particular division.
Data hidden inside files
Files carry more than what you see on the screen. Office documents store the author's name, the company name, revision history, comments, and struck-through or hidden text. Spreadsheets can contain hidden sheets, hidden columns, and formulas that refer to other files. Photos store metadata such as the time they were taken and, on many devices, location coordinates. Screenshots often include notifications, tab names, and other data at the edges of the image. When you upload a whole file to an AI tool, all of that is sent too.
For this reason, uploading a file is riskier than copying a section of text that you chose yourself. If you must upload, make a clean copy first. Accept or reject all tracked changes, delete comments, delete hidden sheets and columns, and then check the document properties. For images, remove the location metadata or crop out the parts you do not need.
Common misconceptions
| Belief | What is more accurate |
|---|---|
| “I deleted the chat, so it's gone.” | Deleting a conversation from your history does not delete share links that were created before. The provider may also keep the data for a set period under its terms. |
| “I turned training off, so it's safe.” | Training is only one stage. Storage, staff access, and the risk of hacking remain. Data that already went into training earlier is not removed when you change the setting. |
| “The biggest risk is that the AI leaks my data to other users.” | The UK's National Cyber Security Centre (NCSC) states that models do not automatically add your queries to share with other users. The more realistic risks are that queries are stored, read by the provider or its partners, hacked, or accidentally made public. |
| “I removed the name, so it's anonymous.” | A combination of job title, division, age, and dates can point to one person. Indirect identifiers need to be checked too. |
| “A local model is automatically safe.” | The data does not leave your machine, but the device, plugins, and access to that machine can still be weak points. |
Analysis framework
This framework is called the three-zone method. It makes the decision to send or not send fast enough to use every day. It is also structured enough that you do not have to rely on a feeling when you are in a hurry. Each step names the information you need to prepare.
- Task and needsWhat you ask the AI to do
- Data inventoryTypes and owners
- ZoneRed, yellow, green
- RouteAccount, contract, local
- MaskingReduce and replace
- Settings and recordsAccount, log, plan
Step 1. Write down the task and the minimum data it needs
Start with a question. What am I actually asking the AI to do, and what data does it need to see to do it? Often the task is general, such as tidying the wording, summarizing the structure, or suggesting a format. The personal data only gets copied along because it is in the document. Write the task in one sentence. Then mark the parts of the document that are truly needed. Data minimization is the cheapest and most effective step, because data you do not send cannot leak.
Step 2. Make a data inventory
List the types of data in the material you plan to send, and who owns them. Some guiding questions follow. Are there names, identity numbers, contact details, financial data, health data, children's data, trade secrets, code, or credentials? Is this my own data or someone else's? Is there a confidentiality marking, such as “confidential” or “internal only”, or a clause in a contract? Data that belongs to other people raises the stakes, because they bear the consequences, not you.
Step 3. Assign a red, yellow, or green zone
| Zone | Characteristics | Examples | Treatment |
|---|---|---|---|
| Red | Large or hard-to-reverse harm if exposed | Passwords, PINs, OTPs, API keys, crypto recovery phrases; card numbers and CVV; national identity number (NIK), passport, taxpayer number (NPWP); medical records with identity; children's data; NDA documents and company source code; customer or employee data that still identifies people | Do not send as is to a personal account. Credentials are never sent. Anything else only after masking, or through a contracted and approved route. |
| Yellow | Limited harm, but it can become large when combined with other data | Work names and emails, phone numbers, job titles and divisions, email drafts with client names, internal meeting summaries, aggregated sales data | Mask it or send only what is needed. Check the route and the account settings first. |
| Green | Already public, or it points to no one and has no secret value | General questions, published text, rough ideas without names, fake example data | Allowed on any route your policy permits. |
If one piece of material contains several zones, the treatment follows the highest zone. A single NIK in the middle of a table turns the whole table red.
Step 4. Choose the route
Once the zone is clear, match it to a route. The matrix below is a general guide, not law. Your office rules, contracts, and sector regulations may be stricter.
| Type of data | Personal account, default settings | Personal account, training turned off | Business account or contracted API | Local model |
|---|---|---|---|---|
| Credentials | Do not send | Do not send | Do not send | Do not send |
| Specific personal data or business secrets with identity | Do not send | Mask first | Careful, written approval | Careful |
| General personal data | Mask first | Only what is needed | Allowed per policy | Allowed |
| Internal material without identity | Careful | Allowed | Allowed | Allowed |
| Public or general information | Allowed | Allowed | Allowed | Allowed |
Credentials are marked “do not send” on every route because there is no strong reason for an AI to hold them. If the task is to fix code that uses a key, replace the key with a placeholder such as API_KEY_DISINI. Then fill in the real value yourself in a safe place.
