PII Redaction
Finding and masking personally identifiable information within a dataset.
What this means for trainers
When in doubt, redact — the cost of over-redacting a borderline case is far lower than the cost of leaking someone's real personal information into a training set.
PII redaction identifies spans of text containing information that could identify a specific individual, such as names, addresses, phone numbers, email addresses, and government ID numbers, and masks or removes them before data is used for training or shared with a broader team. It is a required step for any dataset sourced from real user interactions rather than synthetic or already-anonymized text.
This work is closely related to named entity recognition in technique, since both involve finding spans of a specific type, but the two differ in purpose. NER labels entities to preserve information, while PII redaction removes it. Getting the boundary right matters just as much either way: under-redaction leaves sensitive data exposed, while over-redaction can strip context a model needs to answer correctly.
PII redaction is one part of the broader practice of privacy-preserving annotation, which also covers access controls and minimizing what annotators see in the first place, so redaction is only one layer of a larger set of protections around sensitive data.
Related terms
Put this into practice
Browse open red-teaming, safety evaluation, and model-alignment roles from Alignerr, Mercor, Outlier, and more.
Browse AI training jobs