Anthropic and Music Publishers Clash Over AI Training and Copyright Fair Use

The case involves music publishers including Universal Music Group, Concord and ABKCO. They accuse Anthropic of copying lyrics from songs linked to artists including Beyoncé, The Rolling Stones and The Beach Boys to train its Claude AI models.

US District Judge Eumi Lee is hearing arguments in San Jose, California, about whether Anthropic's use of the material is allowed under the fair use rule in US copyright law.

The decision could affect dozens of similar lawsuits against AI companies.

Publishers Say Claude Used Lyrics Without Licences

The music publishers say Anthropic copied copyrighted song lyrics without getting licences and then used those works to build AI systems that make money.

They say Anthropic gained an advantage by using protected material without paying the licensing fees that services using song lyrics normally have to pay.

The publishers also say Claude can repeat or create material that is similar to copyrighted song lyrics.

This raises another concern.

The case is not just about whether Anthropic copied material while training its AI model. It is also about whether the system it created can compete with businesses that normally license song lyrics.

The publishers argue that letting AI companies use lyrics without paying could hurt legitimate licensing businesses and reduce the financial rewards for songwriters and copyright owners.

They want the court to reject Anthropic's claim that its AI training is allowed under fair use.

Anthropic Says AI Training Is Transformative

Anthropic argues that training an AI model is very different from simply publishing a song's lyrics again.

The company says the material is part of a much larger collection of data used to teach the model how language works and what patterns appear in it.

According to Anthropic, Claude is not designed to act as a database that replaces individual songs.

Instead, the copyrighted material helps train a system that can perform many different tasks.

This is the main point of Anthropic's fair use defence.

Fair use is a part of US copyright law that sometimes allows people or companies to use copyrighted material without permission.

Courts usually consider several factors, including why the material was used, how much of the original work was copied and whether the new use harms the market for the original work.

Anthropic says AI training changes how the material is used because the model does not simply store works for people to read or listen to in the usual way.

The publishers strongly disagree.

US Courts Are Already Split on AI Training

The case comes at a particularly important time.

US judges have started reaching different conclusions about whether training AI models is allowed under fair use.

In an earlier case involving Anthropic and authors, US District Judge William Alsup said training AI systems could be considered highly transformative, meaning it uses material in a substantially different way.

However, he separately found problems with Anthropic keeping pirated books in a central library when that storage was not directly justified by training its AI models.

That dispute later ended with a $1.5 billion settlement.

In a separate case involving Meta, another judge, Vince Chhabria, raised wider concerns.

He warned that generative AI could eventually fill creative markets with competing content and reduce the economic value of work created by people.

These different views have left AI companies and copyright owners without one clear legal rule to follow.

The case brought by the music publishers could help clarify where the legal limits lie.

Market Harm Could Become the Deciding Issue

One of the most important questions is likely to be whether content created by AI harms existing markets for copyrighted music.

Anthropic says its models do not replace the original works.

The publishers argue that a model trained on song lyrics can create competing material and reduce demand for licensed lyric services.

This difference could become a key issue in the case.

Courts often pay close attention to whether a use claimed to be transformative harms an existing or possible future market for licensing the original work.

The publishers say there is already an established market in which companies pay to use song lyrics.

They argue that if AI developers can avoid paying through that market, the economic harm to copyright owners is direct.

Anthropic responds that creating such a broad licensing requirement could make large-scale AI training almost impossible because developers would need permission to use huge amounts of material.

This is one reason the case could have effects far beyond the music industry.

Ruling Could Shape AI Training Across Industries

The dispute could affect books, journalism, photography, films and other copyrighted material used to train AI models.

OpenAI, Meta, Google and several AI startups are already defending similar lawsuits.

If courts broadly accept the fair use argument, AI companies could continue training models on large collections of data without negotiating licences for every individual work.

If courts reject that broad approach, developers could face higher costs and pressure to make licensing deals before using copyrighted material.

Some companies have already begun making such deals.

Publishers, media companies and entertainment groups increasingly see data used to train AI models as something with commercial value, rather than material that companies should be free to collect from the internet without permission.

The Anthropic case is therefore part of a much bigger question: will copyright law treat AI training as a new kind of use that changes the purpose of the original material, or as a commercial activity that requires permission?

Judge Lee's final decision may not settle the question across the entire United States, but it could become another important legal precedent in this fast-changing area of law.

What this means for you:
The case could affect how much AI companies have to pay to use books, music, journalism and other material to train future models. If courts reject a broad interpretation of fair use, licensed training data could become much more important.