Step 5. Reduce and mask
Masking uses five techniques, starting from the simplest.
- Delete. Remove columns or paragraphs that the task does not need.
- Replace with codes. Names become K01 to K40, and a client company becomes “Client A”. Keep the mapping table somewhere separate.
- Generalize. Dates of birth become age ranges, salaries become bands, and addresses become a city or district.
- Merge small groups. A group with fewer than five members is combined with another group so that it does not point to one individual.
- Replace with fake data. If the AI only needs the shape of the data, for example to write a formula or code, send a made-up example with the same structure.
After masking, reread the whole text before sending. Automatic scanners catch patterns such as an NIK or an API key. You have to check names, addresses, and context that points to a person yourself.
Step 6. Set up your account and habits
Check the settings that affect the five stages of the data journey. These are model training (turn it off if available and suited to your needs), history and memory, sharing links (delete the ones you do not use), connectors to email, cloud storage, and calendar (limit access), account security (unique passwords and two-step authentication), and separation of work and personal accounts. Do not use a personal account for work material unless office rules allow it.
Step 7. Keep a record and prepare a plan for when data is sent by mistake
Briefly record the route and the reason for any decision that involves other people's data, so you can explain it if asked. Also prepare a plan for when something is sent by mistake. This plan is covered in the examples and the FAQ. If your organization is a data controller and a failure of personal data protection occurs, Article 46 of the PDP Law requires written notice to the data subjects and the authority within 3 x 24 hours at the latest. The notice must contain at least the data that was exposed, when and how it was exposed, and the steps taken to handle and recover from it. This should be coordinated with the legal or data protection team, not decided alone.
The five-second test before you press Enter
On a busy day, the following five questions are enough to catch most mistakes.
- Is there anything here that someone else could use to get into my accounts or systems?
- Is there a name, number, or combination of attributes that points to a person who does not know I am sending their data?
- Is the content covered by an NDA, a contract, or a confidentiality marking?
- Would it be a problem if this page were read by the provider's staff or appeared in search results?
- Does the AI really need to see this part to do its task?
If any of the first four questions is answered yes, or the fifth is answered no, stop and mask first.
The formula above is not an arithmetic calculation. It is a way of reading the decision. The zone gives the starting point. The route can raise or lower your level of caution. A permission or a prohibition from the authorities overrides everything else. Masking can lower the zone before you send.
Examples and simulation
All names, numbers, and data in this section are illustrations and fake data. Numbers that look like an NIK or a card are only examples to show how the method works.
Case study: Rina and the table of 40 employees
Back to Rina. Her table holds 40 employees and six columns, so there are 240 data cells. Her task is to summarize performance review patterns by division and draft a fair salary increase proposal. Note that for this task the AI does not need to know anyone's name or NIK.
| Column | Cells | Type of data | Starting zone | Needed for the task? | Treatment |
|---|---|---|---|---|---|
| Name | 40 | Direct identifier | Red | No | Replace with codes K01 to K40 |
| NIK | 40 | Identity number | Red | No | Delete |
| Date of birth | 40 | Indirect identifier | Yellow | Partly (age) | Change to an age range |
| Salary | 40 | Personal finances | Red | Yes, in relative terms | Change to bands A to D |
| Performance score | 40 | Job assessment | Yellow | Yes | Keep, after the identifiers are removed |
| Disciplinary notes | 40 | Sensitive records | Red | No | Delete, or summarize as a category code |
| Total | 240 | ||||
Before masking, 160 of the 240 cells (the name, NIK, date of birth, and salary columns) were identifiers or personal financial data. After masking, direct identifiers are zero, but combinations of attributes still need to be checked. Rina counts the group size for each division.
| Division | Number of people | Safe from re-identification? (minimum 5) |
|---|---|---|
| Warehouse | 12 | Yes |
| Sales | 10 | Yes |
| Finance | 8 | Yes |
| Logistics | 7 | Yes |
| Management | 3 | No, merged |
| Total | 40 |
The management division has only three members. If it stayed separate and carried salary band D, anyone in the company could guess who they are. Rina merges management with finance into one group of 11 people. She keeps the table that maps codes to names in a separate file on her computer, and it is not sent.
The prompt she finally sends reads roughly like this.
The result is still useful because the task is about patterns and rules. Rina checks the output by hand against the original data before using it, because AI can be wrong and she remains responsible for the final decision.
