For a few years the question of whether to tell listeners you used an AI voice was a matter of taste. Some shows disclosed, most did not, and nobody was checking.
That changed on 2 August 2026, at least in Europe, when the transparency obligations in Article 50 of the EU AI Act became applicable (European Commission FAQ). Platform rules had already moved earlier. So the practical question for anyone producing audio is no longer whether disclosure is nice, but where exactly the line sits between a production tool that needs no announcement and content that has to carry a label.
I want to be upfront that I am not a lawyer and none of this is legal advice. If you publish into the EU at any scale, talk to someone qualified. What I can do is lay out the shape of the rules and, more usefully, a way of thinking about the line that holds up regardless of jurisdiction.
The distinction that does most of the work
Almost every disclosure framework, legal or platform, turns on the same question, phrased differently each time:
Could a reasonable listener mistake this for a real person actually saying this?
That single test sorts most cases correctly.
An AI voice reading your article in a voice that is obviously a stock synthetic narrator, on a show where that is the format, is not deceiving anyone about a real person. A cloned version of your co-host's voice reading lines they never recorded is a different thing entirely, even if the co-host consented, because the listener hears a specific human being saying specific words.
Everything else is elaboration on that distinction.
What the EU AI Act actually requires
Article 50 sets transparency obligations across several categories. Two matter for audio.
The first applies to deployers, meaning the people using the system to make content. If you use AI to generate or manipulate audio that resembles existing persons and would falsely appear authentic, you have to disclose that it has been artificially generated or manipulated. The Commission's guidance is notably broad here: content that sounds like a real person needs labelling even where there was no intent to deceive (Commission guidelines).
The second applies to providers, meaning the companies building the generation systems. They carry an obligation around marking outputs in a machine-readable way so that generated content can be detected. There is a limited grace period on that particular obligation for systems already on the market before 2 August 2026, running to 2 December 2026. A readable practitioner summary sits at artificialintelligenceact.eu.
There is a further provision covering AI-generated text published to inform the public on matters of public interest. If your show publishes AI-drafted written material about news or public affairs, that is worth reading closely with counsel rather than taking my summary for it.
The part I would emphasise to producers: the obligation attaches to the resemblance and the potential to appear authentic, not to your intentions. A well-meaning use that produces audio sounding like a real person is still in scope.
What platforms already required
Platform policy moved before the law did, and if you publish video versions of your show it has been binding on you for a while.
YouTube requires creators to disclose realistic altered or synthetic content, using a toggle at upload that produces an "Altered or synthetic content" label. The policy is explicitly scoped: it applies where content is realistic enough that a viewer could mistake it for a real person, place, or event. It does not apply to clearly unrealistic or animated content, to standard special effects, or to using generative AI for production assistance (YouTube announcement, help centre).
That production assistance carve-out is the most useful thing in the policy, and it maps closely to the reasonable-listener test above. Using AI to clean up audio, generate a script outline, or write show notes is not synthetic content in the sense the policy cares about. Generating a voice that sounds like a person is.
Other platforms have their own versions, and podcast directories are less consistent than video platforms. Check the terms wherever you actually distribute rather than assuming a single standard.
A working line for producers
Here is how I would sort typical podcast uses. This is my reading, not a legal classification.
Clearly needs disclosure:
A cloned voice of a real person, including your own, saying words they did not record. Consent does not remove the obligation. The listener is the one being informed, not the voice owner.
A synthetic host or narrator presented as a person, with a name and a persona, where nothing on the show tells you otherwise.
AI-generated audio of a public figure, in any framing where it could be taken as authentic. Satire and commentary may be treated differently in some regimes but this is exactly the territory to get advice on rather than guess.
Reconstructed audio in a documentary or true crime context, where a voice performs words from a transcript. This is a strong candidate for disclosure in the episode itself, not just in a description field, because it is precisely the case where a listener will otherwise assume they are hearing a recording.
Generally does not need disclosure, but disclose anyway if it is central to the show:
AI translation or dubbing into another language, where the show is openly a translated edition. Worth flagging that if the dub uses a clone of the original host's voice, you are back in the first category.
An obviously synthetic narrator on a show whose format is transparently automated, like an AI-read article feed. The format itself discloses. A line in the show description costs nothing and removes ambiguity.
Does not need disclosure:
Noise reduction, de-reverb, level matching, filler word removal, and the rest of ordinary post-production, including the machine learning based versions of those tools.
AI-assisted scripting, research, outlining, or show notes, where a human performs the audio.
Transcription and chaptering.
The principle: tools that shape a human performance are production. Tools that generate a performance are content.
