Is There AI In This Audio? What HUZIP Asks and What the Law Asks

On 4 September the Croatian performers’ society told its members that when they register a recording they must now declare whether any part of it was generated by AI, and pointed at the EU AI Act as the reason. The two are not the same thing, and the difference matters a great deal if you are the one commissioning the audio. Here is what the Regulation actually says, what it does not say, and the one page we now send out with every job.

The notice went out on 4 September 2026. HUZIP — the society that collects performers’ royalties in Croatia — has added a mandatory field to its registration form. When a performer registers a recorded performance, they now have to state the mode of participation, and one of the options is “AI generated”. If any role or any instrument was generated, it has to be flagged as such. The sanctions named are HUZIP’s own: payment blocked on the disputed recording, proceedings against whoever filed it, and a claim for the return of money already paid out.

The notice cites Article 50 of Regulation (EU) 2024/1689, the AI Act, and says its provisions came into full application on 2 August 2026.

If you run production at an agency, your first question is presumably whether this lands on you. The short answer: HUZIP’s requirement does not, and the AI Act might, but far more narrowly than the notice suggests.

What Article 50 actually says

It creates two duties for two different groups.

The first is on providers — the people who build the tool. “Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated.” That is a watermark, and it is the tool vendor’s job. Not yours, not ours.

The second is on deployers — whoever uses such a system in a professional capacity. “Deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake, shall disclose that the content has been artificially generated or manipulated.”

Read that sentence twice, because of what it does not say. It does not say “any content made with AI”. It says deep fake, which the Regulation defines as content resembling existing persons, objects, places, entities or events, which would falsely appear to someone to be authentic.

There is also a lighter rule for creative work. Where synthetic content forms part of “an evidently artistic, creative, satirical, fictional or analogous work or programme”, the disclosure only has to be made “in an appropriate manner that does not hamper the display or enjoyment of the work”. The Commission’s guidelines of 20 July 2026 read the word evidently strictly, and take the view that content which is exclusively informative or commercial does not qualify. An advertisement is not automatically an artistic work.

And nowhere in Article 50 is there an obligation on a performer, or on a phonogram producer, to declare anything at all. The Article addresses providers and deployers. That is the whole list.

So why is HUZIP asking?

Because they have a different problem, and a legitimate one. HUZIP’s job is to decide who gets paid for what. A synthetic performance has no performer behind it, and money set aside for performers should not be paid out on a part that nobody played. Their requirement rests on their own distribution rules and on the Croatian Copyright Act, where a performer is defined as a person. It is real and it binds their members.

It is simply not Article 50, and three details in the notice are worth keeping straight. The Regulation does not oblige all authors, performers and phonogram producers to mark anything. The machine-readable marking is the tool provider’s duty, not the filer’s. And “full application” on 2 August is not quite right — parts of the Act apply on later dates.

None of that makes HUZIP wrong to ask. It makes the mail imprecise about why.

What it means for the audio in a spot

Three cases come up in commercial work, and they are not equivalent.

A cloned voice of a real person is the clear one. That resembles an existing person and would plausibly be taken as authentic, so it is a deep fake in the sense the Regulation uses. Consent from the person whose voice it is, is a separate problem and an older one.

A generic synthetic voice that imitates nobody in particular sits in a grey area. The guidelines discuss realistic AI-generated personas, and the honest answer today is that there is no case law. In our own statements we treat a realistic synthetic voice as if it were in scope and say so plainly, because being over-declared has never cost anyone anything.

A fully generated music bed that resembles no specific artist is probably not a deep fake at all. If it deliberately imitates a recognisable artist, the question reopens.

There is no deadline, and we are not going to invent one

As of today, according to the public implementation trackers and the ministry’s own pages, Croatia has not designated the market surveillance authority the Act requires, and no implementing law has appeared in Narodne novine. The penalties in Article 99 are real on paper — up to 15 million euros or 3% of worldwide annual turnover, with lower figures for smaller companies — but at the time of writing there is nobody here to impose them.

So this is not urgent, and anyone telling you otherwise is selling something.

The part that is actually worth doing

What is true is duller and more useful. The question “what is in this audio?” costs about five minutes to answer on the day of delivery and is close to unanswerable a year later.

Who sang it. Which day, in which room. Whether the guitar was played or generated. Whether anyone’s voice was cloned. Whether the file that arrived from the sound designer carried any provenance metadata, and whether it survived the bounce. None of that lives in a WAV file, and none of it is in anybody’s head eighteen months later when a performer, a client, or a lawyer asks.

That is the whole case for writing it down. Not the fine. The reconstruction.

What now goes out with every delivery

One A4 page with every commercial job. It lists the project and the files, who performed and on what day, whether a generative system was used and for what, which processing tools were used, whether any real person’s voice was cloned, and who decided that a tool would be used. It is signed by the studio.

It records facts, not conclusions. It does not say who the deployer is in your chain, because that depends on who decided to use the tool, and that is between you and your client. We write down what we know: that the decision came with the brief, or that it was ours.

We list our processing tools on it too — iZotope RX, oeksound soothe and bloom, stem separation — because “no AI was involved” would not be a careful thing to say, and “AI was used” would be misleading. Those tools repair, separate and balance what was performed. They do not generate anything, and the Regulation’s marking duty is aimed at systems that do.

For most of what leaves this room the line that matters reads: no generative AI was used, and here are the people who played.

It costs nothing and it is part of delivery. For work we delivered earlier, we can reconstruct one from the session documentation, because we were already keeping it.

What this is not

This is not legal advice, and the statement is not a legal document. Neither tells you what your obligations are. If you need that answered for a specific campaign, it is a question for your lawyer, and the useful thing you can hand them is a record of what was actually in the audio.

The sources, if you want to read them rather than take our word for it: Regulation (EU) 2024/1689 (Article 50 is the relevant one), the Commission’s guidelines on transparency obligations of 20 July 2026, the Code of Practice on transparency of AI-generated content, and HUZIP’s notice itself.

If you would like the blank form or a filled example, write to us and we will send both. And if you have an AI track that needs to become a real recording rather than a declaration, we wrote about that last month.