The year the platforms started labelling AI music

The year the platforms started labelling AI music

By Rachel Cohen, Head of Music Curation, RadioPro LLC

For most of the streaming era the question "who made this recording" had an answer in the credits. In 2025 the answer started needing a second line. The two largest independent signals of that change came from Deezer, which built a detector and started counting, and from Spotify, which published a policy and backed a metadata standard. Between the two sat a band with a million monthly listeners that nobody could place. This article sets out what the platforms now do, what their numbers say, and where a performance model that renders human-written songs sits in the new scheme.

Deezer counted first

Deezer announced an in-house detection tool on 24 January 2025 and gave the first hard number: roughly 10,000 fully AI-generated tracks were being delivered to the platform every day, about 10 percent of daily deliveries. The company said it had filed two patent applications in December 2024, and its chief innovation officer explained why it had built a model-agnostic detector rather than one tuned to a single generator, noting that "the detection rate drastically decreases as soon as the tool is subjected to a new model or new data." The release also set the policy that still applies: detected content would be removed from algorithmic and editorial recommendation. (Deezer Newsroom, January 2025)

The series since then is the clearest public measure of the flood.

  • 16 April 2025: over 20,000 fully AI-generated tracks a day, over 18 percent of uploads. (Deezer Newsroom, April 2025)
  • 20 June 2025: 18 percent, and the launch of album tagging. Deezer said every album containing fully AI-generated tracks was now labelled, and that up to 70 percent of the streams on such tracks were fraudulent. Fully AI-generated music was about 0.5 percent of all streams. (Deezer Newsroom, June 2025)
  • 11 September 2025: over 30,000 tracks a day, more than 28 percent of daily delivery. (Deezer Newsroom, September 2025)
  • 12 November 2025: roughly 50,000 tracks a day, 34 percent of all daily deliveries. (Deezer Newsroom, November 2025)

Two features of this series matter more than the headline growth. The first is that consumption did not follow supply. Through 2025 Deezer kept reporting that fully AI-generated music was a small fraction of streams, around half a percent, and that most of those streams were manipulated. Deezer's chief executive, Alexis Lanternier, put the policy consequence plainly in September: "we remove fully AI-generated content from algorithmic recommendations and we don't include it in editorial playlists," so that "the impact on the royalty pool remains minimal." The second is that the category Deezer measures is narrow. The tool detects tracks that are 100 percent AI-generated. There is no partial category on Deezer. A recording with a human composition and an AI performance is simply not what the counter counts.

The November release added a listener study. Ipsos surveyed 9,000 people in eight countries for Deezer. Ninety-seven percent could not tell fully AI-generated music from human recordings. Eighty percent agreed that fully AI-generated music should be clearly labelled. Only 11 percent thought it should be treated equally on the charts. (Deezer Newsroom, November 2025) Those three numbers explain why labelling, not detection alone, became the industry's answer this year. Deezer also cited a CISAC and PMP Strategy estimate that nearly a quarter of creators' revenues are at risk by 2028.

The band that forced the question

Between those releases, a case arrived that made the abstract concrete. The Velvet Sundown released a debut album on 5 June 2025 and a second on 20 June, and by early July had roughly a million monthly listeners on Spotify. On 29 June the band's account insisted it "is not AI-generated in any capacity." Deezer's detector disagreed. Its profile carried the platform's label and the sentence "Some tracks on this album may have been created using artificial intelligence." (MusicTech, 2 July 2025; The Fader, 7 July 2025)

The story then took a turn that showed how thin verification had become. A person using the name Andrew Frelon presented himself as the band's spokesman, told Rolling Stone that some tracks had been generated with Suno, and then admitted he had no connection to the act, citing "numerous and large gaps in the verification process used by the majority of journalists." On 5 July the band itself posted a new description: "The Velvet Sundown is a synthetic music project guided by human creative direction," with music "composed, voiced, and visualized with the support of artificial intelligence." Spotify, which had neither labelled nor removed the music, told the German press agency dpa on 21 July: "As soon as a standardised approach is available, we will carefully examine how we can implement it." (dpa via Yahoo News, 21 July 2025)

That sentence is the hinge of the year. A platform with the largest audience in music declined to act on its own detection and asked for a standard. Two months later it had one.

