Italy wrote "human authorship" into law: what its AI act means for producers using AI

Italy has put the word "human" into its copyright law. Law No. 132 of 23 Sep 2025, the country's general law on artificial intelligence, was published in the Gazzetta Ufficiale on 25 Sep and came into force on 10 Oct 2025. The law is designed to help apply the EU AI Act, but for musicians the most useful parts are three short changes in Articles 25 and 26: who counts as an author, when AI can mine your work, and what happens to someone who spreads a fake of your voice.

What exactly did Italy change in its copyright law?

Article 25 edits Article 1 of Italy's 1941 copyright law (Law 633/1941) in two places. It inserts "umano" (human) after "opere dell'ingegno" (works of intellect), so the law now protects creative works of human intellect. And it adds that this covers works "even where created with the aid of artificial intelligence tools, provided they are the result of the author's intellectual work" (our translation of the official text). The amended article still names music among the protected fields, alongside literature, the visual arts, architecture, theatre and cinema.

What the law doesn't do is define how much human work is enough. Ernesto Apa and Chiara Marchisotti of Portolano Cavallo, writing for the European Audiovisual Observatory, expect the standard to develop through case law as courts distinguish between levels of human involvement. The ICT Legal Consulting authors of a Chambers practice guide read it as requiring "a sufficient human creative contribution". That is their reading, not the statute's wording.

Does the law cover cloned voices?

Yes, by name. Article 26 adds a new Article 612-quater to the Italian Criminal Code. In our translation, it punishes anyone who causes unjust harm to a person by handing over, publishing or otherwise spreading, without that person's consent, images, videos or voices falsified or altered with AI systems and capable of misleading people about whether they are genuine. The penalty is one to five years in prison.

The details matter for an artist. The offence is spreading the fake, not making it. All four conditions have to be met: dissemination, no consent, a convincing AI fake, and unjust harm. And it is prosecuted only on a complaint from the person harmed, except where it is connected to another offence prosecuted automatically, or where the victim cannot act because of age or infirmity, or is a public authority targeted for their role. In practice, a singer whose cloned vocal turns up on a release under their name has to file the complaint themselves. Whether a clearly labelled AI "tribute" is "capable of misleading" is the kind of question the courts will have to settle.

Can AI companies still train on Italian music?

Within limits. The new Article 70-septies allows reproductions and extractions from works lawfully accessible online or in databases for text and data mining through AI models and systems, "including generative" ones, in line with Articles 70-ter and 70-quater. The general exception in Article 70-quater applies only where the use has not been "expressly reserved" by rights holders. In other words, the opt-out survives.

Article 26 also adds a new letter a-ter to Article 171 of the copyright law, making it an offence to mine text or data in breach of those two articles, "including through artificial intelligence systems". Here is the detail the headlines skip: that paragraph of Article 171 sets a fine of €51 to €2,065. "Criminal sanctions for unlawful mining" is accurate. But on its face the basic penalty is a small number next to the cost of training a model, and it only bites if the rights holder reserved their works in the first place.

How does this compare with the US position?

The outcome on paper is close. In Part 2 of its AI report, published on 29 Jan 2025, the U.S. Copyright Office said that using AI to assist creation doesn't bar copyright, that protection applies only where a human author has determined sufficient expressive elements, and that "the mere provision of prompts" isn't enough. It added that a human-authored work perceptible in an AI output, or creative arrangements or modifications of the output, can be protected. The Office concluded that "the case has not been made" for changing the law.

The difference is the instrument. The US position is agency guidance under existing law. Italy's is statute, which judges must apply. Under the Berne Convention, which Italy ratified, protection can't depend on any formality, so nobody checks your human contribution when you finish a track. Our reading: in Italy, the question is most likely to come up later, in a dispute, when you have to show what you did.

What should producers who use AI actually do?

Treat your process as evidence. Concretely:

• Keep dated drafts and session files, not just the final bounce. Project files, alternate takes, stems (separate tracks for vocals, drums and so on) and lyric drafts show decisions being made over time.
• Note what the AI generated and what you changed. A short log per song ("AI drum pattern, re-programmed bars 9 to 16; melody and lyrics mine") is easy to keep and hard to reconstruct later.
• Don't rely on prompts as proof. The US Office says prompts alone aren't enough, and nothing in the Italian text suggests a prompt counts as "intellectual work" by itself.
• Get consent in writing before training a model on, or releasing, anyone else's voice, including a collaborator's.

Working in a tool where AI assists inside a project you control, such as Veena, makes the first two habits easier, because your edits, takes and arrangement choices stay in the session rather than disappearing into a one-shot prompt.

What we think

Italy's law is more useful to artists than a lot of AI regulation, and less protective than its headlines suggest.

The human-authorship rule cuts in musicians' favour. A catalogue generated end to end has, on this text, a weak claim to copyright in Italy, which undermines the business of flooding platforms with machine-made tracks and treating them as owned assets. Producers who use AI as one instrument among many are not shut out; the law says so explicitly. The winners are people who can show their working. The losers are anyone whose "work" is a prompt and a download button.

The enforcement side puts the load on artists. The deepfake offence needs harm and a complaint from the victim, which favours performers with lawyers over the session singer whose voice is cloned for a cheap cover. The mining offence depends on rights holders having reserved their works, and carries a basic fine that a well-funded AI company won't notice. None of this is unfair in principle, but it means the protections are only as strong as an artist's paperwork and willingness to act.

The second-order effect we expect is quieter: contracts. Once human authorship is a legal condition, labels, publishers and sync buyers have a reason to ask for warranties about how a track was made, and to want the evidence behind them. Producers who start keeping records now will find that conversation easy. Those who don't may find their AI-assisted catalogue worth less than they thought, not because a court said so, but because nobody can prove otherwise.