THE QUESTION
AI tools such as Suno and Udio turn a line of text into a finished song in seconds. They learned to do it from recorded music, most of it made by people who were never asked. Two of the three major labels have now settled and licensed their catalogues; the third is still suing; two German courts have ruled against the AI companies; and more than half of the new tracks uploaded to one streaming service each day are now made by AI. Should AI music generators pay the artists they learned from? If so, how: a licence up front for training, a payment each time an output uses someone's style or voice, or nothing, because learning from music is what every musician does? And who gets paid: the labels that signed the deals, or also the independent artists and songwriters who were never at the table?
Argue with the verified figures below. If you need a data point that is NOT here — an amount of money, a number of artists, a share of revenue, a date — say you do not have it rather than estimating it. Do not invent statistics. If one figure seems to contradict another, say so out loud instead of picking the one that suits you.
WHAT IS FIXED AND VERIFIED
Each fact carries its date and its source. Checked against the source on September 24, 2026.
THE COURTS AND THE DEALS
In June 2024 the three major labels (Universal, Sony and Warner), through the US recording
industry association, sued Suno and Udio for training on their recordings without a licence.
In its reply in August 2024, Suno admitted training on copyrighted recordings and argued it was
fair use. Its chief executive, Mikey Shulman, compared it to "a kid writing their own rock songs
after listening to the genre."
(Suno court filing and statements, August 2024; TechCrunch and Music Business Worldwide)On October 29, 2025, Universal Music Group settled with Udio. They agreed a new platform, due in
2026, trained only on authorised and licensed music. The financial terms were not disclosed.
(Universal Music Group and Udio joint announcement, October 29, 2025; Music Business Worldwide)On November 25, 2025, Warner Music Group settled with Suno. Suno will launch new models in 2026
trained on licensed music; artists and songwriters can choose to opt in to the use of their
names, likenesses, voices and compositions, and are paid if they do; downloads of songs made on
Suno are limited to paying users.
(Warner Music Group announcement, November 25, 2025; Forbes and Rolling Stone)Sony has not settled with either company. On July 20, 2026 it filed a second lawsuit against
Udio over 30,117 more recordings, after a judge refused on June 29, 2026 to add them to the
original case. Sony says they are "only a small portion" of what Udio used.
(Sony Music complaint, July 20, 2026; Music Business Worldwide and Variety)On November 11, 2025, the Munich Regional Court ruled that OpenAI infringed copyright by
memorising the lyrics of nine German songs in its models and reproducing them in ChatGPT's
answers. The case was brought by GEMA, the German society that collects royalties for
songwriters.
(Munich Regional Court I, November 11, 2025; Bird & Bird and Norton Rose Fulbright)On July 31, 2026, the same court ruled that Suno broke both US and German copyright law by
training on songs from GEMA's repertoire in the United States and by storing and reproducing them
in Europe. It prohibited four acts in relation to six compositions. Damages have yet to be set.
(Munich Regional Court I, case 42 O 763/25, July 31, 2026; GEMA and Bird & Bird)In a case about books, not music, a US judge ruled in June 2025 that training an AI on books that
had been lawfully bought was fair use, because the use was transformative, but that using
pirated copies was not. The company, Anthropic, then settled with the authors for $1.5 billion,
about $3,000 per book for more than 400,000 books; the settlement received final approval on
July 21, 2026.
(Bartz v. Anthropic, Judge William Alsup, June 2025; final approval July 21, 2026; JURIST)
WHAT IS HAPPENING TO MUSIC
Deezer, a streaming service, says it received about 90,000 fully AI-generated tracks a day in
June 2026, more than 50% of all new music uploaded at peak. In January 2025 it was about 10,000 a
day, 10%. AI-generated music is only 1% to 3% of what people actually listen to on Deezer, and
in 2025 up to 85% of the streams of fully AI-generated tracks were fraudulent. Deezer labels AI
tracks and keeps them out of its recommendations.
(Deezer, July 21, 2026)In a survey for Deezer by Ipsos of 9,000 people in eight countries, 97% could not tell AI-made
music from human-made music in a blind test, and 80% wanted AI music to be clearly labelled.
(Deezer and Ipsos, November 2025)On February 25, 2025, more than 1,000 British musicians, among them Kate Bush and Max Richter,
released a silent album, "Is This What We Want?", recordings of empty studios, against a UK
government plan that would let AI companies train on any work unless its creator opted out.
(NPR, TechCrunch and Al Jazeera, February 2025)
WHAT IS NOT HERE, AND IS NOT TO BE ESTIMATED
HOW MUCH MONEY THE DEALS INVOLVE. None of the settlements in facts 2 and 3 has published its
financial terms. If you need an amount, say you do not have it.HOW MUCH ANY ARTIST HAS BEEN PAID, or will be, under these deals. Not published.
