Measured on 105 synthetic test cases
Pattern matching removed 57% of personal details. Adding a trained model removed 98.7%.
Keeping personal data out of AI prompts
Every prompt that reaches an AI provider is a disclosure.
Share of annotated personal details removed from 105 held-out synthetic test cases, measured on 15 September 2026. This does not measure whole-document anonymity. The guide explains the method and its limits.