Twelve million tracks. Played end to end, with no gaps, no sleep and no skipping the songs you do not like, it would take ninety-one years to listen to that dataset in full. A person born the year the recordings began would be dead long before reaching the last track. And yet the entire collection can be downloaded in an afternoon, copied onto a laptop, and fed into a machine designed to learn from it and then make more music in its image. The recordings inside it represent the accumulated labour of tens of thousands of musicians, living and dead, famous and obscure. Not one of them was asked. Not one of them was paid.
This is the picture that came into focus in mid-June 2026, when The Atlantic's AI Watchdog journalist Alex Reisner published an investigation that did what the industry had demanded for years: it provided proof. Reisner identified four music datasets actively circulating within AI development communities, downloaded thousands of times, and built a searchable tool that let any artist type in their own name and discover whether their work had been swept into the training corpus of the machines now flooding the platforms where they earn their living. The largest holds roughly twelve million tracks; a second around nine million; two more each exceed a hundred thousand — more than twenty-one million recordings in all. The names inside read like a history of recorded music: Taylor Swift, The Beatles, AC/DC, Kylie Minogue, Miles Davis, Bad Bunny, and tens of thousands of working musicians whose names you have never heard.
The reaction was immediate, and it was personal. The rapper Backxwash, finding her work in the database, wrote simply: “I dont like this.” The Toronto producer Tre Mission was blunter: “I'm 100% sure I never consented to this. Anyone who knows me, knows I HATE the use of AI in music, so this is very disappointing.” The producer DJ Sabrina the Teenage DJ landed on the bitter irony at the heart of the whole affair: “to everyone who thought my music sounded like ai slop, did you ever think it was because Suno was using a dataset that contained 22 of my songs?”
That sentence deserves to be read twice. An artist's work is ingested without consent. A machine learns from it and generates derivative output that floods the market. And the human whose labour made the machine possible is accused of sounding like the machine. The appropriation is not only economic. It is a theft of authorship turned back against the author.
This is the story of how creative ownership came to be quietly redefined out from under the people who depend on it, why the technology industry's preferred legal solution would make matters worse, and what a genuine licensing framework — one built on consent rather than its absence — would have to look like.
Two Harms, Not One
To think clearly, you have to separate two injuries often blurred together because they share the same villain. They do not have the same solution.
The first is the output-side harm: the flood of synthetic music now competing with human work on streaming platforms. Deezer reported in April 2026 that AI-generated tracks accounted for forty-four per cent of all new music uploaded to its platform, close to seventy-five thousand fully synthetic tracks every single day. That is nearly half of everything arriving, and the figure had climbed steeply. Deezer began detecting AI music at around ten thousand tracks a day in January 2025; by January 2026 it was sixty thousand. The trajectory points one way.
A peer-reviewed paper published to arXiv on 16 June 2026 — “An Empirical Analysis of AI Slop in Music Streaming,” by Stanley Wu, Josephine Passananti, Viresh Mittal, Wenxin Ding, Haitao Zheng and Ben Y. Zhao — put hard numbers to what this deluge consists of. Analysing 256 million Spotify tracks, around ninety-nine per cent of the platform's catalogue as of mid-2025, they found that the overwhelming majority of AI music — ninety-three per cent of it — receives few if any listener plays and is rarely recommended. AI musicians, they observed, “spray and pray”, releasing enormous volumes across genres in the hope that something catches. AI-only artists uploaded at double the volume of human musicians, averaging twenty-seven tracks against thirteen, and at roughly five times the frequency — five releases a month against one, with an average gap between releases of sixteen days against more than fifty. Publishing their own AI tracks through eleven independent distributors, they found distributor policies inconsistent and largely unenforced, and detection unreliable. Their conclusion was that, as the cost of generation falls toward nothing, AI slop is on course to become a self-sustaining industry — a shadow economy of machine-made noise driven by the same incentives that drive email spam.
Spotify removed seventy-five million tracks it classed as spam in a single twelve-month action disclosed in 2025. The output-side harm is real, measurable and accelerating, and it is visible: you can see the slop, count it, and watch it crowd the shelves.
The second harm is quieter, and in the long run more fundamental. It is the input-side harm: the unlicensed ingestion of human creative work to build the machines in the first place. This is what Reisner's investigation exposed. The output flood is the symptom; the training data is the cause. You cannot have a machine that generates plausible Kylie Minogue without first feeding it Kylie Minogue — or, more precisely, tens of thousands of recordings whose collective stylistic DNA the model distils into something it can recombine on demand. Every synthetic track competing with a working musician is built out of that musician's labour and the labour of their peers, taken without asking.
