The most common mistake people make with transcription privacy settings is assuming there is one switch. There are usually three, they control different things, and the vendor's documentation is the only reliable place to learn which is which.
The clause people miss
“When you rate a dictation or submit feedback through a form or by clicking thumbs up or thumbs down, we collect your feedback and also data from the related session. Session data may include audio recordings, transcripts, your edits, the app where you dictated, and technical information such as device, app version, and performance data. We use this information to diagnose issues, improve Wispr Flow, and develop and train our AI models.”
That is Wispr Flow's Data Controls page. It is a clean, honest paragraph — and it is the single most important one on the page, because it describes a path by which audio recordings reach model training even though the headline training setting is about text.
Clicking a thumbs-down is an ordinary thing to do. It is also, per that paragraph, an act that attaches a recording of the session to a feedback record the company may use to develop and train models.
Training and storage are not the same setting
“Whether Customer Content is synchronized across devices or retained to provide Service functionality is controlled by the applicable Service features, not by your model training preference.”
Wispr says this outright, and it is the sentence to keep. Turning off model improvement does not turn off cloud storage. Turning off cloud storage does not turn off the feedback path. The controls are orthogonal, and only one of them is about training.
The model-improvement setting
“If you opt to share your content with us for model training, we may also collect pseudonymized text, meeting transcripts, meeting outputs, and corrections you provide to improve the performance of Wispr Flow for all users.”
Two things are worth noticing. First, the word is pseudonymized, not anonymized — the two are not synonyms, and only one of them means the data cannot be traced back. Second, the setting is described as something you opt into. Wispr's documentation does not state, on the pages cited here, what the default is for a new account, and a default is exactly the kind of thing a buyer should be able to read in one line rather than infer.
Enterprise is the exception, and it is unambiguous:
“Enterprise accounts have data sharing turned off, and there is no option to enable it.”
So the strongest training guarantee in the product is a contract tier. Everyone else is trusting a switch whose default the documentation does not pin down.
The three switches, separated
| Switch | What it controls | What it does not control |
|---|---|---|
| Model improvement | Whether your content may be used to train or improve models | Storage, sync, or the feedback path |
| Dictation Cloud Storage | Whether transcripts, audio and history sit on Wispr servers | Training, or the subprocessor chain |
| Feedback (thumbs, forms) | Nothing — it is an action, not a setting | Attaches session audio and edits to a record |
| Enterprise contract | Data sharing off, no user-level override | Requires a sales process and a budget |
What the alternative looks like
Jot's App Store listing describes a narrower design: transcription on the device, no text saved to a server, and an optional corrective pass that sends de-identified text to US providers under zero retention agreements. There is no corpus of your dictation to train on, because the dictation is not sent anywhere in the first place.
That is not a stronger promise. It is a smaller surface. A product that never receives your audio cannot train on it, cannot lose it, and cannot change its mind about it.
How to check your own settings
- Open Wispr Flow, then Settings, then Data and Privacy. Read the model-improvement switch and note its current position before you change anything.
- Read the Dictation Cloud Storage switch separately. Decide whether you want a server-side history at all.
- Decide whether you are willing to send feedback. If a session contains something confidential, a thumbs-down is a data-sharing action, not a bug report.
- If you are on a team, ask whether the workspace is on Enterprise, because that is the only tier where the answer is not a toggle.
Verdict
Yes, Wispr Flow can train on your audio — through the feedback path, per its own Data Controls page, regardless of how you feel about the model-improvement toggle.
The settings are honest and separate. Wispr documents all of this, which is more than most competitors do. The problem is not disclosure; it is that a cloud product needs three switches to approximate what local processing gets by default.
If you do not want a training corpus to be possible, the answer is not a better switch. It is Jot on the App Store.
Sources
Primary sources, read on the date above. All Wispr Flow quotations are verbatim from its Data Controls page.