Anthropic voice data is now something the company asks for directly. Claude users who open voice mode are starting to see a pop-up that reads “Allow us to use your voice data to improve our AI models“, with two buttons: Allow and Not now. BleepingComputer’s Mayank Parmar reported the prompt on 4 October 2026, along with a new switch in Claude’s privacy settings that, in the version his team saw, is turned off by default.
On its face this is a modest, optional change. It matters for three reasons.
First, Anthropic’s own privacy pages still say the opposite: that audio is deleted once it has been turned into text, and that “we do not use your voice for training our models”. Second, a voice recording carries far more than a typed message does. Your voice can identify you, it can reveal your health and mood, and it can pick up other people in the room. Third, the request lands days after a class action accused Anthropic of collecting Illinois users’ biometric data without the written disclosures that state law requires. Voiceprints sit on the same list.
Below we set out exactly what the Anthropic voice data prompt says, what Anthropic did with your voice until now, how it compares with the chat-training switch Anthropic introduced in 2025, why audio is a different kind of data, how OpenAI, Google, Meta and Amazon handle the same question, and what businesses using Claude should do.
Table of contents
- What the Anthropic Voice Data Prompt Says
- What Anthropic Did With Your Voice Until Now
- Anthropic Voice Data Follows the 2025 Chat-Training Switch
- Why Anthropic Voice Data Is Different From a Typed Chat
- The Illinois Lawsuit Behind the Anthropic Voice Data Debate
- How Anthropic Voice Data Compares With OpenAI, Google, Meta and Amazon
- What Anthropic Voice Data Means Under UK and EU Data Protection Law
- What Anthropic Voice Data Means for Businesses Using Claude
- How to Check or Change Your Anthropic Voice Data Setting
- Questions Anthropic Has Not Answered Yet
- Anthropic Voice Data FAQ
- References and Further Reading
What the Anthropic Voice Data Prompt Says
The prompt appears inside Claude’s voice features. It asks once, in plain terms, and gives a reason. According to BleepingComputer, the full text reads: “Allow us to use your voice data to improve our AI models. Audio recordings and voice chat data help improve how Anthropic AI models understand and respond to speech. Turn this off or delete this data anytime in settings.”
Two buttons, and a setting behind them
You can tap Allow or Not now. The same choice lives permanently under Settings > Privacy, where the Anthropic voice data option is labelled “Allow us to use your voice data”. Its description is slightly different from the pop-up: “For approved sessions, allow the use of your audio recordings and voice chat data to improve Anthropic AI models.”
Off unless you turn it on
In the interface BleepingComputer saw, the switch was disabled by default. Anthropic is therefore not quietly opting anyone in to sharing Anthropic voice data. You have to choose it, and you can turn it off again later.
Separate from the chat-training switch
The voice setting is independent of Anthropic’s existing “Help improve Claude” option, which controls whether your typed chats and Claude Code sessions are used for training. You could share Anthropic voice data while keeping your chats out of training, or the reverse.
| Where | What it says | Choices |
|---|---|---|
| Pop-up in voice features | “Allow us to use your voice data to improve our AI models” | Allow / Not now |
| Pop-up body text | “Audio recordings and voice chat data help improve how Anthropic AI models understand and respond to speech” | “Turn this off or delete this data anytime in settings” |
| Settings > Privacy | “For approved sessions, allow the use of your audio recordings and voice chat data” | Toggle, off by default in the build seen |
| Existing chat setting | “Help improve Claude” (chats and coding sessions) | Separate toggle, unchanged |
Two phrases in the Anthropic voice data wording are worth reading slowly. “Audio recordings” means the sound of your voice, not just the words. “For approved sessions” suggests consent applies to sessions from the point you allow it, rather than to everything you have ever said to Claude. Anthropic has not published a help article explaining either phrase, so both readings rest on the interface text alone.
What Anthropic Did With Your Voice Until Now
To see why the Anthropic voice data prompt is a real change, you need the baseline. Claude has two ways to use your voice, and Anthropic has documented them separately.
