PII redaction finds personally identifiable information (PII), such as names, emails and card numbers, and masks it before it reaches a model or leaves in a response.
How it works
Detection comes first, and two methods are usually combined:
- Patterns for data with a fixed shape: emails, phone numbers, card numbers (confirmed with a checksum), national ID numbers.
- Models for data without one: names, addresses, health details. A model reads the context, so it can tell Jordan the person from Jordan the country.
Each span found is then replaced, usually with a typed placeholder: Email me at [email protected] becomes Email me at [EMAIL], and the sentence stays usable for the model.
Detect, redact or block
Once PII is found, the action is a separate choice:
- Detect. Let the text through and log or flag the finding. Use it to measure what flows through the app.
- Redact. Mask the spans and pass the rest. Use it when the task does not need the values.
- Block. Stop the whole message. Use it when the data must never enter or leave.
Where to put it
On the input, before the model call, so personal data does not reach a third-party model. On the output, so the model does not repeat personal data from its context (data leakage).
With White Circle
Control detects PII before the model call, and redact is one of the actions a policy can take.
