The all important question as to whether AI training constitutes fair use under American copyright law was being considered in a US courtroom again yesterday, and finally in the context of music, albeit lyrics. This is the legal dispute between Anthropic and the Universal, Concord and ABKCO music publishers, with both sides currently trying to convince Judge Eumi Lee to grant summary judgements in their favour. 

If AI training is fair use, then Anthropic can make copies of existing works when putting together a training dataset for its AI models without getting permission from any copyright owners, including lyrics owned by Universal, Concord and ABKCO. Dozens of copyright lawsuits filed against AI companies in the US swing on this fair use question, including other music industry cases against Suno and Udio. 

The most interesting part of yesterday’s hearing centred on whether or not Claude outputting new lyrics – having been trained on Universal, Concord and ABKCO’s existing human-created lyrics – dilutes the market value of the original songs. 

Market dilution is one of the key criteria when courts consider whether or not the use of a copyright protected work is fair use. If the use of a work negatively impacts on its market value, that use is less likely to be fair use. Here the question is whether new lyrics generated by Anthropic’s Claude negatively impact on the value of the human-created lyrics copied when training the AI model. 

Speaking for the publishers, lawyer Matthew J Oppenheim insisted that Claude’s lyrical outputs are negatively impacting on the music publishing business and the streaming market, while Anthropic’s refusal to secure licences for its training is also damaging the music industry’s emerging AI licensing market. 

According to Law360, Oppenheim stressed the scale of generative AI outputs, telling Lee that – while Claude has many uses – there is evidence that the Anthropic AI received nearly 120,000 requests for what he called “synthetic lyrics” in a six month period. “What’s crucial is Claude’s scale here”, he added, arguing that AI-generated music “could easily overwhelm the market”. 

On one level that’s a large number of lyrical requests. Except, Anthropic is reported to now have somewhere in the region of 30 million weekly users, and in that context 120,000 requests for synthetic lyrics over six months actually seems quite modest. 

Meanwhile, Judge Lee seemed to want tangible evidence of this market dilution, which hasn’t really been provided as yet as part of this case. And Anthropic’s lawyer, Sonal Mehta, was adamant “there isn’t anything in the record that shows doom for the music industry”.

According to Courthouse News, she added that the publishers have failed to show that Claude is having a negative impact on the music licensing market, or on lyric platforms that work with the music industry like LyricFind and Genius. Meanwhile, Mehta added, Universal has told its investors that AI is “overwhelmingly a net positive” for the music industry..

In the first two generative AI copyright cases to get to trial in the US, both involving book authors, one against Meta and the other against Anthropic, the judges accepted the fair use defences put forward by the two tech companies. But with important caveats. 

In the Anthropic case, the judge said AI training was only fair use if the AI company began with legitimately sourced copies of each work. Anthropic initially used millions of pirated ebooks, including pirated books that contained the music companies’ lyrics. 

With that in mind, Universal, Concord and ABKCO tried to amend their Anthropic lawsuit to hone in on the pirated source materials. But the judge wouldn’t let them make those amendments, forcing the publishers to file a second lawsuit earlier this year, which was not being considered in yesterday’s hearing. 

In the Meta case, the judge said that while he was accepting the fair use defence, he felt such a defence could be defeated on market dilution grounds if better arguments were presented to that effect. Hence why, with the initial music industry v Anthropic lawsuit, the market dilution debate is crucial. 

However, when it comes to proving market dilution, that will probably be easier in the cases involving the record labels and music AI platforms Suno and Udio. 

With the music generation platforms, you can show that tracks that have been generated were uploaded onto the streaming services, where they take a share of the finite streaming royalty pool each month if and when they are streamed. Which in turn means less streaming money for human-created tracks. 

That market dilution is still somewhat indirect – AI-generated tracks are not directly competing one-to-one with the human-created tracks used to train the Suno and Udio models – but it is relatively easy to show how AI-generated music in general is taking income away from human-created music in general. 

The publishers can try to make a similar claim, but with Claude generating written lyrics – which then need to be converted into audio form and placed within a track to be taking money out of the streaming royalty pool – the market dilution is even more indirect. 

Which means with Anthropic and Claude’s lyrics, the publishers’ second lawsuit that hones in on the AI company’s pirated source material might be easier for the music companies to win. 

Whereas in the legal battles with Suno and Udio – where the labels still need to demonstrate that the AI companies stream-ripping music off YouTube constitutes piracy – the market dilution claims may be easier to argue and more important in winning the case. 

But this initial legal battle between the publishers and Anthropic is some way ahead of those other cases, having been filed a year earlier than the Suno and Udio lawsuits, and we await to see what conclusions Lee now reaches based on yesterday’s arguments. 

​