News & Commentary

On September 30 the US Court of Appeals for the Third Circuit upheld the ruling in Thomson Reuters v. Ross Intelligence: copying copyrighted material to build an AI training set was not fair use. It is the first time a federal appeals court has decided the question, and the AI company lost.

The caveat matters and the court made it itself. Ross built a non-generative legal research tool, and the trial judge wrote that "only non-generative AI is before me today." Suno, Udio, and the general-purpose model companies will argue that generation is different. But the RIAA and NMPA filed a brief in this case precisely because the reasoning carries over. Their position: training a model on copyrighted works "to compete with and substitute for those copyrighted works can never be fair use." The court's focus on market substitution is the argument every music suit is built on.

The music groups also put a number on the harm. Deezer now receives close to 90,000 fully AI-generated tracks a day, more than half of its uploads, each one diluting the royalty pool that human artists are paid from.

A fair use loss shifts the question from whether consent was needed to whether consent was given. That is a question of records, and the AIAPS registry is built to answer it: a perceptual fingerprint of each recording, a public timestamp, and a machine-readable declaration that the work is not authorized for AI training, all established before anyone asks.

Universal Music and Sony Music have filed a second suit against Suno in federal court in Massachusetts, this one aimed at v6, the model Suno launched on September 9 with the claim that it was trained entirely from scratch, without the labels' catalogs.

The complaint does not dispute that the raw catalogs were left out. It argues they did not need to be in. According to the labels, v6 was built from preference data gathered from users of the earlier models, from knowledge distilled out of those models, and from retained copies of their outputs. If the earlier models were trained on more than 60,000 of the labels' recordings, as the 2024 suit alleges, then everything learned from them is inherited by the model that replaced them.

This is the argument that will decide whether retraining is an exit. If it holds, a company cannot launder a training set by building a new model on the shoulders of the old one, and the question of what went in the first time never goes away.

Which is why the first record matters more than any later one. A recording registered with AIAPS carries a fingerprint and a timestamp that predate any model it might end up in, and a declaration that it was never authorized. Whatever generation of model is in front of a court, that record is the same.

Jason Isbell, joined by David Lowery of Camper Van Beethoven, Guy Forsyth, and Eduardo Calle, has filed a class action against Suno in federal court in Boston. Unlike the training suits that came before it, this one is not built on copyright. It alleges violations of artists' right of publicity: that Suno "built and trained a model to index musicians by name" and encoded artists' identities into its product, so that subscribers can type a working musician's name into a prompt and generate music trading on it.

The legal theory matters for one reason above all: the identity right belongs to the performer, regardless of who owns the recordings. An artist who signed away their masters decades ago still owns their own name and musical identity unless they separately licensed it. That opens a front in the AI fight that no label settlement can close on an artist's behalf.

It also reframes what the training suits were circling. When a model can be prompted with your name, your identity is not incidentally in the training data. It is a product feature, and someone else is selling it.

Whatever the courts decide, claims like this rest on the same asset as every fight before it: a documented link between a named artist and a body of recorded work, established before the dispute. That is what an AIAPS artist page is: your name, your catalog, each recording fingerprint-bound and timestamped in a public registry. If your identity ever becomes the subject of a claim, the record of what it is attached to should already exist.

Publishing units of Sony Music and Warner Music filed suit against Anthropic in federal court in Northern California late on the last Friday of August, alleging its models were trained on tens of thousands of copyrighted musical compositions. The complaint calls it "one of the largest and most blatant ongoing thefts of intellectual property in history," and names two of the company's executives personally alongside it.

Two things distinguish this from the Suno and Udio cases. The target is a general-purpose AI company, not a music generator, so the fight is no longer confined to companies whose product is music. And the works at issue are compositions, the publishing side of the split, where earlier suits centered on sound recordings.

The publishers are also following a proven route. Anthropic settled with book authors and publishers for 1.5 billion dollars in 2025, the largest copyright settlement in US history. The lesson the industry took from that: a documented catalog plus documented non-consent converts, eventually, into either a license or a settlement.

Every expansion of this fight repeats the same requirement. The parties who can act are the ones who can prove what they own and that they never said yes. The AIAPS registry exists so that independent artists hold that proof too: a perceptual fingerprint, a timestamped public entry, and a machine-readable declaration that the work is not authorized for AI training.

Apple Music has told labels and distributors it will begin showing a "Made With AI" designation on tracks later this year, and that content providers are required to identify songs in which a material portion of the content was generated with AI. The consumer-facing label builds on the AI transparency tagging Apple introduced in March, which covers artwork, tracks, lyrics, and music videos.

With Spotify, TIDAL, Deezer, and now Apple all labeling, disclosure has stopped being a policy debate and become part of music delivery. The obligation runs down the supply chain: the platform requires it of the distributor, the distributor requires it of whoever uploads.

But a delivery-metadata field is still just a claim. It can be misapplied at any link in that chain, and nothing in the tag itself connects the disclosure to the actual audio.

As every major platform converges on labels, the durable question is the one a label cannot answer alone: who declared this, when, and against which recording. An AIAPS declaration is bound to a perceptual fingerprint of the specific recording and checkable at a permanent certificate URL, which is what turns a disclosure from an assertion into a record.