The two harms compound each other in a cruel loop. The input harm produces the machine; the machine produces the output harm; the output harm dilutes the income of the artists whose work constituted the input. The musician is taxed twice — once when their work is taken to build the competitor, again when the competitor thins out their royalties. DJ Sabrina the Teenage DJ's complaint captures the closing of that loop exactly. The slop she was accused of resembling was partly made of her.
What “Ownership” Was Supposed to Mean
To understand what is being lost, it helps to be precise about what creative ownership has meant in music, because it is more intricate than the word “copyright” suggests.
A single recorded song is not one piece of property but a bundle of rights, layered and separately owned. There is the composition — the melody and lyrics — and there is the sound recording, the specific captured performance, each carrying its own copyright and often held by different parties. From these flow distinct streams of payment: the mechanical right, the performance right, and the rights in the master recording. When you hear a song on the radio, this hidden machinery routes fractions of a penny to the songwriter, the publisher, the performer and the label, each according to a right they own.
The edifice rests on a principle that has held, in one form or another, for more than a century: that the creator controls whether and how their work is used, and is entitled to be paid when it is. That principle is now under quiet assault, because the defining feature of the training-data regime is that the creator controlled nothing and was paid nothing. The bundle of rights was simply bypassed. The work was treated not as property to be licensed but as raw material lying around to be scraped.
Sampling, the closest historical analogue, built an entire clearance industry precisely because the law insisted that creators retained control. In the golden age of hip-hop sampling in the late 1980s, producers freely looped fragments of older records in a grey zone. Then the courts closed it. The 1991 decision in Grand Upright Music v. Warner Bros., concerning Biz Markie's use of a Gilbert O'Sullivan song, established that you could not simply take; the 2005 ruling in Bridgeport Music v. Dimension Films held that even a two-second sample required a licence. The industry's response was not to abandon sampling but to build the infrastructure to clear it: clearance specialists, negotiated fees, credit and royalty splits. The principle of control was upheld, and a market grew up around honouring it.
The training-data regime inverts that history. It is sampling at planetary scale with the clearance step deleted. Where a lawyer once negotiated for a four-bar loop, the AI developer ingests twelve million whole recordings and asks no one. The difference in scale is so vast it becomes a difference in kind. And the industry's argument for why this is acceptable is not that it has found a way to clear twenty-one million tracks. It is that it should not have to.
The Carve-Out
That argument has a name, and a lobbying strategy behind it. The name is the text-and-data-mining exception — TDM for short — and the strategy is to persuade governments to write it into copyright law before the courts or the public can object.
The premise is seductively technical. Training an AI model, the argument runs, is not “copying” in the ordinary infringing sense; it is “mining” — extracting statistical patterns, analysing rather than reproducing. Ingesting a song to learn from it is, on this view, more like a scholar reading a library than a bootlegger pressing counterfeit discs. The European Union opened the door in its 2019 Copyright Directive, whose Article 4 permits mining by default unless the rights holder has expressly opted out through machine-readable means. The burden sits on the creator. Silence is consent. You must actively forbid the taking, in a format the machines can read, or the taking is permitted.
In December 2024 the United Kingdom government proposed importing exactly this model: a broad TDM exception with an opt-out for rights holders. The creative industries revolted. More than a thousand musicians released a silent protest album, “Is This What We Want?”, its tracks recordings of empty studios; McCartney, Elton John, Dua Lipa and Kate Bush lent their names to the opposition. When the consultation closed, the government had received 11,520 submissions, and the verdict was emphatic. Eighty-one per cent named mandatory licensing as their preferred approach; a mere three per cent backed the government's own stated preference. More than ninety per cent agreed that AI developers should be required to disclose the sources of their training material — a figure worth holding on to, because it identifies the remedy the sector actually wants. The exception was shelved, and by January 2026 the Secretaries of State Liz Kendall and Lisa Nandy were telling the House of Lords that the government had been “wrong” to express its original preference at all. Australia had already gone further, ruling out a TDM exception outright and telling AI firms they would have to negotiate licences like everyone else.
