Do AI dictation apps send your words to a language model?

Some AI dictation apps send your words to a language model, and some do not. The difference is whether the app adds a language model step after speech recognition. A recognizer turns audio into words. A language model can then clean up, punctuate, reformat or rewrite those words, and that second step is the one that sends your text to a model provider. Many AI dictation apps do both, so you need to check each step separately.

Plain recognition sends audio to the recognizer

ChatGPT's dictation is a clear example of the recognizer step. When you press the microphone icon, the recorded audio is sent to OpenAI's models to be transcribed, and the returned text can be edited before you send it ¹. That is audio going to a model to become text. Whether a further language model step runs on the text is a separate question, and the product's policy has to answer it.

Many AI dictation apps add a model that edits your words

Wispr Flow's intelligence was described in a June 2025 report as based on a mix of Meta's Llama and OpenAI's Whisper models ². A model that edits the text works on it after recognition, so the words you dictated reach that model. The company told the same report that it may bring in other model providers as it expands its features ³, which means the list of providers can change.

Wispr Flow says it deletes your audio as soon as it has been transcribed, unless you have turned on its Dictation cloud storage setting ⁴. That statement is about the audio. It does not by itself tell you how the text is handled by a model, so read the section on processing in the policy and the list of providers, not only the deletion line.

Some apps keep the AI step off until you switch it on

Superwhisper shows that the same app can run in two ways. Its free plan dictates with local Whisper models without AI processing, and the company says nothing dictated leaves the device ⁵. Its Pro plan adds cloud models and AI processing ⁶. Its Voice to Text mode outputs raw transcribed text with no language model involved ⁷.

Superwhisper also describes bring-your-own-key use. When you supply your own API keys, it acts as a client of your account with the provider, and the provider's agreement, not Superwhisper's terms, governs how the data is handled ⁸. That is useful for an organization that already has an agreement with a provider, but it means you must read that provider's terms.

Table 1. Where text can go in common AI dictation setups

It depends on the mode, the plan and the provider

Three things decide the answer. The first is the mode: a raw mode skips the model. The second is the plan, because a free tier and a paid tier can run different steps. The third is the provider, which may change as the app grows. A short test is to dictate one sentence with Airplane Mode on. If the words appear, the recognizer runs on the device, but that does not tell you whether a model step is applied later, so you still need the policy.

What to do before you dictate work

1. Find out whether the app has a raw or "voice to text" mode, and select it when you do not want editing.
2. Run the airplane test from the guide linked below on the mode you plan to use.
3. Read the policy for the named model providers, not for the word "AI" alone.
4. Check how long audio and text are kept, using the checklist for a dictation app's privacy policy.
5. For confidential or client content, check whether the text can be used to train models, using the question on how to stop apps training on your voice.

Sending text to a model is not automatically unsafe. A model provider with a clear agreement and a short retention period can be a reasonable choice for ordinary notes. The risk comes from not knowing which provider holds the text, or whether it is kept to train later. The question on whether AI models can leak what you dictated explains why that matters.

If you use an AI dictation app, check the cloud dictation entry too, because it explains the difference between a server step and an on-device step. Run the same checks on Yap, which we make, before you rely on it for anything sensitive. The guide to whether voice to text is safe covers the wider picture.