Strip a user-defined list of sensitive names and terms from transcripts before any information is sent to cloud AI.

Today, audio and speech recognition stay on your Mac, but transcript text still goes to cloud AI for speaker attribution, summaries, and real-time meeting assistance. For teams that discuss specific clients, vendors, projects, or people every day, those recurring terms are the weakest link in the privacy chain.

This feature would let you maintain a list of strings — company names, client names, project codenames, internal identifiers — and have them stripped from the transcript before it leaves your device. Each match is replaced with a neutral placeholder like [redacted] so the cloud model produces summaries and answers using only the sanitized text. Sensitive terms never touch a third-party model.

Scope being considered:

• User-managed list in Settings
• Redaction applied post-ASR, pre-cloud — and optionally pre-database-write for users who want them stripped from local history too
• Works across Insert mode, Transform mode, and meeting transcripts
• Case-insensitive matching with an optional whole-word mode

Not in scope for this request: automatic detection of generic PII (emails, phone numbers, SSNs) via an ML model. That's a separate, larger effort — if demand is there, we'll track it as its own post.