Concepts
Pseudonymization vs redaction
Both hide identity. They fail differently when the next step is an AI model that needs to reason about parties, obligations, and timelines.
Redaction
Permanently removes or obscures content — asterisks, black boxes burned into a PDF, or deleted spans. Best when the audience is a public docket, opposing counsel production, or any context where reversibility is a liability.
Cost for AI: the model often cannot tell who did what. Summaries become vague or invent structure to fill gaps.
Pseudonymization
Replaces identifiers with stable tokens ([PERSON_01], [ORG_02]) and keeps a mapping under your control. Best when you want ChatGPT/Claude to analyze a matter and you may later re-identify outputs under authorization.
Cost for compliance: the mapping is sensitive. SafeToAI keeps mapping.json as a local download — it is not stored on our servers on the default path.
Quick chooser
| Situation | Prefer |
|---|---|
| Public filing / production | Irreversible redaction |
| Private ChatGPT brief of a contract | Pseudonyms |
| Need to restore names into an AI draft | Pseudonyms + mapping |
| Maximum visual concealment in a PDF screenshot | Asterisk / black-box redaction |
SafeToAI supports both modes in-browser. Start with pseudonyms for AI workflows; switch to asterisks when you need stronger obliteration and do not care about model readability.
Try both modes