Which brings us to the most pointed institutional response to Reisner's investigation. APRA AMCOS, which manages the rights of more than a hundred thousand Australian and New Zealand songwriters and composers, seized on the findings as proof of a charge it had been making for months. AI companies, its chief executive Dean Ormston pointed out, were at that very moment lobbying the Australian and New Zealand governments for a copyright carve-out — dangling promises of data-centre investment and “productivity gains” — while simultaneously, demonstrably, using work they had already taken without licence. “No permission. No licence. No payment,” the organisation stated. “These are not bargaining chips — they are the life's work of Australian and New Zealand songwriters.” Ormston added that the major tech platforms had “not come to the table — not once.”
That last point is the tell. The carve-out strategy is not an attempt to negotiate a fair price. It is an attempt to remove the obligation to negotiate at all — to retroactively legalise a taking that has already occurred, and to license all future takings by default. The datasets Reisner identified are downloaded, copied and in use. The carve-out would not authorise a hypothetical future practice. It would launder a present one.
The geopolitical shape of the campaign reveals the strategy more clearly than any single statement. The lobbying happens at the tables where governments weigh data-centre investment, jobs and the prestige of hosting frontier technology against a sector that is, in raw economic terms, smaller and less able to promise headline capital expenditure. The pitch is unmistakable: relax the copyright rules, and the investment follows. That turns a question of property rights into one of industrial policy, on ground where creators are structurally disadvantaged. A songwriters' collective cannot promise a billion-pound server farm in a marginal electorate. A technology company can. The carve-out is sought not because it is just but because it can be traded for things governments want, and those whose work would be expropriated are not party to the deal.
That the strategy is not invincible was demonstrated on 15 July 2026, when the Australian Prime Minister, Anthony Albanese, speaking at the University of Sydney, made the plainest statement any head of government has yet made. Unlicensed training on creative work, he said, is theft: “No company should use Australian books, music, art or news to build or train AI without the artist's control.” He announced a new Office of AI within the Department of the Prime Minister and Cabinet, put the framework to a National Cabinet meeting in August 2026, and committed to legislation to be introduced to Parliament in early 2027 — the first such framework any government has moved to legislate.
One detail deserves particular attention. Albanese rejected outright a proposal from Anthropic for a “shared creative fund” as an alternative to individual licensing: a pooled payment offered in place of consent. The refusal rests on precisely the ground this argument turns on. A fund is a price paid to a category. A licence is permission sought from a person. The first can be generous and still leave the creator with no say in whether their work is used at all — which is the thing that was taken.
What Is Wrong With “Fair Use”
The scholar-in-a-library analogy did not survive July 2026. On the fifteenth of that month, 404 Media published the contents of Suno's source code, obtained by a hacker who had used a supply-chain attack in November 2025 to acquire an employee's credentials. What the code documents is not analysis. It is industrial-scale harvesting, itemised by the hour: 113,879 hours logged as “youtube_music” and a further 152,162 as “ytm_tagged”; 62,117 hours from the stock library Pond5; 19,514 from the International Music Score Library Project; 17,615 from Genius; 12,287 from Deezer. More damning still, the code shows scraping requests routed through commercial proxy services using rotating IP addresses, for the express purpose of defeating YouTube's anti-bot protections. Suno confirmed a security incident, describing it as involving outdated source code and saying it had been contained quickly.
A scholar in a library does not rotate IP addresses to defeat the lock on the door. That detail collapses the distinction the mining argument depends on, converting a passive claim — that developers merely observed what was publicly available — into an active one. It also carries a legal consequence the fair-use debate has largely overlooked. Under Section 1201 of the United States Digital Millennium Copyright Act, circumventing a technological measure controlling access to a copyrighted work is independently actionable. It requires no proof of underlying infringement, and there is no fair-use defence to it. The mining question and the circumvention question are separate, and only one has ever been arguable.
The disclosure also gives the input-side harm something it lacked: a named commercial generator, rather than a dataset of uncertain provenance circulating among developers. And it lands hardest on the artists least equipped to respond. Deezer and YouTube are exactly where the independent catalogue lives — where the producer with a few dozen releases and no label puts their work so that people can hear it. DJ Sabrina the Teenage DJ's complaint about Suno was made before any of this was known. The code suggests the mechanism.
