Sanremo puts AI on the blacklist, whilst the industry signs it up

The rules for Sanremo 2027, presented to the press on Monday, include a new provision: songs must be the result and expression of human creativity, and the use of artificial intelligence systems that replace, alter or distort the author’s creative contribution is not permitted. The ban covers the conception, composition and production of the songs. To verify compliance, the regulations grant Rai and the artistic director, Stefano De Martino, the authority to use any available technological means, without specifying which ones and without providing for an appeals procedure, as is the case with the requirement for songs to be original.

Before even considering how to verify this, however, it is worth establishing what is prohibited. The distinction that Enzo Mazza, CEO of FIMI (Italian Music Industry Federation), considers crucial at the Ariston is between creative use and what he calls ‘black hat’ – the adoption of unethical practices. “The rules do not ban AI outright, but state that it must not replace human creativity,” emphasises the head of the Confindustria association representing the leading record labels and the music industry in Italy. “Many artists already use artificial intelligence today, because it is simply another tool that aids creativity.” On the other hand, there is the industrial phenomenon: “Songs created entirely by AI that are flooding the platforms, either for personal gain or to defraud the system of royalties. And then there are cases of songs created using artificial intelligence and listened to by bots”, in order to manipulate streaming figures and climb the charts.

But does a model that cleans up a vocal track, generates a demo arrangement or suggests a harmonic progression fall under the ban? Where does the line between legitimate and illegitimate use of AI lie? “It hinges on the fact that it must be certain that whoever has used artificial intelligence as a complement to creativity has done so on legitimate platforms,” argues Mazza. “That is to say, authorised platforms that hold a licence, which are known to have used negotiated content and which grant rights to the artists. And then it must be the author themselves who declares its use: there is an issue of transparency towards the fans. If I were an artist, I’d want to say, ‘Look, I used artificial intelligence in this part’, rather than risk being caught out with something inauthentic.”

There is already a precedent in Italy. On 12 August, FIMI adopted the principles governing the eligibility of tracks developed using generative AI for the official ‘Top of the Music’ charts, following a common approach outlined by Sony, Universal and Warner. There are three criteria: the AI service used must be authorised and lawful; the track must be substantially created by a human being; and there must be no suspicion of manipulation of streaming figures. Added to these is the obligation to inform consumers of the use of AI on streaming platforms. The rule has already claimed its first victim: a track generated entirely by AI has been suspended from the Niq/Fimi charts, and the three radio stations in the RTL 102.5 group have removed it from their playlists. And Rai’s initiative, explains Mazza, responds to a request that came from there: “We had already highlighted this issue, because we already apply these principles.”

The most uncertain aspect remains that of verification. Some detection tools do exist, but they do not always guarantee a reliable analysis. Classifiers that analyse the spectrogram, such as Ircam Amplify or ACRCloud, are good at recognising a track fresh out of a generator, but they falter when the track has been post-produced, mixed with live-played parts and mastered. This is precisely the state of every demo that arrives at the Ariston. Some AI systems, such as Suno and Udio, embed inaudible signals in their output – watermarks that can, however, be removed and which, in any case, reveal nothing about a track created on other platforms. There remains the option of requesting process documentation from the artists: the stems (separate audio files for the various instrument groups), session files and dated demo recordings.… It is a somewhat laborious, forensic-style process. Furthermore, if the ban also covers the conceptualisation and composition of a track, what should be done about a songwriter who uses a language model to suggest lyrics and then sings them? “Detection systems clearly have their limitations,” admits Mazza, who places the onus of verification primarily at the early stages. “During the selection phase, should any suspicions arise, these tools can be used to carry out a check. The aim at this stage is primarily deterrence and raising awareness. We’re not in a phase of policing, of hunting down offenders.” And if suspicion arises later, reputation takes care of the rest: the track doesn’t make it into the charts and, he adds, “it’s clear you’ve made a complete fool of yourself”.

If an accusation were to arise during the Festival – perhaps on social media – how would an author who had been wrongly accused defend themselves? “The procedure is more or less the same as that used for cases of plagiarism, namely to investigate the matter. It’s not as though a song is removed from the Festival simply on the basis of a report received via social media. A request for information will be made to the person who submitted the song and to the record label, to clarify the matter, together with technical experts and record industry representatives. This, after all, has already happened in the history of Sanremo: we have had hundreds of cases of plagiarism reported by those whose job it is to spot it. The verification procedures, for better or worse, will be the same,” emphasises Mazza.

The fact that licensing is the key issue is demonstrated by what happened a few days before Rai published the festival regulations. Whilst Universal and Sony are continuing their legal action against Suno, other major labels are striking deals with the generative music creation platform: on 9 September, Suno launched v6, the first family of models trained on licensed music, developed in collaboration with Warner Music Group, BMG and Believe. This is the result of the settlement agreement under which, in November 2025, Warner settled its case against Suno, allowing artists and songwriters to opt in voluntarily and receive remuneration. Believe, which had blocked the distribution of tracks created with Suno in April, eventually signed the agreement. The following day, 10 September, Universal Music announced a multi-year licensing agreement with ElevenLabs – the artificial intelligence company’s first such deal with a major label – for a music creation platform based on licensed music and artist participation, featuring remixes, mashups and personalised vocals. This is the comparison that is closest to Mazza’s heart: the industry emerged from piracy when streaming platforms arrived. “It is not a question of rejecting the technologies, but of adopting them in accordance with rules that ensure the entire supply chain is satisfied with their use – and, above all, remunerated.”

Which brings us back to that line in the regulations. ‘Every available technological means’ sounds like a laboratory promise, but the tools that will actually decide the outcome in February are the very same ones the Festival has been using to deal with suspected plagiarism for half a century: a request for clarification from the author and the record label, a team of technicians listening to the tracks, and the fear of making a fool of oneself in front of twelve million viewers. Faced with the most modern technology available, Sanremo defends itself with its word of honour and its reputation.