New EU transparency rules landed in August 2026. Here is what a firm actually needs to tell listeners when AI is involved in making its podcast.
By Will Nash
7 September 2026
Someone at your firm has probably asked it already, most likely in a meeting about the podcast: if we use AI to help make this, do we have to tell people? Nobody in the room knows, and the honest answer comes in two parts that tend to get tangled together.
There is what the law requires, which is narrower than most people assume. And there is what your audience will forgive, which is wider. The second one is far more likely to cost you.
What changed in August 2026
Article 50 of the EU AI Act became applicable on 2 August 2026. It places transparency duties on the organisations that build AI systems generating synthetic audio, images or video, and on the organisations that deploy them.
The part that matters for a podcast is the deep fake provision. Where AI-generated or manipulated audio resembles a real person and would falsely appear to be authentic, whoever deploys it has to disclose that it was artificially generated. That disclosure has to be clear, and it has to arrive no later than the first time someone is exposed to the content.
There is a carve-out worth knowing about. The obligation is relaxed where the content has been through human review or editorial control and a named person or organisation holds editorial responsibility for publishing it. For a firm that signs off every episode before it goes out, that is a meaningful distinction.
Separately, the providers of these systems have to mark their outputs in a machine-readable format. That is a duty on the tool rather than on you, but it is worth asking your supplier whether they do it.
Sitting outside the EU does not automatically put you outside this. The Act can reach organisations whose output is used inside the EU, so a UK firm with EU listeners should check rather than assume. What you are reading is a short summary of a long piece of legislation, not advice on your particular situation.
Whether it applies turns on what you are actually doing
Cloning a named partner's voice, so that a listener would take the audio for a recording of that partner, is the case these rules were written for. A conversation performed by synthetic voices that are not impersonating any specific real person sits somewhere else, and the analysis is different.
The other variable is editorial control. An automated feed that publishes without anyone reading it is a different proposition from an episode a named person approved. Most firms are closer to the second than they realise, and they get no credit for it because they have never written it down anywhere.
The trust question is the one that will cost you
Set the law aside for a moment and ask a simpler question. What happens if a client works out, from something other than you telling them, that the voice they were listening to was synthetic?
Professional-services firms sell judgment, and the reputational damage rarely comes from using the technology. It comes from the appearance of having concealed it. Firms that say plainly what they did tend to get a shrug. Firms that get found out get a conversation they did not want to have.
How to say it without making it the story
Keep it short and factual, and put it somewhere a listener can find without hunting for it.
A line in the show description covers most of it. Something along the lines of: episodes are written from our own research and reviewed by a named person before publication, and the voices are produced with AI. A sentence at the top of the first episode does the rest.
There is now infrastructure for this. RSS.com added an AI disclosure feature to the open RSS feed in early 2026 and Spreaker followed shortly after, so the disclosure can travel with the feed rather than living only on your website.
The useful move is to describe the human work alongside the machine's. Saying "AI-generated" tells a listener nothing about whether anyone at your firm stood behind the episode. Naming who researched it and who approved it tells them the thing they actually want to know.
Where this gets harder
Some situations need more than a line in the show notes. Regulated communications may need compliance sign-off on the disclosure wording itself, and if your episodes touch client matters then confidentiality gets there long before AI disclosure does.
If your audience is one that will reject synthetic audio outright, a disclosure line will not change their mind, and you are better off making a different kind of show.
The hardest case is a firm that cannot name anyone who takes responsibility for what goes out. Disclosure does not fix that. Naming an owner does, and it is worth doing whatever the law asks of you.
In short
The legal question is narrower than the corridor version suggests, and it turns on whether the audio impersonates a real person and on who holds editorial responsibility. The trust question is wider, and the answer to it is straightforward: say what you did, name who stood behind it, and get on with the episode. Firms that handle this early never have to have the awkward version of the conversation later.
We are an AI-native podcast agency, and we settle disclosure wording with clients at the start rather than leaving it to the end. Every episode has a named person at the firm who signs it off before it publishes, which is the part that makes a disclosure line worth anything. If you want help working out what yours should say and where it should sit, that is a straightforward conversation. Get in touch.
Related: What Is Synthetic Audio?
