Anthropic began asking Claude users on October 4, 2026 to let it use their voice data to improve its AI models. The setting is voluntary, switched off by default, and separate from the existing option to train on text chats. If you talk to Claude, it is worth a minute to check what yours is set to.
This post covers what the prompt says, what reporting has and has not confirmed, how it relates to the policy Anthropic introduced in 2025, and a practical checklist for individuals and teams. The facts below come from BleepingComputer and gHacks, plus Anthropic's own consumer terms announcement. We could not find a dedicated Anthropic post for the voice change at time of writing, so gaps are flagged rather than filled in.
TL;DR: the voice-data opt-in at a glance
| Question | Answer |
|---|---|
| What changed? | Claude's voice features now show a prompt: "Allow us to use your voice data to improve our AI models," with Allow or Not now. |
| When? | Reported October 4 to 5, 2026. |
| Is it on by default? | No. It is opt-in and off by default. |
| Where is the setting? | Settings, then Privacy, then "Allow us to use your voice data." |
| Is it tied to chat training? | No. It is a separate toggle from training on chats and Claude Code sessions. |
| Can I undo it? | The prompt says you can turn it off or delete the data anytime in settings. |
| How long are recordings kept? | Not stated in the prompt or setting text, per gHacks. |
| Who is affected? | Users of Claude's voice features. Whether every account sees the prompt is unclear. |
What exactly does the prompt say?
According to the coverage, the prompt reads: "Allow us to use your voice data to improve our AI models." The supporting text explains that "Audio recordings and voice chat data help improve how Anthropic AI models understand and respond to speech," and that you can turn the setting off or delete the data at any time in settings.
Two details matter. First, the default is no. Second, the choice is independent. A user can consent to voice training while excluding text chats and Claude Code sessions, or the reverse. That is a finer-grained control than a single "improve Claude for everyone" switch, and it matches how people actually feel about the data: many are comfortable sharing a coding prompt but not their voice.
Why does Anthropic want voice data?
Speech is a hard modality. Accents, background noise, interruptions, filler words and overlapping speakers are all things a text model never sees. Real recordings of people talking to an assistant are the most direct training signal for how well a model understands and answers speech, which is exactly the reason Anthropic gives in the prompt.
The competitive context explains the timing. Voice has become a front line for assistants. OpenAI shipped GPT Realtime 2 voice models for developers and brought ChatGPT Voice with GPT-Live to the desktop, where you can steer coding agents by speaking. Open and independent projects such as Miso One push real-time speech synthesis, and Microsoft's MAI-Transcribe 2 targets streaming transcription accuracy. A lab that wants a competitive voice assistant needs speech data, and consenting users are the cleanest source.
How does this relate to Anthropic's 2025 training policy?
Voice is a new branch of a policy that already exists. On August 28, 2025, Anthropic changed its consumer terms so that people on Claude Free, Pro and Max could choose whether their data is used to improve Claude. In Anthropic's words, the change gave users "the choice to allow their data to be used to improve Claude and strengthen our safeguards."
The details from that announcement are useful context now:
- Retention: for users who opted in, retention extends to five years, versus the standard 30 days for those who declined. This applied to new or resumed conversations.
- Scope: the choice excluded commercial services, including Claude for Work, Government, Education and API access through third-party platforms.
- Control: users could change the preference at any time in Privacy Settings, and deleting a conversation excluded it from future training.
What nobody has confirmed publicly is whether the five-year retention applies to opted-in voice recordings. gHacks notes the prompt and setting text do not say how long approved recordings are kept. Until Anthropic answers, assume the longer window is possible and decide accordingly. This is an inference about risk, not a reported fact.
What are the privacy risks of voice specifically?
Voice differs from text in ways that raise the stakes of a retention window:
- Biometric character. A voice recording carries identity information beyond its words. It can be used to recognize you, and it is the raw material for cloning.
- Bystanders. Voice chats often capture other people in the room, who never consented.
- Context leakage. People say things aloud they would never type, including names, addresses and emotional detail.
- Hard to anonymize. Scrubbing a transcript is easy. Scrubbing a voice is not.
None of this means Anthropic is doing anything improper. The setting is opt-in, which is the right default. It means the decision deserves a deliberate yes or no instead of a reflexive tap on a pop-up.
There is a related reminder in how chat content is handled generally. Our look at Anthropic's chat-ending policy and a viral Claude conversation shows that consumer conversations are subject to usage policies and review. Voice does not sit outside that framework. If you want a technical approach to keeping sensitive prompts out of a provider's hands entirely, see how AgentCloak adds a privacy layer to AI prompts and the private personal AI launch from Underdog.
How to check and change your setting
- Open Claude on web or mobile.
- Go to Settings, then Privacy.
- Find Allow us to use your voice data. Confirm it is off unless you want it on.
- Separately, check the existing toggle for using your chats and Claude Code sessions to improve models. These are independent.
- If you previously chose Allow and changed your mind, switch it off and use the delete option the prompt describes.
- If you see the pop-up mid-conversation, Not now keeps the default of no training.
If the toggle does not appear for you, the rollout may not have reached your account yet. Reports do not make clear whether every user sees it.
What should teams and builders do?
- Individuals on consumer plans: decide per modality. A reasonable stance is yes to nothing until retention is clarified.
- Companies: if staff use personal Claude accounts for work calls or dictation, voice recordings of internal discussions may be in scope for a consumer opt-in. Set a policy: work audio goes through an approved commercial plan, where Anthropic's 2025 terms excluded training on the consumer toggle.
- App builders using the API: the reporting concerns Claude's own voice features, not API traffic. Still, review the data terms of your own speech stack. Our guide to Claude for Work covers the business side of the product.
- Regulated industries: voice can contain health or financial detail. Confirm with your compliance team before anyone uses consumer voice mode.
What is still unanswered
- How long approved voice recordings are retained.
- Whether recordings are transcribed, filtered for personal information, or reviewed by humans before training.
- Whether the setting applies to every plan and region, or only some accounts.
- What happens to data already used in a training run when you later turn the toggle off. The prompt promises deletion of the data, but whether already-trained models are affected is the standard open question for any opt-in.
These are exactly the questions a regulator or journalist will ask. Regulators are already looking at the labs' safety and data practices, as in the FTC probe of OpenAI and Anthropic, so expect scrutiny of how consent is collected.
Anthropic's consent record in context
Anthropic's relationship with user trust has been a recurring theme. Our timeline of Anthropic controversies tracks earlier flashpoints, and the pattern is consistent: the company tends to ship policy changes quietly through in-app prompts and terms updates, and the community reads the fine print afterward. The 2025 consumer terms update followed that path too, with a hard deadline of October 8, 2025 for existing users to choose a setting.
The voice prompt is gentler: Allow or Not now, off by default, no forced deadline reported. That is the better pattern. The remaining test is transparency on retention, which is what privacy-minded users will judge it on.
Bottom line
Nothing changes unless you tap Allow. If you use Claude's voice features, take a minute to open Settings, then Privacy, and confirm both the voice and chat toggles say what you intend. If Anthropic publishes a retention window for voice, we will update this post.
Related reading
- OpenAI GPT Realtime 2 voice models and API
- ChatGPT Voice on desktop with GPT-Live
- Miso One open-source real-time voice model
- Microsoft MAI-Transcribe 2 streaming transcription
- Claude chat-ending policy and a viral post
- AgentCloak privacy layer for AI prompts
- FTC probe of OpenAI and Anthropic
- Official: Anthropic consumer terms update
Details reflect reporting available on October 5, 2026. Check Claude's Privacy settings and Anthropic's policies for current terms.