Dictation: audio deleted after transcription
Dictation turns your speech into a typed prompt. Anthropic’s privacy centre article on dictation in the Claude mobile apps, last updated on 16 March 2026, says: “After converting your speech input to text, we delete your audio recording but the text of your chat will be retained in accordance with our retention periods.” It adds: “We do not use your voice for training our models. If you have allowed us to use your chats or coding sessions to improve Claude, the transcribed text may be used in accordance with your privacy settings.”
So until now, the words you spoke could reach training, if you had opted in for chats. The sound of your voice, which is what makes Anthropic voice data different, could not.
Voice mode: transcripts kept like chats
Voice mode is a full spoken conversation. Anthropic’s help article on voice mode, updated on 16 September 2026, says it is a beta feature on every plan, including Free, Pro, Max, Team and Enterprise, on mobile, desktop and the web. Asked whether voice conversations are saved, it answers: “Yes. Textual transcripts of your audio conversations are saved in your chat history just like text conversations.” It says nothing about keeping the audio itself, let alone training on Anthropic voice data.
The privacy pages have not caught up
That leaves an awkward gap in how Anthropic voice data is documented. The consumer dictation article still says Anthropic does not use your voice for training. The commercial version of the same article, for Claude for Work and the API, says the same. Anthropic’s privacy policy, effective 10 September 2026, does not mention audio, voice, recordings or speech at all; we searched its text for each word. It does describe identity verification data, including “facial geometry templates”, which it notes “may be considered ‘biometric data’ in some jurisdictions”. Anyone reading the documents today would conclude that sharing Anthropic voice data for training is not possible.
| Anthropic page | Last updated | What it says about voice |
|---|---|---|
| Dictation privacy article (consumer) | 16 March 2026 | Audio deleted after transcription; “We do not use your voice for training our models” |
| Dictation privacy article (commercial) | 16 March 2026 | Same deletion rule; same no-training statement |
| Is my data used for model training? | 16 March 2026 | Covers chats and coding sessions only; no mention of audio |
| How long do you store my data? | 1 July 2026 | Retention periods for chats, feedback and flagged content; none for audio |
| Use voice mode (help centre) | 16 September 2026 | Text transcripts saved in chat history; audio not addressed |
| Privacy Policy | Effective 10 September 2026 | No mention of audio, voice or recordings |
Product changes often run ahead of help pages, and Anthropic may well publish updated wording when the prompt reaches everyone. Until it does, the pop-up is the only description of what the Anthropic voice data setting covers.
Anthropic Voice Data Follows the 2025 Chat-Training Switch
This is the second time in just over a year that Anthropic has asked Claude users for training data they were not previously giving.
What changed on 28 August 2025
Until August 2025, Anthropic did not train on consumer chats at all. On 28 August it updated its Consumer Terms and Privacy Policy so that users of Free, Pro and Max plans, including Claude Code on those accounts, could allow their chats and coding sessions to be used for training. Anyone who agreed had their retention period extended from 30 days to five years. Existing users initially had until 28 September 2025 to decide; Anthropic’s announcement now gives 8 October 2025. Commercial products, including the API and Claude for Work, were left out.
Default on then, default off now
The design of the choice is where the two requests differ most. The Verge described the 2025 pop-up: a large black “Accept” button, with a smaller line beneath it reading “Allow the use of your chats and coding sessions to train and improve Anthropic AI models” and a toggle that was “automatically set to ‘On'”. Anyone who clicked Accept without reading opted in.
