# What is PII Redaction?

Also called Data Masking, Anonymization, Pseudonymization.

PII redaction is the removal or replacement of personal data in text before it is stored, logged, or sent onward. Methods include masking with placeholders, replacing values with reversible tokens, generalizing a value to a coarser range, and substituting realistic synthetic values. The method chosen determines whether the original can ever be recovered.

Reversibility is the first decision. Tokenization keeps a mapping so the original can be restored, which preserves workflows where a reply must reach a real person, and which means the data remains personal data with the vault as a new asset to protect. Irreversible masking discards the original entirely, which is safer and forecloses any later need for it.

Under European data protection law the distinction matters legally. Pseudonymized data, where re-identification remains possible using additional information held separately, is still personal data and stays in scope. Anonymized data, where re-identification is not reasonably possible by any party, falls outside the regime. Meeting the anonymization standard is harder than most implementations assume.

Utility loss is the constant trade-off. Masking every entity to an identical placeholder can strip the coreference a summarizer needs to keep two people apart. Consistent per-entity tokens preserve structure at the cost of a stable pseudonym across the document. Synthetic substitution keeps text natural for downstream models but can mislead a human reader who does not know values were replaced.

Placement matters as much as method. Redaction applied in the application layer but not in logging middleware leaves the raw payload in log storage, which is a common and easily missed gap. Applying it once at the boundary, before any component can persist the text, is more reliable than asking every component to remember.

## Key points

- Reversible tokenization keeps a vault that must itself be protected
- Pseudonymized data is still personal data under GDPR
- True anonymization is a higher bar than most systems meet
- Over-masking destroys the structure downstream tasks need
- Redact once at the boundary, before anything can persist text

## In practice

A conversation pipeline replaces each detected person with a stable token such as PERSON_1, keeps a mapping in a separate encrypted store, and passes the masked text to the model. Summaries stay coherent because two speakers remain distinguishable. When a reply must be sent, the mapping is applied in reverse at the delivery step only. The mapping store carries the same access controls as the original data.

## Related terms

- [PII Detection](/en/glossary/pii-detection)
- [Data Leakage Prevention](/en/glossary/data-leakage-prevention)
- [GDPR](/en/glossary/gdpr)
- [Audit Trail](/en/glossary/audit-trail)
- [Data Residency](/en/glossary/data-residency)

[Back to the AI Glossary](/en/glossary)