Why Disclosure Became a Practical Question
For most of the history of dubbing, nobody disclosed anything. Audiences understood that a foreign-language film had been dubbed by actors, and the convention was so well established that stating it would have been odd.
Synthetic voice changed the question in two ways. It made voice replication of a specific person possible, which raises consent and identity questions that generic dubbing never did. And it made high-quality voice generation cheap enough to be used at a scale where audiences can no longer assume they know what they are hearing.
The result is a set of expectations that have formed quickly and are now partly codified: platform policies requiring labels on realistic synthetic media, regulatory transparency obligations in several jurisdictions, and audience norms that punish undisclosed synthesis more than they punish synthesis itself.
For anyone running a multilingual video programme, this is a workflow question rather than a philosophical one. What gets labelled, where the label lives, and who is responsible for applying it.
The Distinctions That Determine the Answer
Not all synthetic voice raises the same questions, and treating it as one category produces either over-labelling or under-labelling.
Generic synthetic narration. A stock or custom voice that is not modelled on an identifiable person, reading scripted content. This is the least sensitive case — closest to traditional voice-over — and the disclosure expectation is mild.
Cloned voice with the speaker's consent. A presenter's own voice reproduced in languages they do not speak. The identity being used belongs to the person using it. This is the central case for creator and corporate localization, and audience reaction is generally positive when it is explained and negative when it is discovered.
Cloned voice of a third party with consent and licence. An actor or voice professional licensing their voice for synthetic use. Consent exists but the audience has no way to know it does.
Cloned voice without consent. Not a disclosure question. It is an impersonation and rights problem, and disclosure does not cure it.
Synthetic voice attributed to a real person who did not say the words. Putting statements into someone's mouth. This is the case regulators are most focused on, and it is where the strictest obligations sit regardless of consent.
The line that matters most in practice is between synthesis that changes how something is said and synthesis that changes who is understood to have said it. Dubbing a presenter's own script into Spanish with their cloned voice does the former. Generating a statement the presenter never made does the latter.
What Platforms and Regulators Now Expect
The landscape varies by jurisdiction and changes, so treat specifics as things to verify rather than as settled facts. The direction of travel is consistent.
Major video and social platforms require disclosure of realistic synthetic media in their upload flows, typically as a checkbox that produces a label on the published item. Policies generally focus on content that could mislead about real events or real people rather than on all AI involvement, and enforcement includes automatic labelling where detection systems flag content the uploader did not declare.
The European Union's AI Act includes transparency obligations for AI systems that generate or manipulate content, requiring that synthetic content be disclosed in a machine-readable way and that deep fakes be labelled. Compliance timelines are phased.
Several jurisdictions regulate synthetic media in political advertising specifically, with disclosure requirements and in some cases prohibitions near elections.
Advertising standards bodies in multiple markets treat undisclosed synthetic endorsement as misleading advertising under existing rules, without needing new AI-specific regulation.
Voice performer agreements increasingly include explicit terms on synthetic replication, and collective agreements in the performance industries have made consent and compensation for voice cloning a standard negotiating point.
The practical reading: if the content could reasonably lead someone to believe a specific real person said something they did not, disclosure obligations are likely to apply somewhere. If it is generic narration, they mostly do not, but platform-level declaration may still be requested.
Where the Label Goes
There are four places disclosure can live, and they serve different purposes. Most programmes should use several.
In the metadata. Platform upload fields, structured metadata in the file, and content credentials embedded in the asset. This is machine-readable, survives redistribution reasonably well where the format supports it, and is what regulatory transparency obligations increasingly target. It is also invisible to the viewer, so it cannot be the only measure.
In the description. A line in the video description or show notes explaining that the audio is an AI-generated translation of the original. Cheap, easy to keep accurate, and read by the fraction of the audience that reads descriptions.
On screen. A brief card at the start or a persistent corner mark. Unmissable, and correspondingly intrusive. Appropriate where the stakes are high — political content, health claims, anything attributed to a named individual — and heavy-handed for a localized product tutorial.
In the audio. A spoken line at the start. The most robust against redistribution, since it survives clipping and reposting, and the most disruptive to the content. Reserved for high-stakes cases.
The general principle is to match the prominence of the disclosure to the risk of the content being misunderstood. Over-labelling is not free: a persistent on-screen mark on every localized tutorial trains audiences to ignore labels, which weakens them where they matter.
Wording That Works
Disclosure language that helps rather than alarms shares a few properties.
Say what was done, not that "AI was used." "This video has been dubbed into Spanish using an AI recreation of the presenter's voice" tells the viewer something. "AI-generated content" tells them almost nothing and reads as a warning label.
Name the consent where it exists. "Recreated with the speaker's permission" answers the question that most viewers actually have.
Distinguish translation from fabrication. Making clear that the words are the speaker's own, translated, addresses the concern that matters. Audiences are largely comfortable with translation and uncomfortable with invention.
Keep it short and put it where it will be seen. A paragraph buried at the end of a description satisfies nobody.
