AI Voice Disclosure Rules: A Platform-by-Platform Guide

If you publish audio made with an AI voice, three separate sets of rules can apply to the same file at once: the EU AI Act, the policy of the platform you upload to, and the contract you signed with whoever supplied the voice. They do not say the same thing, and none of them replaces the others.

This page collects what the major platforms currently require, with links to the source documents so you can check the rule instead of guessing at it. Everything below was checked in August 2026. Platform policies change without notice, and the dates given are the dates of the sources, not of this article.

The EU baseline every platform rule sits on top of

Article 50 of the AI Act has applied since 2 August 2026. It creates two duties that are easy to confuse. Providers of generative AI systems must mark their output in a machine-readable format so it can be detected as artificially generated. Deployers who publish a deepfake must disclose that fact to the audience. We cover the mechanics in detail in our guide to Article 50 and synthetic voice.

What is new since then is the interpretive layer. The Commission published the final Code of Practice on Transparency of AI-generated Content on 10 June 2026. It has two sections: one for providers on marking and detection, one for deployers on labelling deepfakes and AI-generated text. The Commission and the AI Board have confirmed the code as an adequate voluntary tool for demonstrating compliance.

Signing is optional. The obligation is not. By the end of July 2026, around 190 organisations had signed, according to the Commission. Anyone who chooses a different route has to show their own measures are adequate, and that gets assessed individually by national market surveillance authorities.

In Germany, that authority now has a name. The Gesetz zur Marktüberwachung und Innovationsförderung von künstlicher Intelligenz (KI-MIG) came into force on 29 July 2026, and the Bundesnetzagentur became market surveillance body, central contact point and complaints body for the AI Act. Media content stays with the Landesmedienanstalten under a deliberately hybrid split, which matters if your work is broadcast or dubbed rather than uploaded.

YouTube: one toggle, and a label you cannot remove

YouTube handles this at upload. In YouTube Studio, under Details, there is an Altered content question. You answer yes if the content is realistic and has been meaningfully altered or synthetically generated. Unrealistic content and minor, primarily aesthetic edits do not require disclosure.

The disclosure surfaces to viewers in the How this content was made section of the expanded description, and in some cases on the player itself. YouTube states that where content is undisclosed, it may proactively apply a label that the creator has no option to remove.

There is a second route into that same label. YouTube carries forward disclosures embedded as Content Credentials under the C2PA standard, version 2.1 or higher, where they indicate the entire video was made with AI. When the disclosure arrives that way, viewers see an attribution to the signing authority alongside it.

For voice work specifically: the toggle asks about the content, not about the video track. A corporate explainer whose only synthetic element is the narration is still realistic content that has been meaningfully altered. Read the question as covering the audio, because nothing in the policy carves audio out.

TikTok and Meta: realistic audio counts

TikTok requires creators to label realistic AI-generated content, and its guidance names images, audio and video together. Creators can apply the platform label, which displays as creator labelled as AI-generated, or disclose in the caption, a sticker or the description. TikTok says applying the label does not affect distribution, provided the content otherwise meets its guidelines. The relevant policy sits in the Integrity and Authenticity section of the Community Guidelines.

Meta labels this class of content as AI Info. Its policy requires people to use the disclosure tool when posting organic content containing photorealistic video or realistic-sounding audio that was digitally created or altered, and Meta reserves the right to apply penalties where they do not. Since 1 June 2026, Meta has also run automated detection on ad media, adding an AI Info label in the About this ad panel when third-party generative tools are detected. Meta signed the EU code of practice in July 2026.

The common thread across both is worth naming. Neither platform asks whether the voice was cloned from a specific person. Both ask whether the result sounds real. A wholly invented synthetic narrator that no listener would question is inside the policy.

Audiobooks: the label attaches to the narrator field

Audiobook retailers took a different approach. Rather than a viewer-facing badge, they put the disclosure in the product metadata, where it travels with the title through distribution.

