Where AI puts personal data at risk

Why this matters

Most people think of an AI tool as something that answers a question and forgets. Many do not forget. What an officer types or uploads can be stored, read by staff at the provider, used to improve the system or passed to other companies that help run the service. As a privacy officer you are often the person asked whether a tool is safe. To answer, you first need to see where personal data can go.

What counts as personal data

Any information that identifies a living person, or could identify them when combined with other details. That includes names, CNIC numbers, phone numbers, addresses, photographs, voice recordings, signatures, bank details, health and family information and case details such as a complaint or an application. A single detail may look harmless. Several together often point to one person.

Eight places where it can escape

• Prompts. Text typed or pasted into a chatbot, often a whole letter or case note at a time.
• Uploaded files. Scans, spreadsheets and PDFs, which may hold names in places the officer did not notice, such as headers, comments and hidden columns.
• Chat history and memory. Some tools keep past conversations or a chat memory of things you told them, and may use them later.
• Provider logs and review. Providers may keep records of what was sent and, under some terms, let staff read samples.
• Training. Under some terms, what users type may be used to improve the model. Read the terms for the exact version your office uses: free and paid versions often differ.
• Add-ons. Browser extensions, meeting note-takers, transcription apps and plug-ins that send audio or text to other services.
• Features switched on inside existing software. An email, document or case system that quietly gains an AI feature after an update.
• Outputs. A tool can repeat personal data it was given earlier in the same account or workspace.

Who is at risk

Citizens first: their information leaves the control of the state and may be seen, kept or misused. Staff next: personnel records, performance notes and complaints against colleagues are also personal data. And the department: a leak damages public trust, which is slow to rebuild.

Read the terms, not the marketing

Do not rely on a tool's description of itself. Look for the provider's terms and privacy statement and find answers to four plain questions: Is what we type stored? For how long? Is it used for training? Where is it processed? If the answers are unclear, treat the tool as unsuitable for personal data until you have a clear written answer.

Try it

Pick one AI tool that staff in your office use, or might use. For each of the eight places above, write one line: could personal data reach it, yes, no or unknown? Count the unknowns. Those are your first questions to the provider.

Check yourself

1. Why is a name not the only thing that makes data personal?
2. Name three places, other than a typed prompt, where personal data can escape through an AI tool.
3. What should you do if the answer to "is it used for training?" is unclear?

Answers: (1) Combinations of details, such as area, job and date, can identify a person. (2) Any three of: uploaded files, chat history or memory, provider logs and review, training, add-ons, features inside existing software, outputs. (3) Treat the tool as unsuitable for personal data until you have a clear written answer.