Back to the journal
Xdripia Waves Journal / AI

AI in Music: Labels, Detectors and Who Owns Your Voice

YouTube Music now labels AI songs, Deezer's detector wins a TIME award, labels chase DistroKid's AI data, and voices move into trademark and consent law.

6 min readMusic. Culture. Creation.
Xdripia Waves AI cover: a glowing audio waveform with a red segment under a scanning line, a frosted glass AI label panel with a toggle, a voice-print consent panel and a neural mesh on a dark dotted canvas, with the headline Who Decides What Is AI?

On Tuesday Deezer said it would let listeners hide fully AI-generated music from search, a move we covered in our last AI edition. A day later YouTube Music started putting labels on AI tracks. The rest of the week made one thing clear: the hard question is no longer whether platforms should mark AI music, but who decides what counts as AI, on what evidence, and who owns the voice in the file.

YouTube Music puts a label on the player

On October 14, YouTube announced that YouTube Music is rolling out labels for songs "created primarily using generative AI", for example tracks where a generative model created both the composition and the vocals. The label sits directly on the player page, tapping it opens more information, and the labels will also expand to the main YouTube app.

YouTube says it identifies these tracks with information from music partners plus its own detection tools. As Music Business Worldwide reports, the Help Center page goes further: if partners send no generative AI information, YouTube "will rely solely on those signals and may designate content as fully or partly Gen AI." Neither the blog post nor the Help Center says whether labeled tracks are treated differently in recommendations or royalty payouts.

One song, two definitions

Labels only work if everyone agrees on the categories, and they don't yet. The AI Musicpreneur's weekly briefing compared the rules:

  • YouTube counts a fully human song with an AI-generated music video as "Partly Gen AI", while Auto-Tune and AI mastering count as "No Gen AI".
  • Deezer draws its line at "limited AI-assisted elements".
  • RIAA and IFPI proposed two voluntary labels in July: "AI-generated" for tracks built entirely by AI or with an AI lead vocal or main instrumental, and "AI-assisted" for mostly human work, according to MBW.

The other services are moving at different speeds. MBW notes that Apple Music told partners in August that its AI Transparency Tags would become visible as "Made With AI" labels later this year, based on declarations from labels and distributors. Spotify began testing AI disclosures in song credits in April and announced an "AI Persona" badge in August. TIDAL said in June that it would tag fully AI-generated music and stop paying royalties on it.

The practical result: the same track can land in different buckets on different platforms, and most of the system still depends on someone upstream telling the truth.

The detector gets a trophy

The detection side had its own moment. On October 15, Deezer announced that TIME had named its AI music detection tool to the Best Inventions of 2026 list. Deezer's own figures:

  • nearly 90,000 fully AI-generated tracks delivered to the platform every day;
  • more than 42 million AI tracks detected and removed from algorithmic and editorial recommendations between January 2025 and August 2026;
  • a stated accuracy of 99.8%, missing about two AI tracks in 1,000 and wrongly flagging fewer than one human track in 10,000.

Those numbers come from Deezer, as of July 2026. A detector also answers a narrower question than a label implies. It can say there are traces of a generator in the audio, but not who wrote the lyrics or which parts a person played.

Detection data goes to court

Detection is now evidence. According to Digital Music News, Universal Music Group and Sony Music filed a new action in the Southern District of New York to make DistroKid disclose how many AI-generated songs it delivers to streaming services, particularly from Suno, and details of its AI-detection tools.

DistroKid is not a party to the labels' case against Suno, and it had argued the data is irrelevant to that case. The majors call DistroKid a "primary conduit" for AI music and say the information is "proportional to a case in which the market effect of AI-generated music is a central contested question." The filing claims DistroKid delivered nearly 12 million tracks to a single unnamed service between March and September, more than all other distributors combined. The court had not ruled at the time of writing.

Your voice as a trademark, your face as a right

While platforms sort files, artists and lawmakers are working on the identity behind them.

Taylor Swift. MBW reports that her company, TAS Rights Management, filed sound mark applications for the spoken phrase "Hey, it's Taylor Swift" in the UK, the EU and Australia on October 6. IP Australia accepted the Australian application on October 12, six days later; it is not yet registered and is in its opposition period. The UK and EU filings are still in examination. As IP attorney Josh Gerben noted when the US filings surfaced in April, registering a celebrity's spoken voice "has not been tested in court before." The Austrian Patent Office adds a caution: a sound mark covers one recorded phrase, in commerce and for the listed services, not a voice in its entirety.

Denmark. On October 8, Culture Minister Zenia Stampe tabled bill L 49. It would make it unlawful to share realistic digital imitations of a person's voice or appearance without consent, and adds protection for imitations of performers' performances. PPC Land's analysis highlights that the publisher must prove consent, protection lasts 50 years after death, and satire and parody stay allowed unless the imitation is misinformation that can cause serious harm. The proposed start date is January 1, 2027, if parliament passes it.

What this means if you make music with AI

None of this is legal advice, but a few habits already make sense:

  1. Declare before you're detected. Ask your distributor which AI tags it sends to YouTube, Apple Music and Spotify, and pick the one that matches what you actually did.
  2. Keep a production log. Note which parts were written, sung or played by people and which tools and models were used. Labels and disputes increasingly come down to that record.
  3. Treat voices as someone else's property. Voice models of real artists are moving from a grey zone into trademark filings and consent laws.
  4. Expect different answers on different platforms. Until definitions converge, check each service's help pages rather than assuming one rule applies everywhere.

In the coming editions we will keep watching how these labels behave in practice and whether the DistroKid data ever reaches the public record.

Sources

Keep exploring

More from the journal.

Music, sound design and the ideas behind the work.

Explore all stories