PII (Personally Identifiable Information)

When you will meet it

You will hit it — usually after it has already entered the context — whenever you feed data to a model, build a RAG corpus, or let an agent read support tickets. PII needs special handling not because the model might remember it, but because the act of sending it out can itself be unlawful: most jurisdictions regulate processing, cross-border transfer and secondary use of personal data, and data inside a third-party service cannot be recalled.

Minimal example

原始工單（示意，人物為虛構）：
  「王小明（0912-345-678）反映其住處的網路
    自 3 月 5 日起中斷，身分證字號 A123456789」

兩種處理：
  遮罩（redaction）：
    「[姓名]（[電話]）反映[地址]的網路自 3 月 5 日起中斷，[ID]」
    → 直接拿掉，需要時回原系統查
  假名化（pseudonymisation）：
    「用戶 U-8842 反映其住處網路自 3 月 5 日起中斷」
    → 換成代號，對照表另行保管；分析仍可跨工單追蹤同一人

The difference is reversibility: redaction removes; pseudonymisation substitutes, and the mapping table still exists somewhere. Pseudonymisation keeps the analytical value of linking records — and keeps the re-identification risk — so the mapping needs protection equal to the original data.

What people get wrong

• Assuming only names and ID numbers count. Anything that pins down a person in combination counts: company plus title plus city — each harmless alone — often identifies exactly one human.
• Treating "I skimmed it before pasting" as a process. Human eyes do not catch PII across batches; automate detection in the pipeline, and assume some will still slip through.