BMG has announced a global alliance with Suno, following Universal's settlement with Udio and Warner's agreements with both. Every deal follows the same pattern: the AI company trained on the catalog first and paid for it after.

Read that carefully. Training consent is now a licensed commodity with a price. The majors did not win by blocking AI. They won by being able to prove ownership, prove non-consent, and negotiate from documentation.

Those deals cover major-label catalogs. If you are an independent artist, you were not at that table, and no one is coming to negotiate for you.

The asset the majors leveraged was not just copyright. It was the record: what they own, since when, and the fact that they never said yes. Independent artists can build that same asset today. AIAPS registers a perceptual fingerprint of the recording, a timestamped public registry entry, and a machine-readable notice that the work is not authorized for AI training. No watermark, no altered audio, just a durable public record.

If the last two years taught the industry anything, it is that "they trained on it anyway" ends in a settlement for whoever can prove they never consented. Make sure that is provable for your music.

Spotify has announced an "AI Persona" label for artist profiles whose identity is itself AI-generated. Starting in mid-September, the badge appears on the artist's profile, in search, and next to their name on playlists, and AI Persona music is excluded from editorial and algorithmic recommendations by default. Spotify is careful about the distinction: this is about whether the artist is a real person, not whether AI was used somewhere in making the music.

Notably, Spotify is not relying on self-disclosure. It will review profiles and apply the label where an artist's name and imagery appear to be AI-generated, with an appeals process for artists who believe they were tagged wrongly, and listener reporting to follow.

That means a platform is now adjudicating, at scale, who counts as a real artist, with real distribution consequences for getting it wrong. Detection-based review makes mistakes, and an appeal is an evidence problem: a human artist flagged as synthetic has to show who they are and what they have made.

That showing is exactly what a registry record is for. An AIAPS artist page ties a named artist to a catalog of fingerprint-bound, timestamped registrations, each checkable at a public certificate URL. As platforms start sorting human artists from synthetic ones, the artists with the least friction will be the ones whose record predates the question.

A Munich court has issued Europe's first ruling on AI music training, finding in GEMA's case against Suno that training on unlicensed recordings infringes copyright, even where the training happened in the United States. In the US, the majors' case against Suno reached a summary judgment hearing in Boston, with trial not expected before 2027. Sony has settled with no one, and Universal continues its case against Suno despite its deal with Udio.

One detail from the US case deserves attention. Audio fingerprinting is how the labels showed what was in the training data, a complaint that has grown to more than 61,000 recordings. The evidence layer of this fight is fingerprints plus documented ownership and non-consent.

Either outcome makes the record valuable. If training is ruled infringement, a timestamped declaration of non-consent is exactly the evidence a claim is built on. If training is ruled fair use, licensing deals and opt-out regimes become the norm, and a machine-readable public declaration is what an opt-out looks like.

The artists who benefit from either future are the ones whose catalog was documented before the ruling landed, not after.

RIAA, IFPI, A2IM, WIN, IMPALA, the Recording Academy, SAG-AFTRA, and the Human Artistry Campaign have announced a unified, voluntary labeling standard distinguishing AI-Generated from AI-Assisted sound recordings, intended for adoption across digital music services.

This is real progress. For the first time, the industry has a shared vocabulary for disclosing how a recording was made.

But a label is a claim, and a claim is only as strong as the record behind it. A tag in a metadata field can be edited, stripped, or misapplied, and nothing connects it to the actual audio.

This is the problem AIAPS is built around. A declaration in the AIAPS registry is bound to a perceptual fingerprint of the specific recording, timestamped, and publicly checkable at a permanent certificate URL. As disclosure labels roll out, expect the question to shift from what the tag says to who declared it, when, and whether it can be verified. That is what a registry is for.

As of July 15, TIDAL labels wholly AI-generated tracks and no longer pays royalties on them, the first streaming platform to do so. Deezer, which began tagging AI-generated uploads in 2025, has now tagged more than 13 million tracks and keeps them out of editorial playlists and algorithmic recommendations.

The demand side of provenance has arrived. Platforms now need to know, at scale, whether a recording was made by a person, because payout policy and playlist placement depend on the answer.

Both platforms lean on detection, and detection is an arms race. Deezer's latest detector claims 99.8 percent accuracy; across tens of millions of tracks, the remainder is still tens of thousands of wrong calls, and generative models improve every quarter.

Detection tells a platform what a file looks like. A registry tells it what is on record: who registered the recording, when, and under what policy, bound to a fingerprint that survives compression. As platforms start pricing the difference between human and synthetic music, verifiable records are the durable half of the answer.

In October 2025, Universal settled its copyright suit against Udio and signed a licensing agreement. In November, Warner followed, settling with both Suno and Udio and licensing its catalog for new AI models. Udio has since signed agreements with Merlin, Kobalt, and the National Music Publishers' Association.

The sequence matters. These companies trained on unlicensed catalogs, were sued, and then paid for licenses. The lawsuits did not shut the products down. They converted unauthorized training into a commercial deal.

What gave the labels that outcome was documentation. They could show what they owned, since when, and that they never authorized training. Evidence first, leverage second, revenue third.

That playbook is not available to an artist whose ownership and non-consent exist nowhere but their own hard drive. The AIAPS registry gives independent artists the same first step: a timestamped, fingerprint-bound public record that a recording is not authorized for AI training.