How to redact a document before using ChatGPT or Claude

1. Open a document

Launch Lex Redactus and drag a PDF or DOCX into the app. Everything — detection, redaction, and decoding — runs on your computer. The only network call the app makes is a periodic license check; no document content is transmitted.

2. Detect the PII

The app scans the document with on-device detection (a neural GLiNER model plus pattern libraries tuned for legal text) and highlights what it finds — people, organizations, addresses, SSNs, case numbers, and more.

3. Review every suggestion

This is the step that matters most. Detection is high-recall but not perfect, so you confirm, edit, or add each entity before anything is exported. Human review is the safety gate — see the review guide for how to add a missed name or remove a false positive.

4. Export a tokenized copy

Use the Download control to export an anonymized version where each value is replaced by a reversible token like [[PERSON:01]] or [[ADDRESS:01]]. You can even assign your own labels (e.g. [[PERSON:TD]]). This is the only thing that leaves your machine — and it goes to your clipboard, not a server.

5. Use the AI

Paste the tokenized text into ChatGPT, Claude, or any AI tool and ask for a summary, chronology, issue list, or draft outline. The AI sees only tokens — never a real client name.

6. Decode the answer

Paste the AI's response into the Decode Workspace. The tokens turn back into the real names, color-coded by entity type, using the entity map that stayed on your machine. Keep working — ask a follow-up and it re-encodes automatically.

Where it should not be used

• Do not paste original client identifiers into an AI tool — only the tokenized copy.
• Do not treat AI output as final legal work; it is lawyer-reviewed assistance, not a substitute for your judgment.
• Always re-read the exported copy before sharing it — you are the final check.