Invest2 publishers2 min readPublished
Anthropic asks Claude users for the voice recordings it used to delete
Anthropic is asking Claude users to share voice chats for training, extending a 2025 text policy that keeps opted-in data up to five years. Teams with staff on consumer seats now have a second data setting to check before anyone talks to Claude.
The Investor · Invest desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Claude's voice audio used to be deleted after transcription, so only the transcript could reach training, and only if a user's settings allowed it.
- Anthropic did not train on consumer data at all until it changed its consumer terms on August 28, 2025.
- Users who opt out of training keep the earlier arrangement, with their data retained for about 30 days.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- decision Admins with staff on consumer seats have to confirm the voice setting user by user, because the last change billed as voluntary shipped with its toggle already on.
- exposure If opted-in audio follows the text rules, a recording could be held about 60 times longer than one from a user who declines.
- constraint While Anthropic keeps commercial data out to hold business customers, its voice training pool can only grow from consumer volume.
Five years comes to about 1,825 days. That is roughly 60 times the 30 or so days that data from Claude users who decline training is kept [1]. Crypto Briefing describes the guardrails on the 2025 consumer terms: the longer window applies only to new or resumed conversations, deleted chats drop out of future training runs, and the setting can be changed at any time [8]. Those rules were written for typed chats and coding sessions [3].
Crypto Briefing calls the voice request voluntary. Its second report says users can opt in, opt out and delete their data at any time [1][12]. The 2025 change was also framed as consent [3]. It arrived with the "Help improve Claude" toggle already switched on and a deadline for existing users to choose [6][14]. The reports do not say whether the voice setting starts switched on, or whether opted-in audio is kept for the five years that apply to text [7].
The two possible defaults give an operator different jobs. If the voice toggle copies the text one, recordings from staff on Free, Pro or Max seats go into the pool before anyone at the company has made a choice [5][6]. Should it start off, contributions depend on individuals saying yes, and the policy question shrinks to whether staff are allowed to. The audio used to be deleted once it became text [2]. A team that approved Claude voice on that basis needs to recheck the assumption seat by seat.
I think the exposure sits with consumer seats and stops there. The 2025 terms did not cover Anthropic's API or enterprise offerings [9]. Crypto Briefing's explanation is commercial. Businesses paying for that access care intensely where their data goes, and training on it would be a fast way to lose them, while consumers are where the volume lives [10]. On that reading, Anthropic is collecting training material from individual subscribers and keeping its commercial contracts out of the pool.
The counter-case is that voice is a more personal form of data than typed text, by Crypto Briefing's own description [15]. A lab that frames real user interactions as a route to better capability and safety [11] has reason to want speech from every source it can reach, and the second report says the aim is better speech recognition and responses [13]. I would be wrong if the voice request reaches API or enterprise accounts.
What to watch
- Opt-in rates and user sentiment across the Free, Pro and Max plans, the measure Crypto Briefing names for whether a voluntary ask produces volume.
- Whether Anthropic applies the 2025 deletion rule to voice, so that deleting a voice chat also removes its recording from future training runs.