The Anthropic voice data prompt is built the other way round. The choice itself is the headline, Not now sits beside Allow, and the setting starts off. Whether that reflects criticism of the 2025 design, the sensitivity of audio, or legal advice, Anthropic has not said. It is still the more defensible design.
| Question | Chats and coding sessions (2025) | Anthropic voice data (2026) |
|---|---|---|
| Setting name | Help improve Claude | Allow us to use your voice data |
| Default | On in the pop-up, per The Verge | Off, per BleepingComputer |
| Main button | Accept (the terms) | Allow or Not now (the data use) |
| Deadline | 8 October 2025 (first 28 September) | None reported |
| Published retention | Up to five years, de-identified | Not yet published |
| Plans covered | Free, Pro, Max | Not yet stated |
The retention question nobody has answered
Anthropic publishes unusually specific retention periods for most consumer data. Its privacy centre lists them: 30 days for a deleted chat, up to five years for chats in training pipelines when you have opted in, five years for feedback, two years for inputs and outputs flagged under its Usage Policy, and seven years for the trust and safety scores attached to those flags. There is no line yet for Anthropic voice data.
How long Anthropic says it keeps consumer data, in days (privacy centre, July 2026)
The figures are converted at 365 days a year, and each bar is scaled against the seven-year maximum. If audio follows the chat rule, a voice clip you share today could sit in a training pipeline until 2031. The pop-up says you can “delete this data anytime”, but the existing retention page also says that turning a setting off does not pull data out of training runs already in progress or models already trained. Expect the same to apply to Anthropic voice data.
Why Anthropic Voice Data Is Different From a Typed Chat
A transcript and a recording of the same sentence are not the same data. The difference is the reason voice privacy rules exist at all.
Your voice identifies you
A recording of your voice is a biometric signal. Speaker recognition systems turn it into a voiceprint, a numerical template that can pick you out of other recordings. Text can identify you through what you say. Audio identifies you through how you sound, even when you say nothing personal.
Voices carry more than words
Tone, pace, accent, breathing and hesitation all survive in a recording and disappear in a transcript. Those signals can hint at age, regional origin, mood, fatigue or illness, and Anthropic voice data shared for training would carry all of them. A speech model trained on them learns from all of it, whether or not anyone intended it to.
Other people in the room
Voice mode’s default is hands-free: Claude “listens continuously and responds to natural pauses”, in Anthropic’s words. Its help page advises a quiet space because hands-free “works best when Claude can detect only one voice”. In a shared office, a kitchen or a car, a recording can capture people who never saw the Anthropic voice data prompt and never agreed to anything. Google’s own Gemini privacy hub says it plainly: “Please respect others’ privacy and ask permission before you record or include them in a Live chat.”
Connected tools widen the net
Since July 2026, voice mode has run on Claude Opus, Sonnet and Haiku and can use connected tools such as Gmail, Google Calendar, Google Docs and Slack. Someone might ask Claude aloud to read out today’s emails and draft replies. Anthropic’s text-training rules say raw connector content is not used for training unless it is copied into the conversation. It has not said whether Anthropic voice data includes spoken summaries of that content, or whether Claude’s own spoken replies count as recordings.
| What a recording holds | Survives in a transcript? | Why it matters |
|---|---|---|
| The words you said | Yes | Already covered by the chat-training switch |
| The sound of your voice | No | Can be turned into a voiceprint that identifies you |
| Accent, tone and pace | No | Can hint at origin, mood or health |
| Background voices | Rarely | Bystanders have not consented |
| Room sounds | No | Can reveal location or company |
None of this makes sharing Anthropic voice data a bad idea for everyone. Speech models do get better with real speech, and people with strong accents, speech differences or less common languages stand to gain most from that. It does mean the choice deserves more thought than the chat switch did.
The Illinois Lawsuit Behind the Anthropic Voice Data Debate
Anthropic is already defending a claim under the strictest biometric privacy law in the United States, and the prompt arrives in that context.
What Colon v. Anthropic alleges
Courthouse News reported on 30 September that Jose Enrique Ortiz Colon, a Chicago resident, had filed a proposed class action against Anthropic in San Francisco County Superior Court. The case is not about Anthropic voice data at all. Colon says Claude locked him out until he completed identity verification. He photographed both sides of his driving licence and held his face to the camera for a “live image”. The check ran through Anthropic’s vendor, Persona Identities.
