Anthropic has started asking Claude users to voluntarily share their voice conversations to help train and improve its AI models. The new prompt appears when people use Claude's voice features and explicitly requests permission to use their voice data for training, according to a report by BleepingComputer.
The prompt reads: "Allow us to use your voice data to improve our AI models," followed by an explanation that "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." Users can choose Allow or Not now, and the request can also be managed later through a dedicated toggle inside Claude's privacy settings. If you are tracking how major AI labs handle user consent, this is one of the more concrete examples of granular, reversible data controls shipping in a mainstream assistant. For more context, see our coverage of the latest AI developments.
A separate switch for voice
The most notable detail is that voice training is handled separately from Anthropic's existing option to use chats and coding sessions for model improvement. That means a user could allow Anthropic to train on their voice conversations while keeping regular chats and Claude Code sessions excluded from training, or the reverse.
Inside Settings > Privacy, Anthropic describes the option as "Allow us to use your voice data," with supporting text stating: "For approved sessions, allow the use of your audio recordings and voice chat data to improve Anthropic AI models."
According to BleepingComputer, the setting is disabled by default in the interface the outlet reviewed, so Anthropic is not automatically enrolling users into voice-data sharing. People who opt in can also turn the option off later, and the company says voice data can be deleted from settings at any time.
Why voice data matters for training
Voice is a fundamentally different data type from text. Speech carries accent, cadence, emotion and background noise, all of which are difficult to synthesize convincingly and expensive to label. Training on real conversations is one of the most direct ways for a lab to make a voice assistant sound more natural, handle interruptions gracefully and recover from mishearings.
That value is exactly why voice data is sensitive. A recording of someone speaking is biometric in nature, can reveal information the speaker never intended to share, and often captures other people in the room who never agreed to anything. Regulators in several jurisdictions treat voice recordings with more caution than text, and any mishandling of them tends to attract outsized scrutiny.
By making the voice toggle distinct from the general training opt-in, Anthropic is acknowledging that users may want different rules for different data types. It is a design choice that privacy-conscious users have long requested: instead of one blanket switch that bundles text, code and audio together, each category can be evaluated on its own terms.
Opt-in by default, but the ask is there
The rollout also reflects where the industry's center of gravity has moved. In the early era of chatbots, user conversations were frequently used for training with little ceremony. A wave of user backlash, regulatory attention and enterprise customer requirements pushed major labs toward opt-out controls, and in some cases opt-in defaults for sensitive data.
Anthropic's approach here sits firmly in the opt-in camp: nothing is shared unless a user actively allows it, the prompt is shown at the moment someone uses voice features, and revocation is available in settings. Critics would still note that prompts appearing inside a product flow tend to convert a meaningful share of users, and that the practical pressure to click Allow is real for people who want the feature to work well.
The company's own wording tries to preempt those concerns by emphasizing reversibility: the data can be turned off or deleted at any time, and the permission is scoped to approved sessions rather than treated as a permanent grant.
What users should know
For Claude users, the practical takeaway is straightforward. If you use Claude's voice features, expect to see the prompt. If you decline, nothing about the voice feature itself should change; you simply will not contribute recordings to training. If you accept and later change your mind, the controls live in Settings > Privacy, where the voice toggle sits alongside the existing chat and Claude Code training options.
Anyone particularly cautious about audio should also remember that voice data sharing, like chat training, is a choice that can be revisited. Checking the privacy settings once, deliberately, is a better posture than reacting to prompts as they appear.
For the industry, the rollout is another data point in an ongoing negotiation between AI labs and their users over who funds model improvement with their data. Voice assistants are becoming a primary interface for AI, and how companies ask for speech data now will shape user expectations for years. Anthropic's version of the ask is transparent, segmented and reversible, which is a higher bar than the industry met just a few years ago. Whether that becomes the standard template, or floors get raised further, will depend on how competitors handle the same question and how regulators weigh in on biometric data collected through consumer AI products.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