Beneath the specifics lies a deeper conceptual problem. The TDM and fair-use arguments treat the analytical character of machine learning as if it settled the moral question: because a model “learns patterns” rather than storing literal copies, no meaningful harm is supposedly done. But this conflates the mechanism with the consequence. It does not matter, to the musician whose income is being eroded, whether the machine memorised their song or abstracted its style into a vector of weights. What matters is that it could not exist without their work, and now competes with them. The “transformative” defence was designed for a world in which a transformation produced something that did not substitute for the original. A parody does not compete with the song it mocks. A generative model trained on a catalogue produces output that competes directly with that catalogue, on the same platforms, for the same finite pool of listener attention and royalty money. The substitution is not incidental. It is the entire commercial point.
European courts have begun to answer the question the TDM argument treats as settled. In November 2025 the 42nd Civil Chamber of the Munich Regional Court ruled against OpenAI, finding that storing copyrighted content in a model's parameters during training constituted reproduction, and that the TDM exception did not cover commercial uses producing outputs closely similar to protected works. The same chamber, before Judge Elke Schwager, is due to deliver its verdict against Suno on 31 July 2026, postponed from 12 June for administrative reasons. The claim was filed on 21 January 2025 by GEMA, the German collecting society, which represents more than 95,000 songwriters, composers and publishers and holds a mandate reaching more than two million rights holders worldwide; it is the first European case to test AI training on audio. At a hearing on 9 March 2026 the judge had the original recordings and Suno's outputs played aloud in court, across six works including Alphaville's “Forever Young” and “Big in Japan”, Kristina Bach's “Atemlos”, Lou Bega's “Mambo No. 5” and Boney M's “Rasputin” and “Daddy Cool”. Under German law a first-instance ruling is immediately enforceable while appeals proceed, so GEMA could move for an injunction against Suno's European operations. Whatever the outcome, the proposition that training is analysis and analysis is exempt has already been tested in a European court and found wanting.
There is also a question of consent no technical reframing can dissolve. Opt-out is not consent; it is the absence of refusal, a very different thing. A regime that takes by default and asks the dispossessed to object after the fact, in a machine-readable format most have never heard of, is not a licensing system. It is designed to maximise the volume of work taken before anyone notices.
The Working Musician's Position
It is tempting, when the numbers run to twenty-one million tracks and ninety-one years of audio, to lose sight of the individual at the bottom of the pile. But the individual is the point: the harm falls hardest on those least able to absorb it.
Consider a mid-tier independent musician — not a superstar with a legal department, but a working professional with a few dozen releases, a modest but real audience, and an income assembled from thin streams: a little touring, a little sync licensing, a few hundred pounds of streaming royalties each quarter. This is the artist who distributed their music through an aggregator years ago, ticking a box they barely read, assuming the platform would help people find their work. They did not imagine that “distribution” might one day mean their entire recorded output copied into a research dataset and used to train a system that generates an infinite supply of music in roughly their idiom.
The consent question bites hardest here, because what this artist agreed to and what was done are not the same thing. They granted a distributor the right to place their music on streaming services so that human beings could choose to listen to it. It is a very long way from that to the proposition that the recordings could be aggregated into a corpus, divorced from any listening event, and used as feedstock to manufacture a machine designed to generate competing product at industrial scale. No reasonable reading of “help people find my music” extends to “build the thing that will replace me.” The Free Music Archive, one of the smaller datasets Reisner found, was assembled as a public resource for free and legal downloads; the LAION collection anchoring the twelve-million-track set was, its maintainers insisted, “released for research purposes.” Each was made for one purpose and repurposed for another.
Every link in the chain compounds the injury. The work was taken without consent — the input harm. The machine it helped build floods the platform with synthetic competitors — the output harm. The slop dilutes the shared royalty pool from which they draw. And when their genuinely human work resembles the machine output, they may be dismissed as slop themselves, their authorship questioned by the technology built on their labour. The producer who told MusicTech that “until the major labels go through their lawsuits, there's no way for artists or labels to fight back” was describing a structural powerlessness. A lawsuit is a rich entity's tool. The Reisner database gave this artist evidence — but evidence of a wrong is not a means of redress.
That erosion can now be sized. APRA AMCOS's AI and Music Report, the largest study of its kind, drew responses from 4,274 songwriter, composer and publisher members across Australia and New Zealand. Its central finding is that, without a mandatory licensing framework, creators face a twenty-three per cent hit to their revenues — a cumulative loss estimated at more than AUD$519 million over four years, half a billion dollars taken out of the incomes of people who, in the main, are not wealthy. Eighty-two per cent said they feared no longer being able to make a living from their work; ninety-seven per cent wanted policymakers to pay more attention to AI and copyright. Those are not the numbers of a sector resisting change out of nostalgia. They are the numbers of a workforce watching its economic base removed while being told the removal is progress.
