Module 1 · Start safe
Citizens' data and excepted information stay out of the model
In short
ID numbers, family card numbers, addresses, health, finances, children's data, confidential internal memos and unpublished procurement documents are not pasted into AI tools. Mask first, use an agency-approved tool and ask when unsure.
After this lesson you can
- Recognise citizens' data, specific personal data and excepted information that must not be pasted into an AI tool
- Mask agency working material before asking AI for help
- Check whether the AI tool in use is approved by the agency for that kind of data
Terms in this lesson
Understand
Government offices hold large amounts of citizens' data: national ID numbers, family card numbers, addresses, phone numbers, assistance recipient data, complaint histories, even health and financial records. That data was collected because citizens must or need to give it to get a service. That trust is at stake every time someone pastes a table of citizens' data into an AI tool's chat box.
Law Number 27 of 2022 on Personal Data Protection, usually called the PDP Law, applies to personal data controllers, including public bodies. It distinguishes general personal data from specific personal data. Article 4(2) lists specific personal data as health data and information, biometric data, genetic data, criminal records, children's data, personal financial data, and other data under the laws and regulations. Many files in a public service office hold several kinds at once: an assistance recipient list can carry financial condition, children's data and disability status in one row.
Besides personal data, there is excepted information. Law Number 14 of 2008 on Public Information Disclosure requires public bodies to open access to public information, except the kinds excepted in Article 17. These include letter h, information that would reveal personal secrets such as health history, financial condition and the results of an evaluation of someone's capability, and letter i, memoranda or letters between or within public bodies that are confidential by nature. Such information is not opened even to information requesters, so it is clearly not pasted into a general AI tool you do not control.
The third group people forget is procurement and policy documents not yet announced: estimated prices, draft specifications, lists of candidate suppliers, draft budgets. If they leak early, competition is no longer fair and the agency can be accused of favouritism.
This does not mean AI cannot be used. Most work can be done without sensitive data. To tidy a reply letter template, the model does not need the citizen's name; write ‹nama› and ‹nomor-registrasi›. To summarise complaint patterns, the model does not need ID numbers; categories, dates without times and complaint text with identities removed are enough. This is called pseudonymisation: identities are replaced by codes, and the key stays with you. Remember that pseudonymised data can still be traced back, so it is still handled with care; it is not anonymous data.
The second step is checking the tool. Many agencies have already set which AI tools may be used, for what data and with what settings. An approved tool usually has a data processing agreement, settings so conversations are not used to train the model, and access logs. Your personal account on a free app has none of these. If there is no policy yet, ask the information officer or data protection officer, and meanwhile paste nothing that could not be posted on a public notice board.
Before
The prompt
Here is a list of 40 citizens who complained about assistance not yet paid, with ID numbers, addresses and phone numbers. Please group the complaints and write a reply for each.
What came back
Bad example output, do not use: the model processed the whole table and wrote 40 replies naming each citizen's name, address and ID number. That data now sits in a third-party service the agency never approved.
After
The prompt
Goal: I need to group the reasons in complaints about unpaid assistance and prepare one reply template per group. Reader: my supervisor, then citizens through the templates. Material: complaint text I have masked; identities replaced with codes A01 to A40, no ID numbers, addresses or phones. Limits: do not guess identities; do not promise payment dates. Shape: a table of groups, complaint counts and complaint codes; then one reply template per group with ‹nama›. Check: name complaints that fit no group.
What came back
Draft for review. Group 1, account not yet verified, 17 complaints. Group 2, name not on this stage's list, 12 complaints. Group 3, does not know the schedule, 9 complaints. No fit: A07 and A31, please read manually. Templates per group follow, with ‹nama› and no payment dates.
What changed
- The first request sent 40 citizens' ID numbers, addresses and phones to an unapproved tool. That is already a breach, whatever the output.
- The second request uses only masked complaint text; the grouping works just as well.
- Replies to citizens are made from templates, then names are filled in inside the agency's system, not in the AI tool.
Common mistake
- Mistake
- Pasting the owner's estimate and draft procurement specifications into a general AI tool to tidy them, before the documents are announced.
- Why it hurts
- Candidate suppliers must not know those documents yet. If the tool stores or uses conversations, you cannot pull the information back, and the procurement can be challenged.
- Do this instead
- Tidy structure and wording with an empty template or a fictional sample, then fill in the figures yourself in the agency's system. If you need AI help with sensitive documents, use only a tool your agency has approved for that.
Try
Sorting card
May it go into an agency-approved AI tool, or must it be masked or held back first?
A recipient table with ID and family card numbers
A service announcement already published on the agency website
An internal memo marked confidential
Approved minutes for participants, with no personal data
The draft owner's estimate for next month's tender
A citizen's complaint about their sick child, with the school's name
6 cards left
Your office has no AI tool policy yet. A colleague suggests using their free personal account to summarise unmasked citizen complaints. What is the best step?
Take home
Before pasting work material into an AI tool
- This tool is approved by the agency for this kind of data
- No ID numbers, family card numbers, addresses, phones or unnecessary citizen names
- No health, financial, children's or criminal record data
- No memos or letters that are confidential by nature
- No procurement or budget documents not yet announced
- I keep the pseudonym key myself, outside the AI tool
What went wrong for us
Our lead-finding tool is designed to look for companies only
- Context
- We are designing a tool that helps find prospective customers for a product.
- What tends to go wrong
- The design limits the tool to finding companies and organisations, never private individuals. People enter the customer list only when they agree, or when the user imports contacts they already have a clear basis for.
- What it costs if ignored
- Private individuals' personal data would not be collected in the first place, so there would be nothing to guard, delete or explain later.
- Root cause
- Searching for private individuals would mean collecting personal data with no relationship to them, which the tool's purpose does not need.
- How we do it
- The limit is written into the product design as a rule of the tool, not a setting users can change.
- The rule we keep
- Features planned for that tool may not add searches for private individuals.
The lessonData that is never collected cannot leak.
What it means for your workBefore pasting material into AI, ask whether the task really needs data about people; often it does not.
Sources
- UU No. 27 Tahun 2022 tentang Pelindungan Data Pribadi (JDIH BPK) (accessed 26 October 2026)
- UU No. 14 Tahun 2008 tentang Keterbukaan Informasi Publik (JDIH BPK) (accessed 26 October 2026)