In the summer of 2023, a track called "Heart on My Sleeve" — with vocals cloned to sound like Drake and The Weeknd, neither of whom had anything to do with it — went viral and was pulled from streaming services within days [source: The FADER, 2023]. Three years later, the strange has become the routine. The streaming service Deezer reported in July 2026 that fully AI-generated tracks now make up more than half of the roughly 90,000 songs uploaded to it on peak days [source: Deezer, 2026]. An AI-made R&B act called Xania Monet became the first known AI artist to reach a Billboard airplay chart and signed a record deal reported at up to $3 million [source: Billboard, 2025]. And the major labels that spent 2024 suing AI music companies spent late 2025 signing deals with them.
This is a story about creative labor, copyright, and the machinery of the music business under stress — not a verdict on whether AI music is "good." The honest picture in 2026 is unsettled: some lawsuits are still pending while others have produced their first rulings, company announcements outrun what has actually been built, and the same technology looks like theft to one artist and like a tool to another. The point of this piece is to separate what has been claimed from what has been decided, and to lay out the genuine disagreement underneath.
From novelty to flood
The tools at the center of this are generative music models — most prominently Suno and Udio — that turn a text prompt into a finished song, vocals and all, in seconds. What changed between 2023 and 2026 is less the novelty than the scale. Deezer, which began labeling fully AI-generated tracks and built a detector it says identifies them with 99.8% accuracy, watched daily AI uploads climb through 2025 and into 2026 until, by mid-2026, they crossed half of all new uploads on the platform's busiest days [source: Deezer, 2026].
But volume and listening are not the same thing, and this is the first place the numbers get misread. On Deezer, fully AI-generated music accounted for just 1–3% of total streams even as it dominated uploads [source: Deezer, 2026]. More striking, the platform said that in 2025 up to 85% of the streams going to fully AI tracks were fraudulent — the product of bots and streaming farms rather than listeners — and were excluded from royalty calculations [source: Deezer, 2026]. So the "flood" is real as an upload phenomenon, and part of it is a payout scam aimed at diluting the royalty pool, but the claim that listeners are abandoning human music for machine music is not what this particular data shows. Those are different assertions, and only the first is well supported.
The lawsuits: what is claimed, what is decided
The legal fight began in June 2024, when the major record companies — Sony, Universal, and Warner — brought coordinated copyright suits through the RIAA against Suno (in federal court in Boston) and Udio (in New York) [source: RIAA, 2024]. The core accusation is blunt: that the companies copied vast libraries of recorded music to train their models without permission. "These are straightforward cases of copyright infringement involving unlicensed copying of sound recordings on a massive scale," the RIAA's chief legal officer, Ken Doroshow, said at filing [source: RIAA, 2024]. The suits sought injunctions and damages; under U.S. law, statutory damages for willful infringement can run up to $150,000 per work.
Here the layers matter. As of August 2026, those U.S. cases are still pending — no final judgment on the central question of whether training on copyrighted recordings is infringement or protected "fair use." What did land is a European ruling. On July 31, 2026, a Munich court held that Suno had infringed works represented by GEMA, Germany's music collecting society — the first major European decision to find that training an AI model on protected music without a license can infringe, even when the training happens outside the EU [source: Variety, 2026]. The court found that Suno had "memorized" and reproduced specific GEMA-represented songs, and ordered it to stop and to disclose revenue so damages could be calculated [source: Variety, 2026]. Legal analysts flagged the jurisdictional reach as the bigger story: if serving outputs to EU users pulls a U.S.-trained model under EU law, the exposure is broad [source: Reed Smith, 2026]. Even so, this is a first-instance ruling that Suno can appeal, and the damages figure is not yet set. A pending U.S. suit and a not-yet-final German judgment are not the same thing as settled law.
From courtroom to license
Then the fight changed shape. Rather than wait for verdicts, the labels started cutting deals with the defendants. In October 2025, Universal Music settled with Udio and agreed to jointly build a licensed AI music service, slated for 2026, using only recordings from artists who opt in — and structured as a "walled garden" where creations cannot leave the platform [source: Music Business Worldwide, 2025]. Weeks later, in November 2025, Warner Music settled with Suno in what both called a first-of-its-kind deal: a licensed catalog, licensed AI models replacing the old ones in 2026, paid downloads with caps, and — notably — Suno dropping the "fair use" defense it had built its case around [source: Music Business Worldwide, 2025].