The 15-page complaint, filed by Wilshire Law Firm, brings two counts under the Illinois Biometric Information Privacy Act (BIPA). It says Anthropic never told Colon in writing how long his face scan would be kept, and that it has no public schedule for destroying such data. Its sharpest paragraph points out that Anthropic “knows how to publish a retention schedule”, listing the 30-day, two-year, five-year and seven-year periods it publishes for other data. “For the Biometric Data collected through government-ID and selfie verification, it publishes nothing,” it says.
An Anthropic spokesperson told Courthouse News: “We take the privacy of our users seriously. We’re aware of the complaint and are reviewing it.”
Voiceprints are on the same list
BIPA defines biometric identifiers to include fingerprints, retina or iris scans, scans of hand or face geometry, and voiceprints. Before collecting them, a company must tell people in writing what it is collecting, why and for how long, and get a written release. It must also publish a retention schedule and destruction guidelines.
The statute names a voiceprint, not a voice recording, so a plain audio file is not automatically covered. The risk rises if recordings are used to train or improve speaker recognition. The Anthropic voice data prompt is clear about the purpose in general terms. It says nothing yet about retention, which is exactly the gap the Colon complaint is built on.
What the 2024 amendment changed
In August 2024 Illinois amended BIPA so that repeated collection of the same identifier from the same person counts as one violation, not one per scan, and so that an electronic signature counts as a written release. The Seventh Circuit held in April 2026 that the amendment applies retroactively. The damages per person are unchanged, and so are the written disclosure duties that would apply if Anthropic voice data were ever used to build voiceprints.
BIPA statutory damages, per person and per 1,000 class members (illustrative arithmetic, US dollars)
The per-person figures are the statutory amounts the Colon complaint seeks. The per-1,000 rows simply multiply them; the size of any real class is unknown. The arithmetic explains why a company asking millions of people for audio would want its written disclosures in order before the first recording is kept.
How Anthropic Voice Data Compares With OpenAI, Google, Meta and Amazon
Anthropic is late to this question. Every large assistant maker has already decided whether to keep and train on audio, and they have not all chosen the same way.
| Company | Audio used for training? | Default | Notable detail |
|---|---|---|---|
| Anthropic (Claude) | Only if you allow it | Off | “For approved sessions”; retention not yet published |
| OpenAI (ChatGPT) | Only if you turn on “Include your audio recordings” | Off | Clips kept 30 days with the chat; shared clips may be reviewed by staff |
| Google (Gemini) | Only if you turn on the audio setting in Gemini Apps Activity | Off | Applies to audio already stored as well as new audio; some is reviewed by people |
| Meta (Ray-Ban Meta glasses) | Voice recordings stored to improve products | On, no opt-out since 29 April 2025 | Kept up to one year; delete recordings manually |
| Amazon (Echo) | Recordings go to the cloud | Local-only option removed 28 March 2025 | “Don’t save recordings” deletes after processing |
OpenAI: two switches and human review
OpenAI’s help centre says ChatGPT does not train on audio or video clips unless you share them. In personal Free, Plus and Pro workspaces you first turn on “Improve the model for everyone”, then a second switch, “Include your audio recordings”. If you do, “our teams may review shared clips”, and OpenAI says it takes “steps to reduce the amount of personal information in the clip” before training. Business, Enterprise and Edu users cannot share clips at all. Anthropic has not yet said whether people will listen to shared Anthropic voice data.
Google: off by default, but retroactive
Google’s Gemini privacy hub says “your audio and Gemini Live videos and screenshares aren’t used to improve Google services by default”. Turning the setting on covers audio “already stored in Gemini Apps Activity and added in the future”, and some of it is “reviewed by human reviewers”. Reviewed data is kept for up to three years, disconnected from your account. The Anthropic voice data setting’s “for approved sessions” wording, if it means what it appears to, is narrower.