Spotify began accepting ElevenLabs-narrated audiobooks through Findaway Voices on 20 February 2025. All digitally narrated titles are marked in the metadata on Spotify and at the other retailers Findaway distributes to, and the book description is prepended with a first sentence reading: This audiobook is narrated by a digital voice.

Audible went further and attached the label to the narrator. In a post dated 9 July 2025, ACX opened a US-only beta letting invited narrators create and monetise replicas of their own voices. Titles narrated with a voice replica are labelled in the narrator field of the listing and detail page. Audible states it will not separately use a narrator’s replica for any content without their approval, and narrators choose per title between a per-finished-hour rate, royalty share, or royalty share plus.

That last detail is the one narrators should sit with. On ACX the disclosure is not a generic AI badge somewhere in the footer — it is next to your name, on every title, permanently. If you are weighing a replica deal, our breakdown of what AI voice platforms actually pay is the companion piece to this one.

What a compliant label looks like

The Commission has published a free set of EU icons for labelling AI-generated content, as part of Section 2 of the code of practice. There are three: a basic icon for content where AI was involved, one for fully AI-generated content, and one for content that was partially AI-modified. Each comes in four variants — black, white, and both at 50 per cent transparency. No attribution to the Commission is required.

The placement rules are more specific than most people expect. Summarised from the Commission guidance:

• The icon must be clearly perceivable at the latest at first exposure to the content
• It must sit where no intervening overlay elements cover it
• It must be embedded into the content itself, or delivered through an equivalent alternative such as a user interface overlay, and must remain visible when the content is reshared or downloaded
• It should carry alt text or an ARIA label for assistive technology, use plain language, and stay on screen long enough to be read

The Commission also notes that user testing improved results across all measures when the icon was accompanied by a text label. Its own worked example for audio is a deepfake video carrying the words voices generated with, followed by the basic icon. Using the icons is optional; labelling under Article 50 is not, and the icons alone do not establish compliance.

Where the platform rules stop

Three gaps are worth marking clearly, because they catch producers out.

Not every synthetic voice is a deepfake. Article 50(4) reaches audio that resembles existing persons and would falsely appear authentic. A wholly invented voice that resembles nobody may fall outside the deployer’s disclosure duty while the provider’s machine-readable marking duty still applies. That is a distinction about which obligation bites, not a licence to stay quiet.

Creative work gets a softer duty, not an exemption. Where a deepfake forms part of an evidently artistic, creative, satirical or fictional work, the Commission states that transparency is limited to disclosure in a manner that does not hamper the display or enjoyment of the work. Dubbing and animation sit close to this line. How close is not settled, and anyone telling you dubbing is simply exempt is going further than the text does.

Most voice work never touches a platform at all. IVR prompts, e-learning inside a corporate LMS, in-store audio, museum guides, internal training — none of these pass through an upload form with a disclosure toggle. The AI Act still applies. The only thing missing is the platform that would have reminded you.

Before you publish: a working checklist

1. Establish whether you are the provider, the deployer, or both. Fine-tuning a cloned voice and then publishing it makes you both.
2. Ask your TTS vendor, in writing, whether output carries machine-readable marking, and which standard it uses.
3. Check whether the vendor has signed the code of practice. If not, ask what alternative measures they will rely on.
4. Set the platform disclosure at upload rather than after. Retrofitting a label after a proactive one has been applied is not a route back.
5. Keep the consent paperwork for any cloned voice with the project file, not in an inbox. Our country-by-country look at voice rights covers what that consent needs to say.
6. If your delivery never touches a platform, write the disclosure decision into the project brief so it is a decision rather than an omission.

None of this is an argument against synthetic voice. It is an argument for knowing which of the three rulebooks you are standing in. If a project needs a human performance and a clean chain of consent, Voicfy connects you with native-language voice talent directly — you can post a brief and have quotes back the same day.