Localize the disclosure. A Spanish-dubbed video with an English disclosure line is a common oversight and defeats the purpose.
Keep it consistent. A standard form of words used across the catalogue is easier to maintain, easier to update when rules change, and reads as policy rather than as an afterthought.
Detection and Provenance
Disclosure is a claim made by the publisher. Provenance is evidence attached to the asset.
Content provenance standards — the C2PA specification and the Content Credentials implementations built on it — attach signed metadata recording how an asset was produced and modified. Support is growing across capture devices, editing tools, generation services, and some platforms.
For a localization programme, provenance is worth attention for a reason beyond compliance: it lets you prove your own content is what you say it is. As synthetic media becomes ubiquitous, the ability to demonstrate that a given clip genuinely came from your organisation is a defence against misattribution.
The limitations are real. Provenance metadata is stripped by many processing steps and by platforms that re-encode. Detection systems that attempt to identify synthetic audio without provenance data are unreliable in both directions — they miss synthetic content and flag genuine content, and audio is harder than images. Watermarking schemes embedded in generated audio are more robust than statistical detection but are not universal and can be degraded.
None of this is a reason to skip provenance. It is a reason not to rely on downstream detection as a substitute for disclosing up front.
Building It Into the Workflow
Disclosure fails when it depends on someone remembering. The fixes are structural.
Record synthesis status as an asset property. Every output carries a flag: which voices were synthetic, whether any were cloned, and whose voice was cloned. This is generated by the pipeline, not entered by hand.
Store consent records against voice assets. Every cloned voice has a linked consent artefact with scope and duration. A voice without one cannot be used by the pipeline.
Attach provenance metadata at render time, automatically.
Template the disclosure text per language, and populate it from the asset properties so the description text cannot drift from what was actually done.
Set the platform declaration in the publishing step, not as a manual checkbox someone ticks.
Review the policy periodically. Platform rules and regulatory requirements in this area change on a timescale of months, and a disclosure policy written eighteen months ago is likely to be out of date.
A Working Checklist
- Distinguish generic synthetic narration, consented voice cloning, and attributed statements — the obligations differ.
- Never treat disclosure as a remedy for missing consent; they are separate requirements.
- Obtain and store written consent for every cloned voice, with defined scope and duration.
- Declare synthetic content in platform upload fields where the platform asks.
- Attach content provenance metadata automatically at render time.
- Put a plain-language disclosure line in the description for cloned-voice content.
- Escalate to on-screen or in-audio disclosure for political, health, and attributed-statement content.
- Say what was done and who consented, rather than labelling content generically as AI.
- Make clear that translated content carries the speaker's own words.
- Localize disclosure text into the language of the dub.
- Use one consistent form of words across the catalogue.
- Generate disclosure from recorded asset properties rather than manual entry.
- Block the pipeline from using any voice asset lacking a consent record.
- Do not rely on downstream detection as a substitute for disclosing up front.
- Review platform and regulatory requirements at least twice a year.
Frequently Asked Questions
Do I have to disclose AI dubbing of my own voice?
Legal obligations vary by jurisdiction and mostly target content that could mislead about real people or events; translating your own words into another language with your own cloned voice is at the mild end. Platform policies may still ask you to declare synthetic media at upload. Beyond compliance, disclosure is worth doing for audience trust: reaction is consistently better when it is explained than when it is discovered.
What is the difference between disclosure and consent?
Consent is permission from the person whose voice is being replicated. Disclosure is telling the audience what they are hearing. They are independent requirements, and disclosure does not cure a consent problem — labelling an unauthorised voice clone as AI-generated does not make using it acceptable. Handle consent as a contractual matter with stored records and disclosure as a publishing matter.
Where should the disclosure appear?
Match prominence to risk. Metadata and platform declaration for everything, a description line for cloned-voice content, on-screen or spoken disclosure for political, health, and anything presented as a statement by a named individual. Over-labelling is not harmless — a persistent mark on routine localized content trains audiences to ignore labels where they matter.
Can synthetic voice be detected automatically?
Not reliably. Statistical detection of synthetic audio produces both false negatives and false positives, and audio is harder than images. Embedded watermarking is more robust but is not universal and can degrade through re-encoding. Provenance metadata is the most dependable signal where it survives, but many processing steps strip it. Detection is not a substitute for disclosing at publication.
What should the disclosure actually say?
Something specific. "This video has been dubbed into German using an AI recreation of the presenter's voice, with their permission. The words are the presenter's own, translated." That answers the questions viewers have. A generic "contains AI-generated content" label answers none of them and reads as a warning.
How do I keep disclosure from being forgotten?
Generate it. Record synthesis and cloning status as properties of each rendered asset, template the disclosure text per language, and populate it from those properties in the publishing step. Anything that depends on a person remembering to tick a box will be missed, and the miss will be discovered by someone else.
Related reading: Ethical Voice Cloning | Clone Your Voice for Multilingual Videos | Video Translation Security and Privacy