There is also a category of injury the language of income and royalties cannot reach, and APRA AMCOS foregrounded it in its response. The datasets did not discriminate. Alongside the pop and rock catalogues they swept up sacred and culturally significant recordings by Aboriginal, Torres Strait Islander and Māori artists — Yothu Yindi, Gurrumul, Warumpi Band, William Barton, Christine Anu, Dan Sultan, Emma Donovan, Barkaa and AB Original among them, and on the New Zealand side Stan Walker, Six60, Maisey Rika, Marlon Williams and Horomona Horo.
For a great deal of this material, the harm is not principally economic. Many such recordings carry cultural protocols governing who may perform them, who may hear them, in what context and at what time; some belong to a community rather than an individual, and some may not be reproduced at all outside particular circumstances. Indigenous Cultural and Intellectual Property is a body of rights and obligations that Western copyright was never designed to hold, and it does not convert into money. A royalty cheque does not answer the copying of a work that should not have been copied at any price, still less a machine producing imitations of it on demand for anyone who types a prompt. This is the clearest demonstration that consent, not compensation, is the load-bearing principle. A framework built purely on payment — the pooled fund, the levy, the blanket rate with no way to say no — would process these recordings as line items and fail completely. Some uses are not underpriced. They are refused.
This is the practical meaning of creative ownership in 2026: for the artist at the top, a portfolio defended by lawyers and eventually licensed on favourable terms; for the artist in the middle, a notional right they cannot afford to enforce, attached to work already taken. The concept has not been abolished. It has been hollowed out — left standing as a word while the substance drains away.
What a Genuine Framework Would Require
So what is the alternative? Not the absence of any framework — that is the status quo, and the status quo is the problem. The alternative is a framework built on the principle the carve-out is designed to erase: that the creator controls the use of their work and is paid for it. It would have to do four things, and do them together, because any one without the others fails.
First, consent — and consent means opt-in, not opt-out. The default must flip. A model trainer should have to obtain permission before ingesting a body of work, not after, and the absence of an objection must never be treated as agreement. This is the most contested point, because it is the most expensive for the technology industry and the most protective of creators. Opt-in is harder. It is also the only version of consent that deserves the name. The sampling-clearance precedent shows it is workable: the industry did not collapse when courts required samples to be cleared; it built the infrastructure to clear them. The objection that opt-in is impossible at the scale of millions of tracks is precisely the objection collective management was invented to answer.
Second, transparency and provenance. A genuine framework would require trainers to disclose what they trained on — to maintain and publish auditable records of the works in their datasets. This is the reform that commanded more than ninety per cent support in the British consultation, and its absence is what made Reisner's investigation necessary: the datasets circulate privately, and artists discover their inclusion only when a journalist builds a search tool, or a hacker leaks the source code. Mandatory provenance would invert that, letting any creator know as a matter of routine whether their work had been used, by whom, and under what licence. It is the precondition for everything else: you cannot pay for what you cannot trace, nor enforce a right you cannot prove has been violated. Credit belongs here too — the humans whose work shaped a model's capabilities should be acknowledged, not erased into a statistical soup.
Third, compensation — and the mechanism matters as much as the principle. The music industry already possesses the most relevant tooling of any creative sector, having spent a century paying large numbers of rights holders small amounts at enormous scale. The performing-rights and collective-management organisations — APRA AMCOS, PRS, GEMA, ASCAP and their counterparts — exist precisely to do what the carve-out lobby calls impossible: to license use collectively, collect the money, and distribute it to the right people. A blanket AI-training licence administered by these bodies, with funds flowing into a royalty pool distributed by usage and representation, is not a fantasy but the obvious extension of an existing model. The compulsory mechanical licence is another precedent: a statutory rate that lets anyone reproduce a published composition provided they pay the set fee. A statutory training licence, set fairly and administered collectively, would marry the certainty the technology industry says it wants with the payment the creator is owed. A statutory licence still compensates; the carve-out does not. One is a price, the other an exemption. And, as the Indigenous material makes plain, even a well-designed compensation mechanism must sit downstream of consent rather than in place of it.
