Guest column: Music Ally publishes guest columns which voice the opinions of notable authors that advance specific perspectives on important issues. These are chosen at the Editorial team’s discretion and are not paid-for. You can explore the guest column archive here.
This guest post was written by Nick Breen, partner at law firm Reed Smith.
A landmark ruling from a German court last month has added new intensity to one of the biggest questions facing the creative industries: how should copyright apply to AI training? And if anyone was still hoping for a unified global framework, the latest developments suggest the opposite – the world is splintering.
The German decision in OpenAI v GEMA represents one of the first major European judgments to directly address whether training generative models on copyright works requires permission. In this case, the court sided with GEMA and found the training to be more than a neutral, technical preprocessing step, but instead an act of copying that engages authorial rights.
This conclusion stands in stark contrast to other recent decisions in Europe and beyond that have taken a narrower approach, focusing on evidentiary and technical details, such as where the training occurred, whether the model contains or generates any memorised protected works, and whether the system is made available remotely or can be imported as a download. In several of these cases, rights holders’ claims have struggled not because the legal theory was incorrect or uncompelling, but because the factual hurdles were insurmountable.
What we can be sure about is AI copyright cases are extremely unpredictable, and the outcomes can turn sharply on jurisdiction and granular technical facts. A change in model architecture, a different method of memorisation or data retention, or the physical location of servers can determine whether a claim succeeds or fails.
The fragmentation problem
Outside Europe, the divergence widens further. In the U.S., emerging case law and commentary suggest courts evaluate the ingestion of training data primarily through the lens of fair use, asking whether the act of training is transformative or socially beneficial, and/or whether there is measurable market harm. This creates a very different analytical landscape to Europe’s approach.
For developers, this growing fragmentation is more than a legal nuisance, it affects how AI systems are built. Companies may need region-specific training pipelines or separate models for different territories. Some may avoid making models downloadable to limit exposure to importation-style theories of liability. Others are redesigning datasets to minimise memorisation or to ensure auditable, licensed provenance.
This results in copyright uncertainty driving, not merely reacting to, technical architecture.
For rights holders, especially in music, this raises strategic questions; which jurisdictions offer the strongest prospects? Are technical disclosures from AI companies sufficient to determine where and how training occurred? And should enforcement efforts focus on training, outputs, voice models, or all of the above?
“Companies may need region-specific training pipelines or separate models for different territories”
Nick Breen
A European first and possibly a turning point
The German ruling matters because it is the first major EU judgment to confirm that training itself requires permission. If the decision is upheld on appeal, it will likely accelerate a shift that is already happening in practice; licensing, not unlicensed scraping, becoming the foundation of commercial AI development, at least in Europe.
Recent market activity supports this trend. High-profile licensing and settlement deals, including the agreements struck by UMG and WMG with Udio and Suno, alongside Warner Music’s “legislate, litigate, and license” strategy show that the majors are moving rapidly to shape the AI ecosystem rather than wait for courts to do it for them.
Collective management organisations are entering the space too, with both GEMA and STIM publicly announcing AI licensing proposals. These initiatives could pave the way for scalable, blanket permissions covering large portions of repertoires, something individual publishers or labels cannot provide alone. That said, we are yet to see a public announcement of a major licence deal reached between an AI platform and a CMO.
Implications for music and beyond
For the music industry, two developments are likely to define the next year:
- More litigation. Rights holders will continue testing theories across jurisdictions, particularly given how fact-dependent these cases are. Technical evidence, such as the extent of memorisation, the structure of the dataset, or whether the model is hosted or downloadable, will increasingly make or break claims.
- More licensing. Developers want certainty; music companies want to protect catalogue value. As legal risks crystallise, licensed training is becoming a competitive advantage, not a regulatory burden.
Beyond music, the implications are much broader. If courts continue to treat training as a rights-relevant act, AI companies in all sectors will need cleaner, licensed datasets and transparent provenance systems. Smaller developers may find the compliance burden challenging, while those offering rights-cleared or permissioned AI tools may thrive.
Where does this leave us?
We are not heading toward global consensus. Courts, regulators, rights holders, and AI companies are moving in different directions, driven by different legal traditions, stakeholder power and technological assumptions.
Nevertheless, over the last few months, one trend has clearly emerged; the market is shifting toward permissioned AI, either because courts require it, or because commercial reality demands it.
The next year will show how far, and how fast, that shift becomes the norm.