Meta and Amazon went the other way
In April 2025 Meta removed the option to stop Ray-Ban Meta glasses storing voice recordings. TechCrunch reported that recordings can be kept “up to one year to help improve Meta’s products” and must be deleted one by one if you object. The same spring, Amazon removed the Echo setting that processed requests on the device without sending recordings to its cloud, ahead of Alexa+. It said fewer than 0.03% of customers used it.
Against that field, the Anthropic voice data design sits with OpenAI and Google: opt-in, off by default, and reversible. It falls short of both on documentation, because neither OpenAI nor Google is contradicting its own help pages.
What Anthropic Voice Data Means Under UK and EU Data Protection Law
For readers in the UK and EU, the legal frame is different from Illinois, but the questions are similar.
Personal data first, biometric data sometimes
Under UK GDPR, a recording of someone’s voice is personal data. The Information Commissioner’s Office biometric guidance says biometric data only becomes special category data when it is processed to uniquely identify someone. A clip used to make a speech model better at understanding accents is still personal data. If recordings are used to recognise who is speaking, the stricter rules for special category data apply.
Consent has to be specific and freely given
Anthropic appears to rely on consent for Anthropic voice data, which is why the design of the prompt matters. Consent under GDPR must be specific, informed, unambiguous and as easy to withdraw as to give. A separate toggle that starts off, sits apart from the chat setting and can be turned off at any time looks much closer to that standard than the 2025 pop-up did. The missing piece is information: people are entitled to know what is kept, for how long and who reviews it, and the published documents do not yet say.
Anthropic’s earlier move to train on chats was framed around improving models and safety classifiers. If the Anthropic voice data programme follows the same logic, a clear statement of purpose, retention and human review would remove most of the ambiguity. A gap between the product and its privacy notice is common in AI rollouts, and it is the first thing a data protection review should check.
What Anthropic Voice Data Means for Businesses Using Claude
Most of the risk for organisations does not come from enterprise contracts. It comes from staff using personal accounts.
Team and Enterprise accounts
Anthropic’s 2025 training change explicitly excluded services under its Commercial Terms, naming Claude for Work (the Team and Enterprise plans), the API, Claude for Government and Claude for Education. Its commercial dictation article still says voice is not used for training. BleepingComputer’s report describes the consumer app, and nothing so far suggests the Anthropic voice data prompt applies to Team or Enterprise workspaces. Voice mode itself is available on those plans, however, and Anthropic’s help page says Enterprise owners who want it switched off must contact Anthropic support. There is no admin toggle.
Staff using personal accounts
The bigger exposure is the employee who uses a personal Claude Pro account on a phone to talk through a client problem on the way to a meeting. If that person taps Allow, the Anthropic voice data they share may include client names, colleagues’ voices or details read out from connected email, and it could enter a consumer training programme. Your organisation has no contract covering it. This is a shadow AI problem, not a new one, but voice makes it easier to overshare without noticing.
A short policy checklist
| Action | Who | Why |
|---|---|---|
| State whether personal AI accounts may be used for work at all | IT and compliance | Consumer training settings sit outside your contracts |
| Ban voice mode near client calls, open-plan desks and meetings | Line managers | Bystanders and client data end up in recordings |
| Ask staff to check the Anthropic voice data setting on any personal account | All users | A tap on Allow is easy to forget |
| Decide whether Enterprise voice mode should be disabled | Claude owners | It requires a support request, not a toggle |
| Add voice and audio to your AI acceptable-use policy | Compliance | Most policies were written for typed prompts |
If your AI policy predates voice assistants, this is a good moment to revisit it as part of wider IT governance work. The Anthropic voice data prompt will not be the last request of its kind. OpenAI, Google and Meta already ask, and as we noted when Grok Bot gained a voice mode, spoken interfaces are spreading fast.
How to Check or Change Your Anthropic Voice Data Setting
The setting is not hard to find once you know where it lives. These steps follow the paths described by BleepingComputer and Anthropic’s help pages; labels may shift as the rollout continues.