These settlements are being read two ways, and both deserve airing. To the labels and the AI companies, they are the responsible outcome: consent, compensation, and guardrails replacing a free-for-all. To a sharp line of critics, the sequence itself is the problem. As one analysis put it, launching on unlicensed data, getting sued, then settling into a paid partnership can make infringement look like a viable business plan — build the model on other people's work first, pay for it only after it is too big to unwind [source: Forbes, 2025]. Both readings can be true at once, which is why the deals have not ended the argument. And the settlements are not unanimous: Sony did not join, and its claims against the AI companies remained active, a reminder that "the industry settled" overstates a still-fractured front [source: Music Business Worldwide, 2025].
Platforms draw lines
While courts and boardrooms worked, the streaming platforms wrote rules — and here, too, an announcement is not the same as an effect. In September 2025, Spotify laid out a package: it said it had removed more than 75 million spam tracks over the prior year, and introduced a "music spam filter" that flags tactics like mass uploads, near-duplicates, and artificially short tracks gaming the payout system, suppressing them in recommendations rather than deleting them [source: TechCrunch, 2025]. It restated a ban on unauthorized AI voice clones and impersonation, and backed a DDEX industry standard for disclosing AI use — in vocals, instrumentation, or post-production — instead of a blunt "AI or not" label [source: Music Ally, 2025].
The design is careful, but its effect is still unproven. A spam filter that suppresses rather than removes depends on how well it classifies; a disclosure standard depends on labels and distributors actually filing accurate metadata; an impersonation ban depends on detection. Deezer's approach — detect, label, and strip fraudulent streams from payouts — is the more aggressive posture, and its own figures suggest why platforms are moving: most AI uploads are not there to be heard so much as to harvest royalties [source: Deezer, 2026]. The policies are real. Whether they meaningfully change what listeners encounter and who gets paid is the part to watch, not assume.
Deepfake voices and the missing law
Running alongside the training-data fight is a distinct problem: cloned voices. "Heart on My Sleeve" was the early warning — a convincing fake of real, named artists, released without their involvement and pulled only after it spread [source: The FADER, 2023]. Voice cloning raises questions copyright does not fully answer, because it targets identity as much as any specific recording.
The main legislative response in the U.S. is the NO FAKES Act, which would create a federal right against unauthorized digital replicas of a person's voice and likeness, with a notice-and-takedown process and carve-outs for news, satire, and other protected speech. In June 2026 the Senate Judiciary Committee advanced it, but — an important layer — as of August 2026 it is not law; it has cleared a committee, not Congress [source: Congress.gov, 2026]. Until something passes, protection is a patchwork of state right-of-publicity laws and platform policy. Spotify's impersonation ban and the labels' insistence on opt-in consent are, in part, the industry trying to hold a line that statute has not yet drawn.
What creators are actually fighting about
Strip away the case numbers and the disagreement is about a few concrete things. The first is consent: whether a model may be trained on an artist's catalog without asking. The second is dilution — not that listeners prefer AI, but that a flood of cheap tracks, some of them fraud, spreads a finite royalty pool thinner and crowds discovery. The third is credit and identity: whether a voice can be reproduced, and whether an AI act competes on the same charts and radio as human performers. Xania Monet's Billboard entry and record deal made that last point vivid, and the backlash from artists such as Kehlani and SZA was less about one song than about what it signals for everyone downstream [source: Billboard, 2025].
The other side is not nothing. Some musicians use these tools to sketch, arrange, and produce, and treat them as instruments rather than replacements. The licensed, opt-in, walled-garden model the labels are now building is an explicit attempt to let AI generation exist without training on unconsenting artists or leaking imitations into the open market [source: Music Business Worldwide, 2025]. And the institutions guarding human authorship have been clear about the boundary: the Recording Academy's rules state that a work with no human authorship is not eligible for a Grammy, and that only human creators can win, even as AI-assisted works remain eligible when the human contribution is meaningful [source: Recording Academy, 2023]. The fight is not "AI versus music." It is about who consented, who gets paid, and whose name is on the work.
What to watch
Three threads will tell you where this is going. First, the pending U.S. lawsuits and any Sony resolution: a court ruling on training-as-fair-use, or a broad settlement, would set the terms the 2025 deals only sketched. Second, whether the licensed 2026 platforms from Suno and Udio actually ship as promised — opt-in, compensated, walled — or whether the gap between announcement and product stays wide. Third, the law: if the NO FAKES Act passes, voice cloning moves from platform policy to federal right, and the deepfake-vocal problem gets a statute instead of a takedown request.
The measured read is neither "AI will replace musicians" nor "nothing has changed." What the evidence supports is narrower and more interesting: the tools are now producing music at industrial scale, most of that volume is not being listened to and some of it is outright fraud, the biggest rightsholders have moved from suing to licensing without resolving the underlying question, and the rules — in courts, on platforms, and in Congress — are being written in real time. Watch the rulings and the shipped products, not the press releases.