PII redaction strips names, numbers, and IDs so the model never has to see them.
DefinitionWhat it means
PII redaction is the process of detecting personally identifiable information, such as names, addresses, or account numbers, in text and masking or removing it either before it is sent to a model or before the model's output is stored or displayed.
Why it mattersWhy you should care
Redaction is often a hard compliance requirement under regulations like GDPR and sector-specific privacy law, so AI products handling customer data typically run automated PII detection at both the input and output boundary rather than relying on the model to withhold sensitive data on its own.
At a glanceSee it
Reversible tokenization routes the real values into a private vault the model never sees, then rehydrates them into the reply so the user still gets the originals back.
One confidence threshold balances two opposite failures — redact too eagerly and you garble the text, redact too timidly and real PII leaks to the model.
Where you see itIn the wild
- Data loss prevention tools scrubbing logs before they reach a model provider.
- Healthcare and finance products masking records before generating a summary.
- Architecture reviews on where to place redaction in an AI data pipeline.