AI glossary · Basic · AI security · Data engineering
Personal data de-identification
Personal data de-identification means removing or replacing information that can identify a specific person (PII, personally identifiable information), such as names, phone numbers and resident registration numbers, so that no one can tell whose data it is. It is a basic protection step data goes through before being used for AI training or analysis.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Security & safety
- Fields
- AI security, Data engineering
- Added
- 2026-10-08
How it works
First, it finds where personal data is in the dataset. Information in fixed positions, like table columns, is easy to find, but in free-form text such as support transcripts and emails, rules (regular expressions) or AI models automatically detect items like names, addresses and account numbers. The detected information is then handled according to the purpose. Methods include masking, which hides "Hong Gil-dong" as "***"; pseudonymization, which replaces it with another name such as "Customer A"; generalization, which widens "34 years old" to "30s"; and deletion, which removes it entirely. Finally, it checks whether the remaining information could be combined to identify a person again (re-identification risk).
Why it matters
LLMs can memorize sentences from their training data verbatim and reproduce them in answers. If customer phone numbers are mixed into the training data, there is a risk that the model will give those numbers to the wrong person. Personal data can also leak out when internal documents are connected through RAG or when prompts are sent to external AI APIs. That is why personal data laws in many countries require appropriate de-identification before data is analyzed or used, and it is the first hurdle companies run into when adopting AI.
Things to know
Pseudonymization and anonymization are different. Pseudonymized data can be traced back to the original person if a separately stored "mapping table" exists, so it is often still treated as personal data; anonymization refers to a state that cannot be reversed by any means. Hiding too much makes the data useless, and hiding too little fails to protect it, so a balance between usefulness and safety is needed. Automatic detection tools also miss things, so for important data it is good to also have a process where people review samples.
From raw data to anonymous
- Masking (partially hidden)
- Pseudonymization (mapping table kept separately)
- Generalization · deletion
- Anonymization
Example
A company building a call summary AI from customer service transcripts would, before training, automatically replace customer names in the transcripts with "Customer A" and phone and card numbers with markers such as "[phone number]" and "[card number]." Another widely used approach is for an LLM gateway in between to find and mask resident registration numbers in prompts before sending them out when employees ask questions to an external chatbot.