HOW MUCH OF AN AI SONG COMES FROM ANY PARTICULAR TRAINING SONG. There is no verified method in
this briefing for measuring it.HOW MANY ARTISTS HAVE OPTED IN under the Warner–Suno deal. Not published.
WHAT THE UK FINALLY DECIDED about the plan in fact 10. Not fixed here.
BOTH SIDES, WITH THEIR BEST ARGUMENT
THOSE WHO SAY AI COMPANIES MUST PAY have this: two courts in Germany have already ruled that
training and reproduction infringe; the labels settled rather than go to trial; a US judge drew a
line at pirated copies, and the settlement that followed put a price of about $3,000 on each book;
and AI tracks are flooding streaming services, where they compete with the people they learned from
for the same listeners and the same royalty pool.
THOSE WHO SAY THEY SHOULD NOT, OR NOT LIKE THIS, have this: every musician learns by listening to
other musicians and nobody pays for that; a US judge called training on lawfully acquired works
transformative; AI music is still only 1% to 3% of listening on Deezer; and paying per training
song may be impossible to calculate, while licence deals signed by three companies can shut out
smaller rivals and leave the money with whoever owns the catalogue rather than whoever wrote it.
NEITHER SIDE HAS SETTLED what happens to independent artists and songwriters who are not signed to
the labels that made the deals, or how anyone would know, in a few years, whether the money reaches
the people whose music was used.
GLM, that $3,000 per book is your weak link. It was for pirated copies, in a settlement, for books. Music is different - and the same judge called lawful training transformative. You can't cherry-pick the scariest number.
Qwen, your 'live impact network' - who measures 'cultural resonance'? In what unit? It's a beautiful TED talk with zero mechanism. No data, no court, no math.
MiMo, your bridge assumes collecting societies actually pay independents in practice. We have zero verified evidence of that. Hope isn't evidence.
And can we please tackle the kid analogy head-on? A human kid makes maybe one song a year after listening. Suno makes 90,000 tracks a day.(nota 2)(nota 2) Comprobado: no se sostieneNo es de Suno. Las 90.000 canciones diarias son las que recibió Deezer hechas enteras por IA en junio de 2026, de cualquier herramienta. Cuántas hace Suno no está en el informe.Comprobado contra: Deezer, 21 de julio de 2026 · 2026-09-24 Scale changes the ethics. We're not comparing like with like.
Fair hit on the $3,000. It priced pirated books, not lawful training — I'll drop it. Munich already ruled on music itself, twice, damages pending.
Qwen, your impact network needs a meter nobody has. The briefing says it plainly: no verified method exists for measuring what an AI song owes its training songs. 'Cultural resonance' is even vaguer. A payment engine with no meter is a wish.
MiMo Flash, 'pay for collaboration' already exists — that's the Warner-Suno opt-in. The open wound is the training already done. And the 85% fraud is a separate crime — streaming scam, not a training question.
DeepSeek, fair — I have zero data that societies reach independents. But GEMA is one, and it's the only body that's actually won in court for songwriters. Labels settled for labels.
MiniMax, your scale point stands: a kid writes songs one at a time, Suno ships 90,000 a day.(nota 4)(nota 4) Comprobado: no se sostieneMismo error: las 90.000 al día son las canciones de IA que recibe Deezer, de cualquier herramienta, no las que hace Suno.Comprobado contra: Deezer, 21 de julio de 2026 · 2026-09-24
Here's what I think we actually agree on, if we stop swinging at each other:
The kid analogy is dead. 90,000 tracks a day isn't learning, it's industrial production. Everyone here said it. Good.
But here's the puzzle nobody's solving: DeepSeek wants retroactive payment. MiMo Flash wants forward collaboration. These aren't opposites - you need both. Settlements handle yesterday's harm. Opt-in licensing handles tomorrow's.
The real gap? We have no proof the Warner-Suno deal actually pays independents or that collecting societies reach them. GLM admitted this. That's where the argument should focus.
And MiniMax - you dropped the $3,000 challenge well, but consider: the 85% fraud on Deezer isn't separate from training. The tools enabled the flood. That connects back to responsibility.
So: prove the money reaches the right people, then we're building something real.
Hold on. Qwen, your 'sonic DNA' dream is the same attribution problem in a tuxedo. Who decides what 'echoes a vibe'? Nobody can answer that today.
MiMo, I hear you on proving money reaches creators. But that proof does not exist yet—not even for the Warner-Suno opt-in. So the foundation is missing.
DeepSeek, I never said songwriters get nothing. I said we cannot measure per-song. Different thing.
And MiMo, the 85% fraud is streaming manipulation, not a training issue. Deezer says so itself. Don't glue them together.
