Anthropic Asks Claude Users to Share Voice Data for AI Model Training

Anthropic Begins Prompting Claude Users to Share Raw Voice Recordings for Training

Anthropic has started showing Claude users a new prompt asking them to voluntarily share their voice conversations to help train and improve the company's AI models, according to reporting by BleepingComputer on October 4, 2026. The prompt, which appears when a user engages Claude's voice features, 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 are presented with two choices — Allow or Not now — and the setting is disabled by default, meaning no one is auto-enrolled. This marks a notable shift: previously, Claude's voice audio was deleted immediately after transcription and was never itself used for training, even for users who had opted into text-chat training. Now, for the first time, the actual audio — not just the resulting transcript — can be retained and used, but only for users who explicitly say yes.

Details

What Changed and Why It Matters

Voice Data Was Previously Off-Limits

Since Claude's voice mode launched, Anthropic treated spoken audio differently from typed conversations. Audio was transcribed to text and then discarded — the recording itself never persisted, and even users who had opted into model training only ever contributed the text transcript, not the sound of their voice. That architecture meant Claude's speech models couldn't learn directly from real user audio: tone, accent, background noise, interruptions, and the acoustic quirks that make speech recognition and generation hard in practice. The new prompt changes that by asking users to let Anthropic keep the actual recordings, which the company says will help its models "understand and respond to speech" more accurately.

An Opt-In Layered on Top of an Earlier Opt-In

This voice-data request does not replace Anthropic's existing text-chat training policy — it sits alongside it as a second, independent decision point. In September 2025, Anthropic ended its longstanding promise not to train on consumer conversations at all, instead asking Free, Pro, and Max users to choose whether their chats and coding sessions could be used for model improvement. Users who opted in had their data retained in de-identified form for up to five years; users who declined kept the standard 30-day retention window with no training use. Existing users had until October 8, 2025 to make that choice to keep using Claude. The voice-data prompt introduced this week follows the same opt-in logic but is scoped specifically to audio, meaning a user who already opted into chat training is not automatically opted into voice training, and vice versa — each requires its own affirmative "Allow."

How It Compares to OpenAI and Google

Anthropic's approach stands out against how its two largest competitors handle similar data. OpenAI's ChatGPT trains on Free and Plus-tier conversations by default, requiring users to manually turn off "Improve the model for everyone" in Settings → Data Controls to stop future conversations from being used — and that opt-out is not retroactive. Google's Gemini behaves similarly on consumer accounts, using conversations to improve its models unless a user disables Web & App Activity through Google's separate account-wide activity controls; Google has also previously paid a US$68 million settlement over its handling of Assistant voice-recording practices. Business and workspace tiers are excluded from training at all three companies. Anthropic's voice prompt, by contrast, requires an active, specific "Allow" rather than defaulting to on and asking users to find the off-switch — a meaningfully different consent model, even as the underlying business incentive (more training data to improve speech models) is the same across the industry.

What Anthropic Says About Safeguards

Anthropic has stated that shared data — covering both the September 2025 chat policy and this voice extension — will not be sold to third parties and is filtered to reduce exposure of sensitive information before being used in training pipelines. Deleted conversations and recordings are not used for training under any circumstance, and users can revisit their choice in Privacy Settings at any time. The company has not published a separate, detailed retention schedule specific to voice data beyond directing users to the same settings panel used for chat-training preferences.

What This Means for Users

Privacy-conscious users should treat this as a deliberate decision point rather than something to click through. Voice recordings are a more sensitive data category than text: they can reveal accent, emotional state, background environment, and — depending on what was said aloud during a session — potentially more candid or offhand information than a typed prompt. Because the toggle is opt-in and off by default, no action is required to avoid participating, but the prompt can appear during active use of voice features, where it's easy to tap "Allow" without reading the explanation.

For Individual Claude Users

• Check Settings → Privacy on the Claude app to confirm whether voice-data sharing is enabled, even if you don't recall explicitly turning it on.
• Treat the voice prompt and the chat-training prompt as two separate decisions — opting out of one does not opt you out of the other.
• If you've previously allowed voice data sharing and change your mind, use the in-app option to delete previously shared voice data, not just disable the toggle going forward.
• Avoid using Claude's voice mode for sensitive conversations (health, financial, legal, or workplace-confidential matters) regardless of your training-data settings, since audio may still be retained short-term for service functionality.

For IT Administrators and Security Teams

• Confirm which Claude plan your organization uses — Work, Enterprise, Education, and Gov tiers are excluded from both the chat and voice training programs, but Free, Pro, and Max accounts used informally by employees are not.
• Add Claude's voice-data and chat-training settings to onboarding/offboarding or BYOD-AI checklists, since individual employees on personal consumer accounts could inadvertently share work-adjacent conversations.
• Update internal AI acceptable-use policies to explicitly address voice assistants, not just text-based chat tools, as audio-based AI features expand across vendors.

For Privacy-Conscious Teams Evaluating AI Vendors

• Compare default training behavior, not just the existence of an opt-out, when assessing AI vendors — Anthropic's opt-in-by-default-off model for voice differs meaningfully from the opt-out models used by OpenAI and Google for text.
• Ask vendors directly whether raw audio (not just transcripts) is retained and for how long, since this distinction is easy to miss in marketing copy.
• Document that deletion requests apply to both already-shared data and future collection, and verify the vendor's own published language on retroactive deletion.

Key Takeaways

1. Anthropic is now asking Claude users to opt in, via an explicit "Allow" prompt, to share raw voice recordings for AI model training — a new category of data beyond the text-chat training option introduced in 2025.
2. The setting is off by default; no user is automatically enrolled, and the choice can be reversed or the shared data deleted at any time in Settings → Privacy.
3. Voice-data sharing is tracked separately from the September 2025 chat/coding-session training opt-in — users must make two independent decisions, not one.
4. Previously, Claude deleted voice audio immediately after transcription; only the resulting text (and only for opted-in users) could ever feed training. Now the audio itself can be retained if a user agrees.
5. Claude for Work, Enterprise, Education, and Gov accounts are excluded from both the chat and voice training programs — this affects consumer Free, Pro, and Max plans only.
6. Anthropic's opt-in, default-off approach contrasts with OpenAI's ChatGPT and Google's Gemini, both of which train on consumer conversations by default and require users to actively opt out.

Sources

• Anthropic asks Claude users to share voice data for AI model training — BleepingComputer
• Anthropic Shifts Privacy Stance, Lets Users Share Data for AI Training — Bitdefender