Anthropic Adds a Separate Claude Voice-Data Training Opt-In
Published by AINave Editorial
Anthropic has started asking Claude users whether it can use their voice conversations to improve its AI models. The choice appears when people use Claude’s voice features, and offers “Allow” or “Not now.” The distinction that matters is that voice consent has its own control, separate from the setting for chats and Claude Code sessions. The report describes the voice request and separate controls.
Voice training has its own switch
In Settings > Privacy, the option is labeled “Allow us to use your voice data.” Its description refers to using audio recordings and voice chat data from approved sessions to improve Anthropic’s models. The report says the toggle was disabled by default in the interface it observed, rather than establishing that every account has the same state. The setting and its observed default are described in the report.
That separation lets users make different choices for different kinds of conversations: the report says someone can allow voice data use while keeping regular chats and Claude Code sessions excluded, or choose the reverse. Consent to one category therefore does not, according to the reported controls, automatically settle consent for the others. Voice training is described as separate from chat and coding-session training.
For users, this is a more specific choice than a single broad training preference. Voice conversations involve audio recordings as well as voice-chat data, and Anthropic’s prompt explicitly asks permission to use that material. The available evidence describes the request and controls, but does not detail how approved recordings are processed or what kinds of model improvements result. Anthropic’s prompt identifies audio recordings and voice chat data as material for improvement.
Turning sharing off is not the same as deleting data
Users can choose “Not now” at the prompt or turn off the voice-data option later in Privacy settings. Anthropic’s prompt also says users can delete voice data from settings. Those are distinct controls: turning off future sharing should not be treated as proof that material previously shared has been deleted. The report describes both changing the setting and deleting voice data.
The useful detail is that voice-data consent is independently manageable, at least in the reported interface. Anyone checking their preferences should look at the voice setting itself rather than infer its status from the controls for ordinary chats or coding sessions. The report identifies a dedicated voice option under Settings > Privacy.