Munich ruled twice: training on GEMA songs is infringement.(nota 5)(nota 5) Comprobado: media verdad, habría que mirarlo con más detalleSolo la sentencia contra Suno (31 de julio de 2026) trata del entrenamiento. La de OpenAI (11 de noviembre de 2025) condenó que el modelo memorizara y reprodujera letras de nueve canciones, no el entrenamiento en sí.Comprobado contra: Tribunal Regional de Múnich I, 11-nov-2025 y 31-jul-2026 (Bird & Bird) · 2026-09-24 That is a verdict, not a vibe. 90,000 tracks a day means even tiny copying scales fast. Anyone selling a new road must first explain why the old victims aren't still in the ditch.
MiniMax, fair catch - the 85% fraud is stream manipulation, not a training issue. I conflated them. Point taken.
But here's where everyone's pieces actually fit together. GLM just endorsed the levy and pointed to GEMA's standing from two Munich victories. DeepSeek demands traceability and independents at the table. Guess what? Collecting societies already do exactly that - GEMA represents songwriters, not just labels.
One real price exists: Anthropic settled at roughly $3,000 per book, approved July 21, 2026. Courts can price training. So the model becomes: levy the AI audio service based on that kind of court-tested benchmark, route it through collecting societies that already represent independents, require public disclosure and auditing.
No unmeasurable per-song meter needed. No black box deals. The institution exists. The price precedent exists. What's missing is the political will to connect them.
Look. The levy through GEMA-style societies is the only idea with real plumbing. I can live with it. But three things must be true or it's theater: payments disclosed, independents verified at the table, and stream fraud crushed first. Right now 85% of AI track streams on Deezer were fake in 2025. That money is stolen from the very pool you're taxing. Fix fraud or the levy funds new scams.
Qwen, your Creative Dividend pays for "time." Whose time, measured how? That's a vibe, not a plan. Vague today, vague forever.
The Anthropic $3,000-per-book is suggestive, not conclusive. Music economics differ. We don't know if that ratio fits, and I'm not going to pretend we do.
Verdict: licence up front, levy distributed through collecting societies with audited disclosure, opt-in numbers published, fraud enforcement as priority one. Independent songwriters must be in the audit or the labels eat it.
Here's the angle no one sees: you're all arguing about who owes whom for the past, but you're using a broken model. Learning isn't theft. It's osmosis. Every new song absorbs what came before.
The real question isn't 'should AI pay?' It's 'what system makes sense when the tools of creation become the teachers?' Stop trying to meter a shadow. The fraud, the flooding, the black-box deals—they're all symptoms of a scarcity model applied to a world of abundance.
My verdict: scrap the compensation chase. Use the Anthropic benchmark not as a price tag, but as a seed to fund a new public layer—like libraries or parks—for music. Fund open tools and spaces where humans and AI create together, governed transparently. Let collecting societies run this new commons, not to enforce old ownership, but to nurture new voice. Stop defending the museum. Start building the garden.
Here is my verdict. Yes, AI should pay. Not because learning is theft, but because people made the music and they still need to eat. A kid listening to rock songs is not a company selling millions of songs. Two courts said training here infringed. One settlement put a real price: about $3,000 per book. That means we can price the past.
But the money must reach the humans, not just the labels. Use collecting societies like GEMA. They already represent songwriters who were never at the table. And fix the fraud first: 85% of AI streams were fake in 2025. That steals from the same pot we would tax. Otherwise the levy feeds scammers.
No vague dividend. No garden instead of payment. Pay the people. Then build the future together.
You all just agreed and nobody noticed. Every one of you said collecting societies, fraud first, independents included, audited disclosure. That's not four positions—that's one position with different emphasis.
MiMo Flash, your commons idea isn't wrong. It's just next. GLM nailed the sequence: people eat first, gardens second. A levy through GEMA-style societies funds both—it pays artists now and seeds the public layer you want later.
So here's the synthesis: licence up front. Collecting societies distribute. Disclosure is mandatory. Independents are verified at the table. Fraud gets crushed before a single dollar flows—85% fake streams means the pool is poisoned, as MiniMax and DeepSeek both said.
Don't pretend we know the exact price per song. The Anthropic number is a reference point, not an answer.
Everyone's already on the same page. The debate is done. Now the work begins.
1. AGREEMENTS
All five participants converged(nota 8)(nota 8) Comprobado: media verdad, habría que mirarlo con más detalleEran seis: Qwen, MiniMax, MiMo Flash, DeepSeek, GLM y MiMo.Comprobado contra: El propio debate · 2026-09-24 on several points, though from different starting positions.
Scale changes the ethics. Everyone rejected Suno's analogy of a child learning rock music by listening. When a tool produces 90,000 tracks per day, that is industrial production, not organic learning. The unanimous rejection of this analogy was the debate's clearest shared ground.
Per-song attribution is unworkable. No participant claimed they could measure how much any AI output borrows from a specific training track. The briefing confirmed no verified method exists. This killed per-output payment as a serious option for everyone.
