Jack Thompson
October 19, 2026
12 min read
The International Confederation of Music Publishers (ICMP) has told Australian policymakers that artificial intelligence systems have already been trained on songs by Kylie Minogue and Nick Cave, without the artists or their publishers being paid. The warning, reported by outlets including The Guardianoosen copyright rules in exchange for AI investment pledges from global tech firms
ICMP director general John Phelan says the pattern is familiar: companies dangle data-centre and AI investment in Australia, then push for carve-outs that let them train models on copyrighted material without permission. “They’re trying to strong-arm Australian policymakers into ‘if you don’t do this, we won’t invest,’” Phelan said, according to reporting on the dispute. The remark has become the defining line of a fight that now touches songwriters, publishers, Google, and the Australian government all at once.
Don’t miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What ICMP says it found in the AI training data
At the centre of the dispute is a dataset containing more than 12 million songs collected from YouTube videos, which ICMP says includes tracks by Kylie Minogue, Nick Cave, and Hilltop Hoods. The scale of that figure is worth sitting with: 12 million tracks is not a boutique archive, it is close to the entire working catalogue of recorded commercial music scraped from a single video platform and repackaged as machine-learning fuel.
ICMP also says it has supplied evidence that Google Gemini can reproduce lyrics associated with Kylie Minogue, G-Flip, and Nick Cave when prompted. That claim is the sharper of the two, because reproducing lyrics verbatim is a much more direct signal of memorisation than simply appearing inside a training corpus. A model that spits out exact lines from “Can’t Get You Out of My Head” or a Nick Cave verse on request suggests the material was ingested at some point in training, not merely referenced in passing. Our guide to using the Gemini app covers how the assistant handles creative-content prompts more generally.
Google has not issued a detailed public rebuttal of the specific lyric-reproduction claims as of October 19, 2026. The company has previously said its AI products are built with safeguards intended to limit verbatim reproduction of copyrighted text, though ICMP’s examples suggest those guardrails are not airtight when it comes to song lyrics.
Kylie Minogue AI training claims: what is confirmed and what isn’t
Separating verified fact from speculation matters here, because the music industry’s case depends on precision. What is confirmed: ICMP says AI systems have been trained on works by Australian performers including Kylie Minogue and Nick Cave without compensation to the authors, and the organisation has provided examples of Google Gemini reproducing lyrics tied to those artists plus G-Flip. The 12-million-song dataset sourced from YouTube videos is also part of the confirmed record, as is Hilltop Hoods’ inclusion in it.
What is not yet independently verified: a figure circulating that puts the number of Kylie Minogue songs identified in one specific dataset at 940, along with claims that individual tracks such as “Spinning Around,” “Can’t Get You Out of My Head,” and “All the Lovers” were named explicitly. Those details appeared in a report from Star Observer but have not been corroborated across the other coverage of this story. Readers should treat the 940-song figure and the named-tracks list as reported but unconfirmed rather than settled fact.
Similarly, an estimate that creators could lose more than $500 million over four years has been attributed to APRA AMCOS, but it is a projection rather than a recorded financial loss. It is useful as a sense of scale for how the industry is framing the stakes, but it should not be reported as a confirmed damages figure.
APRA AMCOS widens the list of affected Australian artists
ICMP is not alone in making this case. APRA AMCOS, the Australian rights organisation representing songwriters, composers, and music publishers, says it has found works by Kylie Minogue, Midnight Oil, Crowded House, and Lorde inside leaked AI-training datasets, according to coverage from Noise11. That list spans five decades of Australasian music, from Midnight Oil’s 1980s political rock through to Lorde’s 2010s pop catalogue, which undercuts any suggestion that this is a narrow problem limited to one genre or one era of recording.
The breadth of the artist list is itself a strategic move by the music bodies. By naming performers as commercially significant and as culturally embedded in Australian identity as Kylie Minogue and Midnight Oil, APRA AMCOS and ICMP are making it harder for policymakers to treat this as a fringe grievance from a handful of niche acts. These are artists whose songs play on commercial radio, at sporting events, and in pub cover sets across the country.
The Atlantic’s four giant datasets
Separate reporting from The Atlantic reportedly found Australian and New Zealand musical works embedded across four large datasets of songs allegedly used in AI model training. That finding runs in parallel to ICMP’s own research rather than duplicating it, and it broadens the picture from a single 12-million-track YouTube scrape to a wider ecosystem of training corpora that music-rights holders say they never agreed to participate in.
Crucially, The Atlantic’s own reporting includes an important caveat that cuts against the most alarmist readings of this story: the publication reportedly could not identify who had actually used most of the datasets it examined to train a commercial model. Inclusion of a song in a dataset does not, by itself, establish that any particular AI model was trained on it. A dataset can exist, be shared, and even be downloaded without ever being fed into a production model that ships to the public.