Fourth, enforcement. A right that cannot be enforced is the right the independent musician already has, which is to say no right at all. A framework that left enforcement to private lawsuits would protect only the wealthy; one empowering collective-management organisations to police compliance, audit datasets and pursue violations for their members would extend protection down to the artist with twenty songs and no lawyer. The arXiv finding that distributor policies are largely unenforced, and detection unreliable, is a finding about the enforcement gap. Closing it requires both technical provenance and an institution with the standing to act.
None of these four is novel. Consent, transparency, collective compensation and enforcement are the pillars of every functioning rights system the music industry has built. The training-data regime does not confront the industry with an unprecedented problem, but with an old one — unlicensed mass use of creative work — at a scale the existing tools already, in principle, exist to address. The obstacle is not capability. It is will.
That will is no longer entirely absent. The Australian commitment of July 2026 moves toward something close to these four requirements: a statutory framework built on the artist's control rather than their failure to object, drafted by a dedicated office of state, and headed for Parliament rather than another consultation. Whether the bill that reaches the floor in 2027 retains that shape is an open question, and the lobbying that produced the carve-out campaign will be brought to bear on the drafting. But the claim that opt-in licensing is administratively impossible has now been contradicted by a government intending to legislate it, and the claim that a pooled fund is an adequate substitute has been put to that government and refused.
One further requirement sits above the other four, and it is the hardest: coordination across borders. Music is global, training is global, and the datasets Reisner found respect no jurisdiction. A robust opt-in regime in one country is worth little if a developer can ingest the same catalogue under a permissive exception in another and deploy the model everywhere — which is why the lobbying targets jurisdictions one at a time, probing for the weakest link. Protecting creators would require the collective-management organisations, which already coordinate through reciprocal agreements, to extend that cooperation to training rights, and governments to resist undercutting one another to host the data centres. That infrastructure is not hypothetical; it has paid songwriters across borders for decades. The question is whether the will exists to point it at this problem before the precedent of the unlicensed taking hardens into something that cannot be reversed.
The Choice Being Made
Strip away the technical vocabulary and the lobbying euphemisms, and what remains is a question of which principle a society chooses to uphold. Either creative work is property its makers control and are paid for, or it is a free resource lying around for whoever has the computing power to scrape it. The text-and-data-mining carve-out is not a compromise between these positions. It is a vote for the second, dressed in the language of the first.
The twelve-million-track dataset that would take ninety-one years to hear is a monument to a category error: the treatment of a century of human creative labour as though it were ambient data, like weather readings or traffic patterns, free for the taking because it happens to exist. It is not ambient data. Every track was made by someone who decided how the chorus should resolve, which take to keep, what the song was about. Each of those decisions is what copyright was built to protect, and each was overridden the moment the work was copied into a dataset without a word to its author.
The remedy is not to ban the technology, nor to pretend the machines can be uninvented. It is to insist that the principle which has governed creative work for a hundred years survives the transition: that you ask first, that you say who you took it from, that you pay, and that the asking, crediting and paying can be enforced by institutions strong enough to act for those too small to act for themselves. APRA AMCOS put it as plainly as it can be put. The songs are not bargaining chips. They are someone's life's work. A framework worthy of the name would refuse the one thing the technology industry is asking for above all others: permission it never sought, granted in advance, for a taking that has already happened.
The unasked permission is the whole of the matter. Restore the asking, and the rest follows.
