PII Redaction
also called Anonymisation
Removing or masking personal identifiers before the model sees them.
Detecting and replacing names, addresses, account numbers and other personal data with placeholders, then restoring them after generation if the application needs the real values.
A clean pattern for regulated contexts: redact on the way in, keep the mapping on your side, rehydrate the model’s output locally. The provider never holds the identifiers.
The hard part is restoration, and it is worth designing deliberately. A placeholder scheme needs to survive the round trip without collisions and without leaking structure — sequential placeholders reveal the number of distinct people, which is itself information in a small dataset.
Redact, generate, restore
Drafting a customer letter that must contain their real details.
Replace identifiers with stable tokens, let the model write the letter, then substitute the real values back on your side. The model never sees the data and the output is still personalised, which is the pattern most privacy-sensitive deployments converge on.
- docsData controls in the OpenAI platform ↗
Training-use and retention controls, including default abuse-monitoring retention and Zero Data Retention eligibility. Vendor-specific, and no substitute for your own privacy review.
- standardAI Risk Management Framework: Generative AI Profile (NIST AI 600-1) ↗
The official record for the cross-sector GenAI companion to the AI RMF. Use it as the governance vocabulary; the PDF itself was not machine-readable in this research pass, so do not quote definitions from here.