Claude now asks to train on voice chats. Your AI policy probably covers only the typed ones

Most small-company AI policies were written about typing. What staff may paste into a chat window, which accounts they may use, whether chats can be used for training. On October 4, a new prompt in Claude showed why that framing is now too narrow: spoken conversations have their own consent switch.

What changed

BleepingComputer reported that Claude has started prompting people who use its voice features to share those conversations for model training. The prompt reads, in part: “Allow us to use your voice data to improve our AI models.”

A matching option sits under Settings > Privacy. Its description, as quoted in the report, says that “for approved sessions” it allows the use of “your audio recordings and voice chat data to improve Anthropic AI models.”

Three details matter for a security lead:

• It is off by default in the interface the reporters saw. Nobody is opted in automatically.
• It is a separate toggle. It works independently of the existing setting for chats and Claude Code sessions. According to gHacks, a user can allow training on voice while keeping typed chats excluded, or the reverse.
• It can be reversed. The prompt says users can turn it off or delete the data anytime in settings.

Several things are still unknown. Neither report quotes an official Anthropic statement; both work from the product screens. gHacks notes that the prompt and the setting text do not state how long approved voice recordings are kept, and that Anthropic has not said whether the request is shown to all Claude users or only to some accounts. Neither report says which plans are affected, or what counts as an “approved session.”

Why a separate toggle is a policy problem

Off by default is the right design. The trouble sits on your side.

If your acceptable-use policy says “do not allow AI vendors to train on company data,” most people will read that as the chat setting they already checked. A new prompt, shown during a voice session and worded as a product improvement, is easy to accept without connecting it to that rule. The person tapping “Allow” may not be breaking a rule as they understand it.

Spoken conversations also carry different material than typed ones. People talk through things they would never type: a customer call being summarized on the drive home, an incident walkthrough, a negotiation position, a personnel issue. And a recording of someone’s voice is the raw material that voice-cloning fraud depends on. That is a reason to know where recordings of your finance lead or CEO are going, whatever any one vendor does with them.

The biggest exposure is personal accounts. If staff use Claude’s voice features on a personal login, your company’s admin settings never touch that toggle. This is shadow AI in its most ordinary form: a phone app, a commute, a helpful prompt.

A 20-minute test you can run this week

Take your current AI policy and answer these five questions in writing. Each “no” or “don’t know” is a gap.

1. Does the policy name voice? Search it for “voice,” “audio,” “recording” and “dictation.” If the only nouns are “prompts” and “chats,” it was written for typing.
2. Does it cover training settings per data type? A rule that says “turn off training” should list every toggle by name, not one. Today, for Claude, that is at least two: chats and Claude Code, and voice.
3. Do you know who uses voice features? Ask, without blame, in your next team meeting or a one-question form: “Do you ever talk to an AI assistant out loud about work?” Include personal accounts and phones.
4. Can you see the setting for company accounts? For any AI tool on a company plan, check whether an admin can view or enforce training settings. If you cannot confirm it from the admin console or vendor documentation, write “unknown” rather than assume.
5. Do the people whose voices matter most know about cloning risk? Your CEO, finance lead and anyone who approves payments should know that their recorded voice is an asset worth controlling.

A simple rule to adopt from the answers: if a new AI consent prompt appears, the default answer for work use is “not now,” and the person forwards a screenshot to whoever owns AI policy. That turns every new vendor prompt into a ten-second report instead of a silent decision.

What this does not mean

This is not a finding that Claude’s voice feature is unsafe, and the setting does nothing until a user turns it on. The point is narrower. Vendors are splitting data consent into more pieces, one input type at a time. Each split is a place where a one-line policy stops matching the product.

What to do this week

• Search your AI policy for the word “voice.” Add a line if it is missing.
• List every training toggle by name for the AI tools your team uses, typed and spoken.
• Ask your team who uses AI voice features for work, on any account.
• Brief your payment approvers on why their recorded voice matters.
• Adopt the “not now, screenshot it” rule for new consent prompts.

This post is general education, not security or legal advice for your company. If you want a draft acceptable-use policy for the AI tools your staff already use, and a short list of agent permissions to tighten, that is the kind of work we do in a consult block.