References
- The Atlantic / Music Business Worldwide, “Four music datasets holding millions of tracks are being shared among AI developers, The Atlantic reports”, 16 June 2026. https://www.musicbusinessworldwide.com/four-music-datasets-holding-millions-of-tracks-are-being-shared-among-ai-developers-the-atlantic-reports/
- Exclaim!, “Backxwash, Titus Andronicus Among Musicians to Find Their Songs in AI-Training Datasets Exposed by 'The Atlantic'”, June 2026. https://exclaim.ca/music/article/backxwash-titus-andronicus-among-musicians-to-find-their-songs-in-ai-training-datasets-exposed-by-the-atlantic
- MusicTech, “A massive music dataset is allegedly allowing AI to train on 12 million+ songs without permission”, June 2026. https://musictech.com/news/industry/ai-12-million-dataset-training/
- Stanley Wu, Josephine Passananti, Viresh Mittal, Wenxin Ding, Haitao Zheng, Ben Y. Zhao, “An Empirical Analysis of AI Slop in Music Streaming”, arXiv:2606.18052, 16 June 2026. https://arxiv.org/abs/2606.18052
- Deezer Newsroom, “AI-generated tracks now represent 44% of all new uploaded music”, April 2026. https://newsroom-deezer.com/2026/04/ai-generated-tracks-represent-44-of-new-uploaded-music/
- Music Business Worldwide, “Spotify has deleted 75m+ tracks in 'spammy' AI music crackdown”, September 2025. https://www.musicbusinessworldwide.com/spotify-has-deleted-75m-spammy-tracks-as-it-unveils-new-ai-music-policies/
- DLA Piper, “Training AI models: UK Government proposes EU style 'opt out' copyright exception”, 2025. https://www.dlapiper.com/en-us/insights/blogs/mse-today/2025/training-ai-models-uk-government-proposes-eu-style-opt-out-copyright-exception
- Reed Smith, “UK copyright and AI report: the 'opt-out' is dead, but what comes next?“. https://www.reedsmith.com/articles/uk-copyright-and-ai-report-the-opt-out-is-dead-but-what-comes-next/
- A&O Shearman, “UK Copyright and AI: Government Drops TDM Exception Plans”, 2025. https://www.aoshearman.com/en/insights/ao-shearman-on-tech/where-are-we-on-copyright-and-ai-in-the-uk
- Attorney-General's Department / APRA AMCOS, “Australian Government rules out Text and Data Mining exception”. https://www.apraamcos.com.au/about-us/news-and-events/tdm-response
- The Music (themusic.com.au), “ARIA, APRA AMCOS Heads Respond To Calls For Pro-AI Copyright Overhaul”, 4 June 2026. https://themusic.com.au/industry/aria-apra-amcos-heads-respond-to-calls-for-pro-ai-copyright-overhaul/vsJ80NPS1dQ/04-06-26
- Complete Music Update, “APRA AMCOS says Atlantic investigation proves AI companies have plundered the 'life's work of Australian and New Zealand songwriters'”, 19 June 2026. https://completemusicupdate.com/apra-amcos-says-atlantic-investigation-proves-ai-companies-have-plundered-the-lifes-work-of-australian-and-new-zealand-songwriters/
- APRA AMCOS, “PROOF OF THEFT: Investigation uncovers songs stolen for AI”. https://www.apraamcos.com.au/about-us/news-and-events/proof-of-theft-ai-investigation-the-atlantic
- APRA AMCOS, “AI and Music Report”. https://www.apraamcos.com.au/about/supporting-the-industry/research-papers/aiandmusic
- Billboard, “Australia's Prime Minister Commits to Copyright Protections In AI Age”, July 2026. https://www.billboard.com/pro/australia-prime-minister-commits-copyright-protections-ai/
- Digital Music News, “Australian Gov't Commits to AI Training Protections With New Law”, 15 July 2026. https://www.digitalmusicnews.com/2026/07/15/australia-ai-law/
- 404 Media, “Hack Reveals Suno AI Music Generator Scraped YouTube, Deezer, and Genius”, 15 July 2026. https://www.404media.co/hack-reveals-suno-ai-music-generator-scraped-youtube-deezer-and-genius/
- Music Business Worldwide, “Suno scraped YouTube, Deezer to train AI, hacked code reveals”, July 2026. https://www.musicbusinessworldwide.com/suno-scraped-youtube-deezer-and-genius-to-train-its-ai-hacked-code-reveals/
- MLex, “GEMA, Suno copyright ruling postponed by Munich court to July 31”. https://www.mlex.com/mlex/articles/2481789/gema-suno-copyright-ruling-postponed-by-munich-court-to-july-31
- Fordham Political Review, “The Legal Limits of Creativity: Sampling, Copyright, and the Future of Music Law”. https://fordhampoliticalreview.org/the-legal-limits-of-creativity-sampling-copyright-and-the-future-of-music-law/
- Wikipedia, “Mechanical license”. https://en.wikipedia.org/wiki/Mechanical_license
- WIPO, “Collective Management of Copyright and Related Rights”. https://www.wipo.int/en/web/copyright/collective-management
Tim Green
UK-based Systems Theorist & Independent Technology Writer
Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.
His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.
ORCID: 0009-0002-0156-9795
Email: tim@smarterarticles.co.uk
Listen to the free weekly SmarterArticles Podcast
