Redact personal information
Detect and remove personal data from transcripts before analysis.
By default, Bristlenose sends transcript text to the AI provider unchanged. PII redaction detects and removes personal data from the transcripts before the analysis pass.
Turn it on
PII redaction is a command-line feature — the macOS app does not offer it yet.
Command line
Add --redact-pii to a run:
bristlenose run interviews --redact-pii
It is off by default, because over-redaction can remove useful research detail. The first time it runs, Bristlenose downloads a small language model used for detection. Redaction is performed by Microsoft Presidio.
What it catches
Names, email addresses, phone numbers, credit-card numbers, and national identifiers such as US Social Security and UK NHS numbers, among others.
What it does not catch
Detection is not complete. It can miss names that are also ordinary words ("Grace", "Will", "Hope"), nicknames, non-Western names, and dictated contact details. It deliberately leaves place names alone — they rarely identify a person, and removing them ("the Oxford Street store") would strip useful context.
Check the result
Every redaction is recorded in pii_summary.txt, in the hidden .bristlenose/ folder, with the original value and a confidence score. Review it. Redaction is an aid, not a guarantee, and should not be relied on alone.
Speaker identification runs before redaction. That step — often where people say their names — sends the raw transcript to the provider. With a cloud provider, that portion is sent unredacted; with a local model, nothing is sent at all. For the strongest assurance that transcript text never leaves the machine, run a local model.
For the difference between this and removing names from a shared report, see Privacy and data.