One thing needs to be recorded. Whether sending employee data to an AI service, even in masked form, requires a particular processing basis under Article 20 of the PDP Law depends on the facts and on company policy. For that reason Rina still reports it to her manager and the legal team, and records the route she used.
Second simulation: a family member's lab results
A reader wants to understand the terms on her father's lab results. Health data is specific personal data, and the AI does not need to know who her father is. The difference before and after masking is shown below.
| Before (do not send) | After (allowed on a reasonable route) |
|---|---|
| Name: Hadi Santoso (example). Date of birth 14-03-1958. NIK 3201011403580001. Hospital and medical record number listed. HbA1c 6.8%, fasting glucose 132 mg/dL, LDL 160 mg/dL. | Male, in his 60s. HbA1c 6.8%, fasting glucose 132 mg/dL, LDL 160 mg/dL. Please explain what the terms mean and what questions I should ask the doctor. |
The lab values stay because they are what needs explaining. What is removed is everything that links those numbers to one person. AI is not a replacement for a doctor, and the result should be taken to a health professional.
Third simulation: an API key sent by mistake
A developer pastes a configuration file into a chatbot to find an error, then realizes the file contains a production API key. Deleting the conversation is not enough, because the key has already left their device. The sequence of actions is as follows.
| Order | Action | Reason |
|---|---|---|
| 1 | Revoke or rotate the key at the original service provider | This is the only step that closes the risk. Deleting the chat does not undo the exposure. |
| 2 | Check the key's usage logs from the time it was sent | Confirms there was no unrecognized use |
| 3 | Delete the conversation and any sharing links | Reduces the remaining traces, and does not replace step 1 |
| 4 | Tell the security team or your manager | The team needs to assess the impact and any reporting obligations |
| 5 | Move the key to an environment variable or a secrets manager, and use placeholders in shared code | So that a similar incident does not happen again |
The lesson is simple. For credentials, prevention costs far less than recovery, and recovery always starts with revoking the key, not with deleting the conversation.
Prompt examples: from risky to safer
Most daily tasks do not need the real data. The table below shows four common tasks. The names, numbers, and addresses in the left column are fake data.
| Task | Risky prompt | Safer version |
|---|---|---|
| Replying to a customer complaint | “Reply to this email. Dear Ms. Dewi Example, order #A-20931 to 12 Melati Street has not arrived and I already transferred Rp2,400,000...” | “Write a polite reply to a customer whose order is three days late. Offer a reshipment or a refund. The tone should be friendly and firm.” |
| Finding an error in code | “Here is my configuration file, why is there an error?” together with the database connection string and its password | “This connection function throws a timeout. I have included the code snippet with the value replaced by DB_URL_DISINI. What are the possible causes?” |
| Writing meeting minutes | Pasting the full transcript of a board meeting, complete with profit figures and participant names | “Create a minutes format with sections for decisions, follow-ups, owners, and deadlines. I will write the content myself.” |
| Improving a CV | Pasting a full CV with home address, phone number, and date of birth | “Improve the following experience descriptions to make them more concise. I have removed my name, contact details, and date of birth.” |
The same pattern repeats. The safer version describes the situation and the desired result without carrying data that is not needed. The quality of an AI's answer rarely depends on the customer's name or the order number. It depends heavily on how clear the task, tone, and limits are. If the AI really needs a data structure, send a made-up example with the same shape, then run the result yourself against the real data on your own computer.
What does this have to do with the law?
The PDP Law prohibits anyone from unlawfully obtaining or collecting another person's personal data for their own or another's benefit in a way that may harm the data subject (Article 65 paragraph 1), unlawfully disclosing another person's personal data (Article 65 paragraph 2), and unlawfully using another person's personal data (Article 65 paragraph 3). The criminal penalties in Article 67 are imprisonment of up to 5 years and/or a fine of up to Rp5 billion for obtaining and using, and imprisonment of up to 4 years and/or a fine of up to Rp4 billion for disclosing.
Whether sending data to an AI service counts as one of these prohibited acts depends heavily on the facts, the purpose, and the processing basis, and I do not draw a conclusion for you. What is certain is that the sentence “I only asked the AI to summarize it” does not answer the question about the processing basis. As of reporting in February 2026, the personal data protection authority mandated by the PDP Law is still being formed, and for now supervision is carried out by the Ministry of Communication and Digital Affairs (Komdigi). This status may have changed, so check for the latest news.
