Do Dictation Apps Train AI on Your Voice?

Some dictation apps use what you say to train their AI, and many do it by default. You can find out for any app in two places: one clause in its privacy policy, and one setting in the app or your account. Apple stores Dictation audio only if you opt in to Improve Siri and Dictation ¹. Microsoft says it uses Copilot data, including voice, for AI training except for certain categories of users and people who opt out ². Otter's policy lists training on de-identified audio ³. This guide shows where each vendor says this, what the settings are, and what de-identification does and does not protect. We checked each policy and help page on October 17, 2026. Vendors change these terms without much notice, so treat every date as a snapshot.

Training on your voice is a policy choice each vendor makes, and many make it opt-in

Training is not automatic in the technical sense. A vendor decides whether to keep recordings or transcripts, and whether to use them to improve its models. Those decisions show up as defaults. A 2025 Stanford study read the privacy policies of six US AI developers and found that all six use users' chat data to train their models by default ⁴. That study covered chatbots, not dictation. It matters for dictation when the text you dictate goes into a chat model.

Defaults differ by product, so compare them one by one:

The table covers different kinds of product. The OpenAI row describes the developer API, not the consumer apps. The Claude row describes the mobile dictation feature, which is separate from the chat terms covered below. Read the table as a starting point. The same company can apply different rules to consumer, business and API accounts.

Apple and Microsoft changed their voice programs in 2019 and 2020

The clearest change came in 2019. After reports that people were listening to Siri recordings, Apple said it immediately suspended human grading of Siri requests and began reviewing its practices ⁹. Apple also said that, by default, it would no longer keep audio recordings of Siri interactions, and that users could opt in ¹⁰. Apple's current page describes the same opt-in structure. Under Improve Siri and Dictation, Apple may store the audio of your interactions with Siri and Dictation ⁵. For people who opt in, only Apple employees, under confidentiality obligations, can access that audio ¹¹. The setting is what decides whether your audio is kept for review and improvement.

Dictation on a Mac is a separate case from Siri. Apple says that, on a Mac, voice inputs and transcripts for general text Dictation are processed on your device and are not sent to Siri servers ¹². That is a statement about where recognition happens. It does not say that nothing is kept for improvement, which is why the setting still matters.

Microsoft's record is similar. It stopped logging voice data for product improvement on October 30, 2020 ¹³. Microsoft also offers a way to contribute voice clips for improvement. If you do contribute, Microsoft says its automated software removes the link between a clip and your Microsoft account or device identifiers ¹⁴. The same policy says Microsoft may keep a sampled clip for longer than two years so that it can keep training its speech recognition models ¹⁵.

The pattern across Apple and Microsoft is that audio improvement is now something you can decline. The choice still sits in settings that you have to find yourself.

AI dictation and meeting apps vary most, so read the clause on model improvement

AI dictation and meeting tools are where the policies differ most. Otter.ai's privacy policy, effective June 16, 2026, lists "training our proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information)" as one of its purposes ³. Read that as Otter listing training on de-identified audio as one of its purposes, and the transcripts may still contain personal details. The policy also commits Otter not to re-identify its de-identified data, except to test whether its de-identification works ¹⁶. It treats aggregated data as personal information again if Otter combines it with your personal information in a way that can identify you ¹⁷.

Otter is also the subject of a lawsuit. A federal class action filed in August 2025 alleges that its Otter Notebook service, by default, does not ask meeting attendees for permission to record ¹⁸. These are allegations, not findings, and we did not check the case's current status. For the question of training specifically, the policy is the document to read, and we did not find an individual opt-out from training on the consumer pages we read. For the full answer on Otter, see whether Otter.ai uses your recordings to train its AI.

Chat-based AI tools show the same range. Stanford's researchers describe a change in which Anthropic's consumer terms began to use conversations with Claude for training by default, unless the user opts out ¹⁹. Anthropic's mobile dictation, by contrast, deletes the audio after it converts speech to text and does not keep a copy or use it for training ⁷. A single company can therefore treat your spoken audio and your typed chat differently, which is why the clause you need is specific to the feature you use.

When you read a policy, check four things, in this order:

Storage periods also matter, because a longer period gives more time for training and for a leak. The stated periods differ widely. Microsoft says it keeps voice clips that people contribute for up to two years ²⁰, which is 730 days ²¹. A sampled clip may be kept longer, so that Microsoft can keep training its speech recognition models ¹⁵. Microsoft's default storage period for Copilot conversation activity is 18 months ²², about 548 days ²³. OpenAI says its API abuse monitoring logs are kept for up to 30 days by default ²⁴.

The three periods cover different data for different purposes. The chart compares stated limits, not the risk of any one product.

