Anthropic wants Claude users' voice chats for training

According to BleepingComputer, Anthropic now wants more from Claude users than typed prompts. It is asking them to voluntarily share their voice conversations to help train and improve its models. The request builds on an earlier change to how Anthropic handles consumer data, which already made every Claude user decide whether their conversations could be used for training.

At a glance

• According to BleepingComputer, Anthropic's new request is voluntary and covers voice conversations specifically, extending the training question from typed chats to the spoken side of Claude.
• TechCrunch reported that Anthropic previously kept consumer chat data out of training, then reversed course and gave every Claude user until September 28 to decide whether conversations could be used.
• The dates do not line up: Anthropic's own page gives existing users until October 8, 2025, after which they must pick a training setting to keep using Claude at all.

If you have not been following, the voice request sits on top of a bigger shift. According to TechCrunch, Anthropic's back end used to delete consumer prompts and outputs within 30 days "unless legally or policy‑required to keep them longer," and kept inputs flagged as policy violations for up to two years. Under the new policy, TechCrunch reports, data from users who do not opt out is retained for five years.

Anthropic set October 8, 2025 as the cutoff for existing Claude users

A Stanford write-up describes the move as a change to Anthropic's terms of service covering conversation data. Anthropic's own documentation says those terms apply to Claude Free, Pro and Max, including Claude Code used from those accounts. They do not apply to services under Commercial Terms: Claude for Work, Government, Education and API use, including access through Amazon Bedrock and Google Cloud's Vertex AI.

The reports disagree on the deadline. TechCrunch, reporting on August 28, gave September 28 as the date by which every Claude user had to decide. Anthropic's page gives existing users until October 8, 2025, and says that after that date they must choose a model-training setting to keep using Claude. The update applies only to new or resumed chats and coding sessions.

TechCrunch offers its own reading. Its view is that LLM companies need large volumes of high-quality conversational data, and that real Claude interactions help Anthropic compete with OpenAI and Google. When TechCrunch asked what prompted the move, the company pointed it to its blog post.

BleepingComputer reports the new request targets Claude's voice conversations

As BleepingComputer describes it, the new request is voluntary: Claude users are invited to share their voice conversations to help train and improve Anthropic's models. According to Claude's help pages, voice mode lets users hold spoken conversations in hands-free or push-to-talk modes. It can use the same Claude models as text chat, along with connected tools such as Gmail, Google Calendar, Docs and Slack.

The same Help Center says voice conversations are saved as text transcripts in chat history, just like typed ones. So a spoken chat with Claude already exists as a text record, sitting next to your typed threads.

Anthropic frames the training push in terms of safety and capability. On its own page, the company says participation makes harmful-content detection more accurate and less likely to flag harmless conversations, and improves future Claude skills such as coding, analysis and reasoning. It also lists safeguards: deleted conversations are not used for training, automated tools filter or obfuscate sensitive data, and it does not sell user data to third parties.

Where does a spoken request travel before a model learns from it?

It passes through several stops. According to Cekura, voice AI pipelines pass audio through multiple layers: speech-to-text, the language model, text-to-speech and telephony. Each layer is a place where data could be stored or leaked. Cekura lists "Your data training someone's model" among the risks, and names training opt-outs as the way to reduce it.

Think of a letter handed through several sorting offices, any of which could keep a copy. A training opt-out decides what may be done with a copy. It does not decide whether the copies exist.

Asking users for voice data is not new outside Anthropic. According to Microsoft, it asks opt-in permission to sample and listen to voice clips to improve speech recognition for accents, dialects and noisy settings, and de-identifies the clips when it stores them. Microsoft says it stopped logging voice data for product improvement from October 30, 2020, unless users agree.

The reports leave open what "voice data" means here. Claude already stores spoken chats as text transcripts, so it is unclear whether Anthropic wants the audio itself, with its accents and background noise, or the transcripts that already sit in chat history. In our view, a voice opt-in only means something if it says plainly which of the two it covers.

What the voice opt-in still has to say

No deadline has been reported for the voice request. Nothing reported so far says whether it follows the consumer-only scope of the chat terms. Two details will show how far it reaches. One is whether the five-year retention TechCrunch described applies to shared voice data. The other is whether Anthropic's rule that deleted conversations are not used for training covers audio as well as transcripts.