Safety
PII
PII is data that identifies a person — names, emails, phone numbers, addresses, identifiers — and it should not reach an API you do not control without a decision behind it.
Also written: personally identifiable information
Provider promises not to train on API data are worth something and are not the whole picture. Retention for abuse monitoring, your own application logs, your error tracker and your contractual obligations to customers are all separate concerns.
Pattern matching reliably catches structured identifiers: emails, cards, phone numbers, keys. It cannot catch a name in a sentence or a medical detail described in prose.
In practice
Most tasks — classification, summarisation, extraction — work identically on pseudonymised text, so names are rarely load-bearing. Redacting before sending is the cheapest item on any privacy review because unsent data needs no protection. Note that redacted free text is pseudonymised, not anonymous: role plus location plus date re-identifies.
Common questions
Do I need to redact data before sending it to a model?
Send the minimum the task needs, which for classification, summarisation and extraction rarely includes names. Unsent data needs no protection, which makes this the cheapest item on any privacy review.
Does redaction make data anonymous?
Almost never. Free text re-identifies through combinations — a role, a location, a date, an unusual detail. Treat redacted text as pseudonymised, which is still personal data.