De-identified data is still your words, and models can memorize rare phrases

De-identified data is not the same as anonymous data, and a trained model can keep some of what it saw. Under EU law, personal data that has been pseudonymised still counts as information about a person when it can be attributed to that person with additional information ²⁵. Otter's own clause makes a similar point: if it combines aggregated data with your personal information in a way that can identify you, it treats the combined data as personal information ¹⁷.

Removing names and account IDs does not remove the words. Some vendors say they de-identify personal information before training, and others do not say so. The 2025 Stanford study found that only some of the six developers state that they de-identify personal information before training ²⁶. A policy that is silent on de-identification does not prove that identifiers are kept, but it means the commitment is not written down.

The stronger concern is memorization. In a 2021 USENIX Security study, Nicholas Carlini and colleagues queried GPT-2 and recovered hundreds of verbatim text sequences from its training data ²⁷. The attack worked even on sequences that appeared in just one document in the training data ²⁸. Rare text can therefore still be recovered from a model, even when it appears only once.

Speech models show the same risk. A 2024 study found that when a speech model has seen a recording during training, a noisy version of that recording can make the model transcribe its memorized sensitive transcript ²⁹. These studies test research models, not the products you use, so they do not show how often a commercial dictation tool leaks what it heard. They do show why a policy that trains on your audio is different from one that does not. Can AI models leak what you dictated? answers the question in more detail.

On-device personalization and federated learning learn from you without uploading audio

Some systems learn from you without sending your audio or text to a server. In a 2018 study of a phone keyboard, Google researchers trained a language model on client devices, a method called federated learning, and they wrote that the work "demonstrates the feasibility and benefit of training language models on client devices without exporting sensitive user data to servers" ³⁰. In federated training, devices send model updates to a server instead of the typed text. It is a design study about next-word prediction, not speech recognition, and it does not show that any dictation product uses this method today.

On-device recognition is a separate question from training. Microsoft says device-based speech recognition in Windows 11 processes your voice locally and sends no voice data to Microsoft ³¹. Recognition that runs in the cloud is the path where your audio may be stored or used, so it is the one to check.

The two questions can have different answers in the same product. On a Mac, general Dictation is processed on the device ¹², while the Improve Siri and Dictation setting governs whether audio samples are stored and reviewed ⁵. A tool can recognize speech on your computer and still have a training setting to check. Personal voice models, where the app builds a voice profile from your own speech, follow the same rule: check where the profile is stored and whether it is used beyond your device.

You can opt out in settings, and on-device tools have nothing to opt out of

Start with the setting, because it is the one thing you control directly. On a Mac running a recent version of macOS, the steps below follow Apple's Mac Help page for Dictation, as checked on October 17, 2026. Older versions use System Preferences with different pane names ³². The setting is shared with Siri:

1. Open System Settings and click Privacy & Security.
2. Click Analytics & Improvements.
3. Turn off Improve Siri & Dictation ³².

On Windows, the steps depend on the app. For the older Copilot app, open its settings and go to Privacy. The Training on voice conversations setting is separate from the setting for text conversations, so you need to change both if you want neither used ⁶. Microsoft's page says it describes only the older app and that a newer version is covered on a different page, so check that page if your app's menus differ.

The on-device case is simpler. If a tool does all its recognition on the device and sends nothing, there is no training setting to change for the audio itself. Check the privacy policy anyway for account data, usage data and crash reports, which may still leave the device. For a full setup that covers the other places your words can go, see whether voice to text is safe and how to stop apps from training on your voice.

Yap, which we make, is an on-device dictation app for Mac and Windows, so recognition runs on your computer. Its settings and policy still decide what you share, and the rules above apply to any app you compare it with.

Some vendors' terms were not verified, so check them yourself

We did not verify Google's current voice typing terms, so this guide makes no claim about whether Google trains on dictated audio. We also did not verify Otter's Business plan terms or any Enterprise default, and the only plan-level opt-out we found was described in secondary sources, not in Otter's own pages. For any app not covered here, follow the same method: read the clause on model improvement and storage, find the matching setting, and write down the date you checked. Our guides on voice data retention, human review of voice recordings and AI dictation apps cover the next steps, and the voice privacy overview covers the whole subject.

If you need an answer in five minutes, check the policy in this order. Search the policy for the words "train", "improve" and "model". Read the sentence that names the setting, if there is one, and confirm which way it is set on the account you use. Then check whether the policy covers the consumer version, the business version or the API, because the rules can differ between them. Last, write down the date and the version of the policy you read. A policy can change between one check and the next, and a dated note is the only record that shows what you were told when you chose.