The legal battle between the music industry and AI giants has reached a fever pitch. BMG, one of the world's largest music publishers, has filed a landmark l...
What the Case Is About
BMG, one of the world's largest music publishers, has sued Anthropic over the use of song lyrics in AI training. The dispute centers on a set of 493 lyrics that BMG says were used without a license. At stake is a simple but unsettled question: when a model is trained on copyrighted words, is that fair use, infringement, or something that needs an explicit deal?
Music publishers control the rights to the underlying compositions—the lyrics and melodies—not just the sound recordings. That makes lyric text especially sensitive. Unlike a casual quote or a review snippet, bulk inclusion in a training corpus can look like wholesale copying of the creative work itself. BMG’s filing frames the 493 examples as proof of that pattern, not a one-off mistake.
Why Lyrics Matter More Than “General Web Text”
Training data often mixes news, forums, code, and other public pages. Lyrics sit in a different category. They are short, highly original, and commercially licensed in almost every other context—karaoke apps, lyric sites, sync deals, and sheet music. Publishers already have markets for that material. Training an AI on the same words without a license cuts across those markets and blurs the line between research and product use.
From the publisher side, the risk is clear: if models can recite or closely paraphrase protected lyrics, they compete with licensed lyric platforms and with the controlled ways fans are supposed to access songs. From the model-builder side, the counterargument is usually that training is transformative learning, not republication, and that outputs can be filtered. Courts have not settled which view wins when the training set includes large amounts of licensed creative text.
What This Means for AI Builders and Rights Holders
Regardless of how this lawsuit ends, the practical takeaways are already usable:
- Inventory training sources. Know whether lyric sites, song databases, or scraped pages that commonly host lyrics appear in your data pipeline.
- Separate research from product. Even if internal experiments use broad crawls, shipping a consumer chat product that can reproduce lyrics raises different legal and brand risk.
- Prefer licensed or synthetic alternatives. Where music text is needed for demos or fine-tuning, licensed subsets or carefully generated non-infringing examples reduce exposure.
- Filter outputs aggressively. Blocking verbatim or near-verbatim lyric reproduction does not erase training-history claims, but it reduces ongoing harm arguments and user-facing copyright risk.
Rights holders, meanwhile, are treating model training as another distribution channel that should be paid for or blocked—not as free raw material. BMG’s move sits inside a wider clash between the music industry and AI companies over who captures value when machines learn from songs.
How to Think About Risk Until the Law Catches Up
There is no single safe recipe yet. Risk depends on what went into the model, what the model can still output, and whether the company monetizes that capability. A defensible approach is documentation plus restraint: log data sources, exclude known high-risk lyric domains when possible, license where feasible, and refuse lyric regurgitation in the product layer.
The BMG–Anthropic fight over 493 lyrics will not answer every copyright question around AI. It does force a concrete choice: treat music text as ordinary internet noise, or treat it as licensed creative property that training must account for. Builders who assume the first will keep getting sued. Builders who plan for the second will be ready if courts and contracts move that way.