How to actually disclose
Getting this wrong is easy in a way that is worth spelling out, because a lot of shows technically disclose and functionally do not.
Put it where the listener is. A note in the RSS description does not reach someone listening in a car. If synthetic voice is material to what the audience is hearing, say it in the audio. A single sentence at the top of the episode is enough.
Say what, not just that. "This episode contains AI-generated audio" leaves a listener wondering which parts, and that uncertainty tends to contaminate their trust in everything else. "The reconstructed 911 call in part two was generated from the official transcript using an AI voice. No recording of the original call exists publicly" is specific, and specificity is what preserves trust rather than eroding it.
Repeat it at the moment. For reconstructions inside a longer episode, a brief line right before the segment does more than a blanket statement at the top that a listener has forgotten by minute 34.
Do not bury it in a legal-sounding block. Disclosure that reads like terms and conditions gets skipped. Disclosure in the show's own voice gets heard.
Set the platform flags too. If you distribute video, use the toggle. In-audio disclosure and platform labelling are not substitutes for each other.
The consent question, which is separate
Legal disclosure obligations and permission to use someone's voice are different problems, and it is worth not conflating them.
If you clone a voice, you need the rights to that voice. This is contract and personality rights territory, it varies significantly by jurisdiction, and several places have introduced or strengthened specific voice and likeness protections in the last couple of years.
Two things I would build into any production that clones a voice, regardless of what the law where you are strictly requires:
Written permission that is specific about scope. Which voice, for what use, for how long, and with what happens when the relationship ends. A blanket grant to use someone's cloned voice indefinitely is going to become a problem for somebody eventually.
A deletion path. If a contributor withdraws consent, can you actually delete the voice model, and can you demonstrate you did? Ask your vendor this before you need the answer.
For a deceased person's voice, get the estate involved and expect that the ethical bar is higher than the legal one.
Why disclosure is also the commercially sensible choice
The argument I would make to a producer who finds all of this tiresome.
Detection tools exist, they are getting better, and your audience has access to them. Anyone can run a clip through a voice detector and get a signal. That signal is imperfect, and I have written at length about why detectors disagree, but imperfect is more than enough to start a conversation you did not want to have on social media.
The failure mode is not that someone proves you used AI. It is that a listener gets a plausible-looking result, posts it, and you are now defending yourself in a context where the accusation is short and the explanation is long. Disclosure removes that entire scenario by making the finding boring.
There is a second effect I find more interesting. As synthetic audio becomes ubiquitous, "this is a real recording of a real person" becomes information worth stating. Shows that are clear about what is real are building something that gets more valuable as the ambient uncertainty rises. Disclosure is not just a defensive obligation, it is a way of making an authenticity claim that will actually be believed.
If you already published without disclosing
Plenty of shows have back catalogues containing synthetic audio produced when none of this was expected of them. That is a normal position to be in and it is fixable.
Do not quietly re-upload. Replacing an episode file with a disclosed version and saying nothing is the version of this that goes badly, because the change is discoverable and the silence is what makes it a story. Podcast feeds are archived, and someone always has the old file.
Work out what you actually did. Before announcing anything, audit. Which episodes, which segments, which voices, whether anything was cloned from a real person. Producers are often surprised in both directions here, and a correction that turns out to be incomplete is worse than the original omission.
Fix the description first, the audio when you can. Updating show notes and episode descriptions is cheap and immediate. Re-cutting audio to add an in-episode line is worth doing for the cases that matter most, particularly reconstructions and cloned voices, and may not be worth it for an obviously synthetic narrator on a format show.
Say it once, plainly, in the current episode. A short segment explaining what you used, where, and what your policy is going forward does more for trust than a written statement nobody reads. It also gets you the benefit of having raised it yourself.
Write the policy down and publish it. A page on your site describing how you use AI, updated when that changes, converts a recurring judgement call into a standing answer. It is also the thing you point at when somebody asks.
The tone that works is matter-of-fact. Not a confession, not a defence. You used a production technique, norms and rules changed, here is where it appears and here is what you do now.
A short checklist
Before you publish an episode containing any synthetic audio:
Does anything in this episode sound like a specific real person saying something they did not say? If yes, disclose in the audio.
Is any voice on this show generated rather than recorded, in a way a listener would not assume? If yes, disclose somewhere the listener will actually encounter it.
If you cloned a voice, do you have written, scoped permission, and a way to delete the model?
If you distribute video, have you set the platform disclosure flag?
If you publish into the EU, have you looked at Article 50 with someone qualified rather than with a blog post?
Would you be comfortable if a listener discovered exactly how this episode was made, from someone other than you?
That last question has always been the real test. The regulation has just made a subset of it mandatory.
Try it on your own writing