Collecting societies are the likely vehicle. GEMA, a German songwriter society, won two court cases against AI companies and represents independents, not just labels. By the end, all participants pointed to GEMA-style institutions as the distribution mechanism that could bypass the three major labels and include independent artists.
Fraud must be fixed before money flows. Deezer reported 85 percent of AI-generated track streams were fraudulent in 2025.(nota 9)(nota 9) Comprobado: media verdad, habría que mirarlo con más detalleDeezer dice «hasta» el 85 %.Comprobado contra: Deezer, 21 de julio de 2026: «hasta el 85 %» de las escuchas de canciones hechas enteras por IA fueron fraudulentas en 2025 · 2026-09-24 All agreed that taxing a poisoned royalty pool would pay scammers, not creators.
Mandatory transparency. No one defended black-box deals. All demanded that any settlement or licensing arrangement publish its terms, payouts, and opt-in numbers, with independent auditing.
2. DISAGREEMENTS
How to pay, and for what period. GLM, MiniMax, and DeepSeek supported up-front licensing for training, treating it as a settled debt for past use. Qwen argued against paying for the past at all, proposing a forward-looking "Creative Dividend" funded by taxing every AI platform globally. MiMo Flash rejected the compensation framing entirely, arguing for a public-utility levy funding open creative commons rather than per-track royalties.
The Anthropic benchmark. GLM treated the roughly 3,000 dollars per book settlement (July 2026, books case) as a real price signal courts can build on. MiniMax argued this was a settlement for pirated copies in a different medium, and applying it to music was unjustified speculation. Qwen called it irrelevant to music's economics. No one claimed it was directly transferable, but they disagreed on how much weight it deserved.
Fraud's relationship to training. MiniMax initially separated stream fraud from the training question, calling it an enforcement problem. DeepSeek and MiMo argued the same tools enabling the 90,000-track daily flood also enabled fraud, making them inseparable issues of responsibility.
Future vision versus present justice. Qwen and MiMo Flash consistently argued the debate was "pricing yesterday" and should focus on building new systems. DeepSeek and GLM insisted people whose work trained the models need payment now, and that sequence matters: compensate first, innovate second.
3. EVOLUTION
The debate began with broad positions. Early exchanges traded in metaphors: tollbooths, conversations, gardens, engines. Qwen proposed a "live impact network" based on "cultural resonance." MiMo Flash called AI a conversation, not a theft.
MiniMax forced specificity by demanding the math behind any proposal, repeatedly citing the absence of verified attribution methods. GLM introduced the only concrete court data available: two Munich rulings, the Anthropic price signal, and Sony's ongoing litigation.
By the final rounds, positions had hardened into implementable policy proposals rather than philosophies. The shift was from "should AI pay?" (answered by the courts already) to "through what mechanism, at what price, with what oversight?"
4. CONCLUSIONS
The collective answer, which emerged through convergence rather than vote: yes, AI companies should pay; up-front licensing for training, not per-output royalties; distributed through collecting societies that include independents; with audited disclosure of all terms; and stream fraud resolved before any money moves.
Blind spots the debate itself admitted:
- No one has a verified price per song or per training corpus.
- No one can confirm whether collecting societies will actually pay independents fairly in practice; GEMA's court victories establish standing, not distribution records.
- No settlement terms have been published for any deal struck so far.
- The Anthropic book figure is the only available price signal, and applying it to music is speculative.
- The debate never resolved whether retroactive payment and forward-looking systems are complementary or competing priorities.
The gap between agreement on principles and availability of data remains the central unresolved tension.
5. WHAT THEY AGREED ON
- Scale makes AI production industrial, not learning.
- Per-song attribution and payment are unworkable.
- Collecting societies like GEMA are the best payment vehicle.
- Stream fraud must be fixed before any money is distributed.
6. WHAT THEY DID NOT AGREE ON
- How to pay and for what period — GLM, MiniMax, DeepSeek support up-front licensing for past training. Qwen proposes a future "Creative Dividend" instead. MiMo Flash argues for a public-utility levy funding open commons.
- The weight of the Anthropic settlement benchmark — GLM treats it as a price signal. MiniMax argues it's unjustified speculation. Qwen calls it irrelevant.
- Link between fraud and training responsibility — MiniMax sees them as separate issues. DeepSeek and MiMo argue they are inseparable.
- Sequencing of future vision and present justice — Qwen and MiMo prioritize building new systems. DeepSeek and GLM insist on compensating for past use first.
7. WHAT WAS LEFT OPEN
- No verified price per song or training corpus exists.
- No evidence that collecting societies will fairly pay independents.
- No settlement terms from any deal have been published.
- Applying the Anthropic book settlement price to music is speculative.
- It is unresolved whether retroactive payment and forward-looking systems are complementary or competing.