ElevenLabs builds agent voices on a self-attested consent claim

What happened

ElevenLabs sells a platform for building conversational voice agents, documented on an ongoing basis rather than tied to one launch date. Its developer documentation, its public site, and its terms of use together describe a company that sells both voice cloning and agent deployment as separate, combinable products. This description reflects those pages as retrieved on 16 September 2026.

What the documents show

The homepage lists voice cloning as a feature of its creative tools: 'Clone your voice, design one from a prompt, or explore 10,000+ voices from the library.' The agent documentation describes a separate product, stating that agents 'accomplish tasks through natural dialogue' and can draw on uploaded documents through retrieval for grounded answers, offering a choice from '5k+ voices across 31 languages.' The terms of use set the conditions for supplying a voice: a user 'may be asked to upload audio recordings of your voice or the voice you are authorized to share with us,' and separately state that a user 'may not provide Input or create Output for which you do not have all the rights necessary to grant us the license described above.' The same document states that ElevenLabs 'will not commercialize your voice on a standalone basis without your permission.'

The mechanism

The consent structure here rests on self-attestation rather than independent verification: the terms ask an uploader to confirm they are 'authorized' to share a given voice, but the pages reviewed do not describe a step that checks that claim against the actual speaker before cloning proceeds. A companion app built on this platform inherits whatever consent chain was used upstream. That is a different object from a licensed voice, individually contracted with a named person, and different again from a voice 'designed from a prompt,' which is synthetic and modeled on no specific real person at all. All three can sound similarly natural to a listener while resting on entirely different permission structures.

What it leaves open

The documents reviewed do not describe a verification mechanism, such as a recorded consent statement matched against the uploaded audio, beyond the written attestation requirement in the terms. Whether such a step exists elsewhere in the product is not established by the pages opened here.

• Was a given synthetic voice cloned from a real person, licensed from them individually, or generated from a text prompt?
• If cloned, what verified the uploader's claim to be authorized to use that voice?
• Does the platform's own documentation distinguish these three cases anywhere a user would see it?

A natural-sounding agent voice does not, on its own, tell a listener which of these consent paths produced it, and the terms place the burden of that distinction on the uploader's word.