There is a version of the Wispr Flow privacy argument that runs on impressions: the app feels private, the settings page has switches, and the company says it does not sell data. Then there is the version that runs on the vendor's own documentation. The second version is one sentence long.
“Transcription always occurs on the cloud. This is the best way for us to provide accurate, low latency transcription.”
(The product was called Whisper Flow until the rename; whisperflow.com redirects to wisprflow.ai, and the documentation below is the same vendor's.)
That is Wispr Flow's Data Controls page, not a critic's summary of it. The company is being unusually direct, and it deserves credit for saying it plainly instead of hiding behind the word "secure". It also settles a question that a lot of people get wrong about this category: the model that turns your voice into text does not run on the device in your hand.
What "always occurs on the cloud" means in practice
Press the dictation key and the following happens, every time, on every platform:
- Your microphone buffer is captured and encoded.
- The audio is uploaded over the network to infrastructure Wispr operates or contracts.
- A speech model runs there, and the text comes back.
- The text is inserted into the window you were typing in.
None of that is a scandal. It is how most dictation products work, and it is why they are fast on hardware that could never run the model itself. But it has four consequences that no setting can undo, because they are properties of the architecture rather than of the policy:
- No offline dictation. No connection, no transcript. Airplane mode is not a privacy feature you can enable — it is a product failure state.
- A copy of your voice exists elsewhere. Briefly is not never. The audio is on someone else's machine for the duration of the request, and what happens after that is governed by a retention setting and a contract, not by physics.
- The subprocessor chain is part of your threat model. Wispr says its third-party AI providers operate under zero data retention agreements, with an explicit carve-out for some features. See the privacy comparison for the full chain.
- The screen you are typing in is sometimes in scope too. With Context Awareness enabled, Wispr may collect text from the active window to improve accuracy.
The three claims, and which one cloud transcription can make
| Claim | What it actually promises | Wispr Flow | Jot |
|---|---|---|---|
| Transport | Audio is encrypted on the way to a server | Yes | Not applicable — no upload |
| Rest | The server deletes it on a schedule you configure | Yes, configurable | Not applicable — no server copy |
| Absence | The audio never leaves the machine | No | Yes |
Only the third survives a determined adversary, a subpoena, or a change of ownership. The first two are policy commitments, and policy is exactly as strong as the vendor's controls on the day you need them.
What on-device actually looks like
Jot runs its speech model on the phone. The audio is decoded locally, the text appears locally, and the recording is never uploaded — there is no server-side copy to retain, delete, or leak. The App Store listing states the architecture in the same blunt register Wispr uses: “No text is saved, transcription is done on device, corrective AI calls are de-identified, use American ZDR providers, and optional.”
That last word matters. Jot does make network calls, but only for an optional corrective pass, and only with de-identified text — never audio — to providers under zero data retention agreements. Turn the feature off and the app is a closed system. That is a different bargain from "we upload your voice but promise to delete it", and it is the bargain most people think they are getting when they read the word "private".
Where the cloud genuinely wins
Accuracy on hard audio. Overlapping speakers, heavy accents, a bad microphone in a noisy room — a frontier model on a server GPU beats anything that fits in a phone's memory budget, and it does so without heating the device or draining the battery. Wispr's speed is real, and it is bought with the upload.
If your recordings are ordinary and your threat model does not include the vendor, that trade is fine. If the recording is a therapy session, a medical intake, a deposition, or a board meeting, the trade is the whole question.
Verdict
Wispr Flow is cloud transcription with a good privacy policy. The policy is a promise about a server that holds your audio; the architecture is what puts it there. You can improve the promise in settings. You cannot change the architecture in settings.
If you need dictation to work without a connection — on a plane, in a basement, in a building that blocks outbound audio — cloud transcription is disqualified before privacy enters the conversation.
If you want the audio to stay on the phone, that is a different product, and Jot on the App Store is the one this site is about.
Sources
Primary sources, read on the date above. The Wispr Flow quotation is verbatim from its Data Controls page; the Jot quotation is verbatim from its App Store listing.