That distinction is the crux of why this dispute is legally messier than it first appears. Proving a song sits in a scraped archive is one thing; proving a specific company trained a specific commercial model on that song, and that the model’s output demonstrably derives from it, is a much higher evidentiary bar. ICMP’s Gemini lyric-reproduction examples are significant precisely because they attempt to clear that higher bar rather than relying solely on dataset membership.
Why ICMP calls it strong-arming
Phelan’s “strong-arm” framing is aimed squarely at a negotiating pattern the music industry says it has watched play out across multiple jurisdictions: AI companies and their government-relations teams present copyright-friendly text-and-data-mining exceptions as a precondition for continued or expanded AI investment. Say no to the carve-out, the implicit message goes, and the data centres, the jobs, and the compute capacity go somewhere more accommodating.
That dynamic has particular weight in Australia right now, where the country has spent much of 2026 courting AI infrastructure investment, including large data-centre commitments from US tech giants chasing capacity across the Asia-Pacific region. Our earlier coverage of Google, Apple, and Meta eyeing a combined $25 billion in Australian AI investment shows just how much leverage is on the table if the government treats continued investment and copyright policy as linked issues.
Music industry groups argue that tying investment decisions to copyright exceptions puts an unfair thumb on the scale during a policy process that is supposed to weigh creators’ rights on their own merits. Reporting from EconoTimes indicates an Australian parliamentary committee has urged a copyright review precisely because of this investment-linked pressure, a sign the strong-arming allegation is gaining institutional traction rather than remaining a talking point confined to industry press releases.
Australia’s music industry closes ranks
This is not the first time in 2026 that Australia’s creative sector has pushed back hard against the government’s AI posture. Earlier reporting on how 90 percent of the music industry opposed aspects of the Albanese government’s AI plan showed the scale of discontent building well before this latest ICMP dataset revelation. What’s changed is the specificity: instead of general objections to a policy direction, the industry is now pointing to a named dataset, a named AI product, and named artists.
According to Variety’s Australian edition, the sector has taken a increasingly united stand, with songwriters, publishers, and performing-rights bodies coordinating messaging rather than issuing scattered individual complaints. That coordination mirrors the global pattern among record labels and publishers who have pursued coordinated litigation and licensing campaigns against AI developers over the past two years.
How this compares with past AI copyright fights
This is not the first time an AI developer has faced a reckoning over training data sourced without consent. Anthropic’s $1.5 billion copyright settlement with authors and publishers, which cleared court earlier in 2026, set a precedent that unauthorised use of copyrighted text in AI training can carry a real, enforceable price tag rather than just reputational risk. The music industry’s current push against ICMP’s Gemini findings reads as an attempt to force a comparable reckoning for song lyrics and recordings, rather than books and journalism.
There is also a competitive-dynamics angle worth noting. Chinese AI lab Alibaba has separately denied accusations that its models improperly distilled outputs from Anthropic’s Claude, a dispute that sent Alibaba’s stock sinking roughly 25 percent when the allegations first surfaced. That episode and the current Kylie Minogue dataset dispute are different in substance, one concerns model-to-model distillation and the other concerns raw training-data sourcing, but both illustrate how quickly an AI training-data controversy can move markets and reputations once it becomes public.
What music creators actually want, according to industry research
Research commissioned by APRA AMCOS and conducted by Goldmedia GmbH, cited in a submission to the Australian Productivity Commission, gives a clearer picture of where rank-and-file songwriters and composers actually stand on AI. An estimated 96 percent of surveyed music creators demanded disclosure when their copyrighted works are used to train AI systems, while an estimated 95 percent wanted copyright-holder consent obtained before that training happens. A further estimated 93 percent wanted remuneration, such as through licensing agreements, when their works are used for AI training.
Interestingly, the same research found an estimated 54 percent of the 4,274 surveyed songwriters, composers, and music publishers agreed that AI technology can assist music creators, a reminder that the industry’s objection is not blanket opposition to AI as a creative tool. The complaint is narrower and sharper: creators want disclosure, consent, and payment, not a ban on the technology itself. That research also projected that by 2028, an estimated 23 percent of music creators’ revenues could be at risk from generative AI, representing approximately AUD 519 million in cumulative potential damage, according to the Goldmedia modelling.
Key entities in the Kylie Minogue AI training dispute
What Australian music creators say they want from AI policy
Market impact: investment pressure meets copyright reform
The timing of this dispute is not incidental. Australia is in the middle of a genuine AI infrastructure boom, with tech giants weighing billions of dollars in data-centre and compute investment. That gives both sides an incentive to move fast: tech firms want copyright clarity before sinking further capital into local AI infrastructure, while music publishers want enforceable protections locked in before any text-and-data-mining exception becomes government policy. OpenAI’s reported $42.6 billion government stake offer, which put several rival labs on notice earlier this year, is part of the same broader pattern of AI companies using large capital commitments as leverage in policy conversations well beyond Australia.