Spotify's three measures

On 25 September 2025 Spotify published its AI policy under the title "Spotify Strengthens AI Protections." It has three parts. (Spotify Newsroom, 25 September 2025)

Impersonation. The policy states that "Vocal impersonation is only allowed in music on Spotify when the impersonated artist has authorized the usage." It also tightened the process for uploads that are fraudulently delivered to another artist's profile, with artists able to report a mismatch before release.

Spam. Spotify said it had "removed over 75 million spammy tracks" in the preceding twelve months and would roll out a filter that identifies uploaders and tracks engaged in "mass uploads, duplicates, SEO hacks, artificially short track abuse," tag them, and stop recommending them. The rollout was described as conservative and gradual.

Disclosure. Spotify committed to an industry standard for AI disclosures in credits, developed through DDEX, the body that maintains the music industry's release metadata formats. The disclosures cover "AI-generated vocals, instrumentation, or post-production," are submitted by labels and distributors, and will be displayed in the app as they arrive. Nineteen distribution and label services companies were named as working on it, including Believe, CD Baby, DistroKid, EMPIRE, FUGA and RouteNote.

The policy's framing is the part most relevant to anyone building or using a music model. Spotify wrote that "the use of AI tools is increasingly a spectrum, not a binary," that the aim is "not about punishing artists who use AI responsibly or down-ranking tracks," and that the company supports "artists' freedom to use AI creatively while actively combating its misuse by content farms and bad actors." Deezer labels a track. Spotify intends to label a contribution.

What the standard actually encodes

The DDEX Electronic Release Notification standard, in version 4.3.1, added two flags: "Communication of the fact that a recording was made (fully or in part) by a generative AI tool," and "Communication of the fact that a contribution was made, fully or in part, by a generative AI tool." (DDEX ERN Part 1 release notes) In practice distributors render those flags as three values for the recording, and the same three values against each credited contributor: all, partly, none.

This two-level design is the first industry mechanism that can describe a human-written song performed by a machine without lying in either direction. A composer and lyricist are people. A vocal and an instrumental performance can be machine-made. The standard lets a release say both.

YouTube reached a similar place earlier and from a different direction. Its November 2023 policy requires creators to disclose realistic synthetic content and gave music partners a route to request removal of AI music that "mimics an artist's unique singing or rapping voice." (YouTube Official Blog, 14 November 2023) The common thread across Deezer, Spotify and YouTube is that identity and provenance are now treated as separate questions: whose voice is this, and how was the recording made.

Where a performance model sits

RadioPro LLC builds AuralVerse, a text-prompted generative audio performance model. It takes lyrics and compositions written by people and produces performed audio. Under the schemes described above, an AuralVerse rendition is not a Deezer "fully AI-generated" track, because the composition and the words are human. Under role-level disclosure it is a recording in which the lyricist and composer are credited as people and the vocal and instrumental performances are credited as AI. We think that is the correct description, and it is the one we will supply wherever a platform asks for it.

Three consequences follow for the people we work with.

  1. Disclose at the role, not the track. The spectrum framing is the right one. A release that hides machine performance behind a human name invites exactly the mismatch reports and removals the platforms have built. A release that declares the performance as AI and the writing as human is on the right side of every policy published this year.
  2. Impersonation is the line that does not move. Every platform now treats an unauthorised voice clone as a removal case. AuralVerse does not render in the voice of an identifiable living performer, and we do not accept prompts that ask it to.
  3. Volume is not a strategy. Deezer's figures show that most of the fully AI-generated supply is never streamed, and most streams it does receive are fraudulent and demonetised. A rendition only earns its place if a person wanted to hear that song performed that way.

What we are not saying

This article describes how the platforms label recordings. It does not describe how AuralVerse makes them. The model's architecture, its weights and the material it was trained on are not published, and nothing in a platform's disclosure field asks for them. A disclosure that a vocal was machine-performed is a statement about the recording, not about the system. We think the industry has drawn that distinction correctly, and we will keep drawing it the same way.

Deezer's November number was 34 percent. The company's chief executive said in June that "AI is not inherently good or bad, but we believe a responsible and transparent approach is key." Transparency about what a listener is hearing is now a requirement. Transparency about how a model works is a different matter, and no platform has asked for it.