Is Wispr Flow safe? What it sends, stores and lets you turn off

What Wispr Flow documents well

It would be wrong to call it careless. The Trust Center provides a SOC 2 Type II report, there is an ISO/IEC 27001:2022 certificate covering Flow dictation, service data is encrypted in transit and at rest, and the page says third-party AI providers may not use your data to train their models. HIPAA-ready dictation needs a signed business associate agreement. These are real controls and a security team will want to know about them.[2]

What leaves your Mac

Everything you dictate. The security FAQ says Flow processes dictated audio in the cloud to turn speech into text, and processes and stores customer data in the United States. It also says the service is not end-to-end encrypted in the strict sense, because it has to process the audio to produce a transcript, and that Wispr manages the encryption keys.[2]

The page does not name the cloud or language-model providers behind this. It says Wispr uses third-party AI providers and keeps a subprocessor list in its Trust Center.[2]

What you can switch off, and what you cannot

There are two settings. 'Improve the model for everyone' is on by default for Free and Pro and can be turned off; it withholds audio, transcripts and edits from model improvement. The documentation is explicit that turning it off does not stop cloud processing, and that metadata such as word count, duration, app and timestamp is still uploaded. A dictation you choose to report is retained and can be used to train and evaluate models regardless.[2][4]

'Dictation cloud storage' keeps transcripts, audio and history on Wispr's servers for cross-device access. Per the privacy article both settings default to on, and signing out resets the storage preference to on. The meeting Notetaker is separate: its audio is kept for 7 days, its transcripts and summaries are stored in the cloud, and there is no opt-out.[4][2]

Where safe is the wrong question

For most dictation this does not matter. Automatic marking of passages where recognition was unsure is not publicly documented for Wispr Flow either, which is a separate question from privacy. It starts to matter when the sentence you just spoke is a client name, a settlement figure, a diagnosis or an unpublished plan, and when someone else's rules decide where it may be processed. Then the useful question is not whether the provider is audited, but whether the content needs to leave your Mac at all. Ask your own policy, not ours.[2]

Witness answers that question by not having the option. Speech recognition and text processing run on your Mac, and so does the licence check. Audio, transcript and vocabulary are never sent anywhere, and audio stays in memory only. What goes out is limited to documented non-content data for activation, licence, payment and updates. The text lands at your cursor at once, a menu-bar marker flags passages that deserve a second look, and one click opens them with the audio behind them and the raw transcript.[1]