For the Australian government, the political calculus is unusually tight. Rejecting a broad AI training exception risks being cast as anti-investment at a moment when competing economies are racing to attract the same data-centre capital. Granting one risks a direct and well-organised backlash from a music and creative industry that has already shown, through the scale of opposition to earlier AI plan proposals, that it can mobilise public pressure effectively.
The legal grey zone: dataset membership versus proven training
One of the more technically important threads in this story is the gap between “a song appears in a dataset” and “a specific commercial model was trained on that song.” The Atlantic’s inability to confirm who used most of the datasets it examined is not a minor footnote, it is the central legal obstacle facing any rights holder trying to bring a claim. Courts generally require evidence that a defendant’s specific model ingested specific copyrighted material, not just that the material existed somewhere in the open AI research ecosystem.
This is exactly why ICMP’s Gemini lyric-reproduction demonstration matters more than the raw 12-million-song figure. A model producing verbatim lyrics on request is circumstantial evidence that training occurred, in a way that dataset membership alone is not. Expect future disputes in this space, in Australia and elsewhere, to increasingly centre on output-based evidence like memorisation tests rather than purely on which songs theoretically sat inside a public dataset.
What happens next: five predictions
- Expect ICMP and APRA AMCOS to publish additional memorisation-test evidence against other AI models beyond Gemini, broadening the dispute from a single company to the wider industry.
- Australia’s parliamentary committee review flagged by EconoTimes is likely to become a flashpoint, with submissions from both music industry bodies and AI companies shaping any eventual text-and-data-mining exception.
- Google will likely face growing pressure to publicly detail what training data Gemini used for music-related content, even if it stops short of a full dataset disclosure.
- More Australian and New Zealand artists will likely be named in leaked-dataset reporting as APRA AMCOS and similar bodies in other countries continue auditing public training corpora.
- Any licensing deal that eventually emerges between major AI developers and Australian music publishers will probably be used as a template reference point in copyright negotiations across other English-speaking markets.
What this means for the AI industry’s data sourcing practices
Beyond the Australian specifics, this dispute adds to a mounting body of evidence that large AI training datasets built from open web and video sources routinely scoop up copyrighted creative work without a clear consent trail. That pattern has already reshaped how some AI developers approach data licensing, with companies increasingly striking direct deals with publishers rather than relying solely on scraped or open-source corpora. The Anthropic settlement set one template for retroactive compensation; the Kylie Minogue case may end up testing whether a similar model can be applied proactively, before a model ships rather than after years of litigation.
For Australian developers and AI companies building on top of foundation models like Gemini, the practical takeaway is to pay closer attention to the provenance of any training or fine-tuning data that includes copyrighted music, lyrics, or audio. As this dispute shows, the line between “publicly available” and “legally clear to use for training” is being tested in real time, and the organisations testing it, ICMP and APRA AMCOS, are showing they have the technical capability to produce concrete reproduction evidence rather than just policy objections.
Frequently asked questions
Did Google confirm that Gemini was trained on Kylie Minogue’s songs?
No. ICMP says it has provided evidence that Gemini can reproduce lyrics associated with Kylie Minogue, G-Flip, and Nick Cave, but Google has not issued a detailed confirmation or denial of training on those specific works as of October 19, 2026.
What is the 12-million-song dataset ICMP is referring to?
Reports describe a dataset containing more than 12 million songs collected from YouTube videos, which ICMP says includes tracks by Kylie Minogue, Nick Cave, and Hilltop Hoods.
Does appearing in an AI training dataset prove a song was used to train a specific model?
No. Reporting from The Atlantic found Australian and New Zealand works across four large song datasets but could not identify who had actually used most of those datasets to train commercial AI models. Dataset membership alone does not establish that a particular model was trained on a particular song.
Which other Australian and New Zealand artists have been named in this dispute?
APRA AMCOS says it found works by Kylie Minogue, Midnight Oil, Crowded House, and Lorde in leaked AI-training datasets, in addition to Nick Cave and Hilltop Hoods cited by ICMP.
Is the claim about 940 Kylie Minogue songs confirmed?
No. That figure was reported by Star Observer but has not been independently corroborated by other outlets covering this story, so it should be treated as unconfirmed.
What does “strong-arming” refer to in this context?
ICMP director general John Phelan used the term to describe what he characterises as tech companies pressuring Australian policymakers by linking continued AI investment to the introduction of more permissive copyright rules for AI training.
Is the $500 million creator-loss estimate a confirmed figure?
No. It is reported as an APRA AMCOS estimate of potential losses over four years, not a recorded or audited financial loss.
What do Australian music creators actually want from AI policy?
Research cited in an APRA AMCOS submission to the Productivity Commission found an estimated 96 percent of surveyed creators want disclosure, 95 percent want consent, and 93 percent want remuneration when their copyrighted works are used to train AI systems.