Interactive tools
The two tools below run entirely in your browser. The script does not send your text anywhere, and you can prove it by disconnecting from the internet before you use them. Even so, test the tools with sample data first. Both tools come filled with fake examples so you can see how they work right away.
Sensitive data scanner
Paste the text you plan to send to an AI tool. The scanner looks for patterns that are easy to recognize: national identity numbers (NIK), card numbers (checked with the Luhn algorithm), taxpayer numbers (NPWP) in dotted format, emails, Indonesian phone numbers, API keys and tokens, phrases like “password:” or “PIN:”, labeled account numbers, labeled dates of birth, salary amounts, IP addresses, and confidentiality markers. Names of people, addresses, and context are not detected, so read the result yourself again.
Decision checker: send or not
Answer six questions. This tool applies the zone and route matrix from the framework section. It then shows the decision if you send the data as it is and if you mask it first.
Application worksheet
This worksheet helps you write your personal rules before the next temptation arrives. Fill it in from top to bottom, then build a summary you can copy into your notes or share with your team. The fields are left empty on purpose so you write from your own situation.
1. Context
2. The data I handle
Enter the real types of data that often cross your desk. The zones and the ways to mask follow the framework section.
3. My personal rules
Write each rule as a short command. Example: “Never paste passwords, keys, or OTPs.”
4. Account settings check
5. Plan if something is already sent
Safe prompt library
A good prompt helps protect your data. The method is simple. You mask the data first. Then you tell the AI that the data is masked, what the task is, and what it must not ask for. This way the AI understands your intent, does not ask for the real data, and answers with the same placeholders, so you fill in the real data yourself on your own computer.
| Part | Purpose | Example sentence |
|---|---|---|
| Context | Tells the AI your situation and role | I work in customer service for an online store. |
| Masking statement | Explains that the data has been replaced with placeholders | I replaced all names and numbers with placeholders in square brackets. |
| Task | States the result you want | Write a polite and firm email reply. |
| Constraints | Stops the AI from asking for or inventing real data | Do not ask for real data. Use the same placeholders in your answer. |
| Format | Sets the shape of the output | Three paragraphs, with no promises I cannot keep. |
Ready-to-copy prompt templates
Choose a task, read when the template applies and what data you must remove first, then copy the prompt. Replace each placeholder in square brackets with data you have already masked.
A prompt does not change what the provider stores or reads. The sentence “do not ask for real data” only steers the AI. It is not a security barrier. Real protection comes from the data you masked before sending, so run the scanner above first and reread the text.
Risks and limitations
The three-zone method and the tools on this page reduce carelessness, but they do not remove risk. You should know some of their assumptions and limits.
What these tools and this method cannot guarantee
- The scanner only recognizes patterns. Names of people, home addresses, names of secret projects, and sentences that reveal identity through context are not detected. A result of “no findings” means no pattern was recognized. It does not mean the text is safe.
- Masking can fail. A combination of remaining attributes can point to one person. The rule of at least five people per group is only a heuristic.
- Provider policies change. Training settings, retention periods, and sharing features are updated from time to time. The quotes in this article reflect the sources at the time of writing and may have changed by the time you read them.
- Settings are not a guarantee. Turning off training does not remove storage, access by staff or partners, or the risk of a breach on the provider’s side.
- The matrix is general guidance. Sector rules, client contracts, and institutional policies can be stricter, and those are the ones that apply to you.
Other risks beyond typing in data
The discussion above focuses on what you put in. Other risks work from the opposite direction. Model output can contain sensitive information, and OWASP lists sensitive information disclosure among the top risks for applications built on large language models, including leaks of personal data, business data, and credentials through responses. Memory features and connectors to email, cloud storage, or calendars widen what the AI can see beyond the text you type. AI agents that act on your behalf add another risk, including hidden instructions in pages or documents the agent reads. Treat every access permission you grant like giving access to a new person.
When this method is not enough
Organizations that process data at large scale, or work in regulated sectors such as health, finance, and law, should write their AI use policy together with their legal, information security, and data protection teams. That policy should include a risk assessment, agreements with providers, and staff training. The method in this article suits individual habits and small teams. It does not replace a compliance program. An emergency involving a real leak also needs incident handling by the proper authorities, not a worksheet.
The article and tools on this page offer a way of thinking and general information. They are not legal advice and not a guarantee of compliance with the PDP Law or any other regulation, and the author is not your legal or security adviser. The final decision is yours, taking into account the rules that apply to your organization and the services you use.
Frequently asked questions
If I turn off model training, is it safe to send my data?
