Anthropic Starts Asking Claude Users to Hand Over Their Voice Data — Opt-In Only, Off by Default

Anthropic has started asking Claude users whether their voice conversations can be used to train its AI models. The request appears as a prompt when people use Claude's voice features, and it is backed by a new setting in the app's privacy panel. The change was first reported by BleepingComputer on October 4 and has since been confirmed by multiple outlets seeing the same interface.

The prompt reads: "Allow us to use your voice data to improve our AI models", with supporting text explaining that audio recordings and voice chat data help Anthropic's models understand and respond to speech. Users can respond with "Allow" or "Not now", and the prompt makes clear both options are available without penalty — declining does not restrict access to the voice features themselves.

Architecturally, the interesting part is that voice gets its own switch. Inside Settings > Privacy, the new option — labeled "Allow us to use your voice data" — is independent of Anthropic's existing training toggles for regular chats and Claude Code sessions. A user can therefore permit voice training while keeping text chats and coding sessions out of training, or make the opposite choice. In the interfaces observed so far, the voice toggle is disabled by default, meaning nothing changes unless a user actively opts in.

Anthropic says users can turn the option off later and delete their shared voice data from settings if they change their mind. What the company has not said is how long approved recordings are retained before deletion takes effect, whether recordings are reviewed by humans before training, or whether the prompt is being shown to every Claude account or rolled out gradually. Those gaps are exactly the details privacy researchers tend to ask about first.

The new control also sits awkwardly next to Anthropic's earlier guidance on dictation. The company's privacy-center material has said Claude's mobile dictation converts speech to text, deletes the audio after transcription, and does not use the user's voice to train models. Voice chat is a different mode — a free-flowing conversation rather than a speech-to-text input method — and the new prompt treats it as a distinct data category with its own consent flow.

Voice raises the stakes in ways typed prompts do not. A recording carries biometric identity information that can be used to recognize or clone a speaker, it routinely captures other people in the room who never agreed to anything, and it is far harder to anonymize than a transcript. People also say things aloud they would never type — names, addresses, emotional detail. An opt-in default addresses the consent problem; it does not by itself answer the retention, bystander-consent and secondary-use questions.

There is also a consumer-versus-enterprise distinction worth keeping straight. The reported prompt concerns Claude's consumer plans, not Claude for Work or the Anthropic API, which historically operate under separate commercial data commitments that exclude training on customer content. Companies whose employees use personal Claude accounts for work calls should not assume the enterprise policy covers them — that mismatch is where most AI privacy incidents actually originate.

The timing is not incidental. Regulators on both sides of the Atlantic are scrutinizing how AI labs collect training data, and the FTC's industry-wide probe of frontier labs including OpenAI and Anthropic is explicitly focused on consumer protection. Anthropic's choice of an off-by-default, independently controllable, deletable toggle is the more cautious pattern — the remaining test is whether the company publishes a clear retention window before someone asks for one.