On the web and desktop
Open Claude, click your name or initials, choose Settings, then Privacy. Look for “Allow us to use your voice data” and set the Anthropic voice data switch the way you want. While you are there, check “Help improve Claude” as well, because it controls your typed chats and Claude Code sessions separately.
On the mobile apps
Open the Claude app, go to Settings, then Privacy, and find the same Anthropic voice data option. If the pop-up appears during a voice conversation, Not now leaves the setting off and lets you decide later.
Deleting what you have shared
The prompt says you can “turn this off or delete this data anytime in settings”. Turning the Anthropic voice data setting off should stop new recordings being used. Based on Anthropic’s existing rules for chats, it will not remove anything already included in a training run that has started, or in a model already trained. If you want to limit what can be shared in future, you can also use typed chat for sensitive topics, or Claude’s incognito chats, which Anthropic says are never used to improve Claude.
Questions Anthropic Has Not Answered Yet
The interface text is clear about the basic choice. The details that would let someone make an informed decision about Anthropic voice data are still missing:
- Retention. How long are shared recordings kept, and are they de-identified like chats?
- Human review. Do people listen to shared clips, as OpenAI and Google say theirs may?
- Scope. Does “for approved sessions” cover only future sessions, and does it include Claude’s spoken replies and summaries of connected tools?
- Plans. Does the prompt appear on Free, Pro and Max only, or on any Team or Enterprise seats?
- Voiceprints. Will recordings be used for speaker recognition, or only speech understanding?
- Documentation. When will the dictation article stop saying “we do not use your voice for training our models”?
We will update this article if Anthropic publishes answers. For context on how courts are starting to treat the human voice itself, see our report on the Japanese ruling that human voices are protected, and for the developer side of synthetic speech, our piece on OpenAI’s Voices API.
Anthropic Voice Data FAQ
Is Anthropic training Claude on my voice?
Only if you agree. The new Anthropic voice data setting is off by default in the build BleepingComputer reviewed, and the pop-up offers Allow or Not now. If you never tap Allow, the existing rules apply: dictation audio is deleted after transcription, and only text can reach training, and then only if “Help improve Claude” is on.
Where is the Anthropic voice data setting?
In Settings > Privacy, labelled “Allow us to use your voice data”. It is separate from “Help improve Claude”, which covers typed chats and Claude Code sessions.
Does Anthropic keep voice recordings if I do not opt in?
Its published documents say audio from dictation is deleted once converted to text, and voice mode saves text transcripts in your chat history. They do not describe keeping audio. The new prompt implies recordings are kept for sessions you approve.
Does the change affect Claude for Work or the API?
Nothing reported so far suggests the Anthropic voice data prompt reaches them. Anthropic’s commercial privacy article still says voice is not used for training, and its 2025 training change excluded commercial services.
Can I delete Anthropic voice data I have already shared?
The prompt says you can delete it in settings. Anthropic’s rules for chats suggest deletion stops future use, but does not remove data from training runs already under way or from models already trained.
References and Further Reading
Anthropic asks Claude users to share voice data for AI model training (BleepingComputer)
How long do you store my data? (Anthropic Privacy Center)
Use voice mode (Claude Help Center)
Think through hard problems in voice mode (Claude blog)
Updates to Consumer Terms and Privacy Policy (Anthropic)
Anthropic will start training its AI models on chat transcripts (The Verge)
Chatbot user files class action against Anthropic over biometric data policy (Courthouse News)
Colon v. Anthropic class action complaint (PDF)
Seventh Circuit holds 2024 BIPA amendment applies retroactively (Sidley Data Matters)
ChatGPT Voice (OpenAI Help Center)
Gemini Apps Privacy Hub (Google)
If you own Ray-Ban Meta glasses, you should double-check your privacy settings (TechCrunch)
Amazon ends little-used privacy feature for Echo recordings (AP via NY1)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