Not necessarily. That setting only answers whether your conversations are used to train models. The data still travels to the provider’s servers and can still be stored. Under the Anthropic consumer terms this article cites, for example, the 30-day retention period still applies to users who do not allow training. Use the setting as an extra layer, not as permission to send red-zone data.
Does deleting a conversation make the data disappear?
Not completely. Deleting a conversation does not delete the share links you created. For ChatGPT, you delete share links through the Data Controls settings and the Shared Links menu, and search results that are already indexed can remain for some time. Providers may also keep data for a set period under their terms. Changing the training setting does not pull back data that was already used.
What about temporary chat or incognito mode?
Modes like these usually reduce the trace in your account history. The text is still sent to the provider for processing. Read the terms for that mode on the service you use, and do not treat it as a reason to send red-zone data.
Can I upload a contract or a company report?
It depends on the contract, the workplace rules, and the route you use. Documents under an NDA or containing customer data should not be uploaded to a personal account. On a business account with a contract, still make sure you have written permission from the authorized party, and send only the parts you need where possible.
I already entered a password or an API key. What should I do?
Change or revoke that credential now, then check its usage logs. After that, delete the conversation and share links, tell the security team, and move the secrets to a safe place. Changing the credential is the only step that truly closes the risk.
Is a local model automatically safe?
Data is not sent to a third party, so the risk drops a lot. However, the security of the device, the plugins, and access to that machine is your responsibility. A local model that many people can reach without controls is no safer than a well-managed business account.
What about memory features and connectors to email or drive?
Both widen the data the AI can see, including data you did not type that day. Turn them on only when needed, limit permissions to specific folders or labels, and review the stored data regularly. On a work account, follow your organization’s policy.
How do I mask data properly?
Remove columns you do not need, replace names with codes, turn exact values into ranges, merge small groups, then reread the whole text. Keep the mapping table in a separate place. If the AI only needs the shape of the data, replace all the content with fake data of the same structure.
Does sending customer or employee data to AI violate the PDP Law?
There is no single answer. It depends on the basis for processing (Article 20 of the PDP Law lists six bases, including valid consent and legal obligation), the purpose, the agreement with the provider, and the type of data. Discuss your case with your organization’s legal or data protection team. For high-risk decisions, ask a legal adviser.
What about photos of ID cards, selfies, or screenshots?
Treat them the same as text. A photo of an ID card contains the NIK, a face photo, and other identity data. Selfies and face photos count as biometric data when used for recognition. Screenshots often carry names, notifications, and data at the edge of the image that you did not realize were there. Crop or cover the parts you do not need before uploading.
What should an AI use policy for a small team include?
One page is enough. It should list the red-zone data that is prohibited, the tools and account types that are allowed, the masking rules, the required account settings, and the steps to take if something is already sent. Use the worksheet section as a first draft, then get approval from leadership and the legal team.
Sources and references
You can check the references below yourself. The examples and figures in the article are illustrative and made up by the author. Provider terms are quoted from official pages at the time of writing and may change, so check the current page before relying on them for a decision.
- Law No. 27 of 2022 on Personal Data Protection, including Article 4 (types of personal data), Article 20 (bases for processing), Article 46 (protection failures), and Articles 65 and 67 (prohibitions and criminal penalties). pasal.id
- OWASP GenAI Security Project. LLM02:2025 Sensitive Information Disclosure, on leaks of personal data, business data, and credentials in large language model applications and how to mitigate them. genai.owasp.org
- UK National Cyber Security Centre (NCSC). ChatGPT and large language models: what’s the risk?, 2023, on stored queries that providers can read and the advice not to enter sensitive information. ncsc.gov.uk
- Bitdefender. Your shared ChatGPT chats may be publicly searchable, 1 August 2025 (updated 22 April 2026), on shared conversations indexed by search engines and how to delete the links. bitdefender.com
- Fortune. Report of 2 May 2023 on Samsung’s ban on employee use of generative AI tools, after source code and meeting recordings were entered into ChatGPT. fortune.com
- OpenAI Help Center. How your data is used to improve model performance, on the use of content for training, opting out, and the differences between individual and business services. help.openai.com
- Anthropic. Updates to our consumer terms, 2025, on model training choices, data retention periods, and which plans are affected. anthropic.com
- CNBC Indonesia. Telat 2 Tahun, Komdigi Janji Badan Penjaga Data Warga RI Berdiri 2026 (Two Years Late, Komdigi Promises the Data Protection Agency Will Be Set Up in 2026), 4 February 2026, on the status of setting up the personal data protection agency. cnbcindonesia.com
