Claude 'can’t help falling in love with’… Elvis’s lyrics

As one redditor reflected, the ‘floodgates of Graceland’ opened last week when a group of rightsholders related to musical compositions recorded by Elvis Presley commenced copyright infringement proceedings against Anthropic. Referring to themselves as the ‘Elvis plaintiffs’, they allege that (1) Anthropic used these works to train Claude models without authorisation, and (2) that such models generate infringing output.
This Kat is busy picking our her own impersonator for Halloween
Photo by Chris Barbalis on Unsplash

The complaint

Much of this complaint relies upon public filings and decisions concerning Anthropic (e.g. Bartz v. Anthropic – IPKat here and here – and Concord 1). It also refers to media reporting and technical documentation about AI training datasets where Anthropic is alleged to have knowledge of the infringing nature of training Claude models and offering them to the public to generate infringing content pursuant to 17 U.S. Code § 106.

Regarding the former, the plaintiffs assert that Anthropic unlawfully accessed training sources through a variety of data extraction methods. From scraping lyric aggregation sites and streaming platform metadata, to large-scale web-crawl archives (e.g. Common Crawl, Book3, and The Pile), the plaintiffs argue that, alongside pre-training behaviour involving synthetic data and model reinforcement feedback, Anthropic reproduced copies of works and incorporated them into Claude models. They also allege that Anthropic continues to ‘stockpile’ works in an ‘all-purpose’ library akin to Bartz. Lastly, in relation to training AI models, following GitHub, Tremblay and Concord 1, the plaintiffs allege that Anthropic has willingly removed copyright management information (CMI), thus contravening 17 U.S. Code § 1201.

Accordingly, when Claude generates lyric outputs, the complaint alleges that they typically omit songwriter credits, publisher identification and copyright notices. The plaintiffs allege that these outputs are possible because the models ‘draw upon the patterns Claude internalised during training’. They describe users as requesting song lyrics by title, ‘in the style of’ artists whose works form part of the Elvis plaintiffs’ catalogue, and ‘original’ compositions. These outputs, alongside Anthropic’s knowledge of its infringing nature, allegedly support a finding of wilful copyright infringement for reproduction, preparing derivative works, distribution, public performance and display of copyright works without authorisation.

When addressing fair use, the plaintiffs allege that it does not apply to Claude models and refer to Bartz in which the court found that the all-purpose library was excluded.

Comment

The characterisation of generative AI-related activities by the Elvis plaintiffs is far from bland. From the Turing test, to unoriginal monkey-Shakespeare metaphors (see commentary here and here), to an old equestrian metaphor referring to sharp, metal-studded hooves (‘ran roughshod’), the plaintiffs appear keen to paint a picture of absolute carnage inflicted by Anthropic. Despite this, there is lacking precision in this Kat’s view to outline a detailed and strong case of copyright infringement. Perhaps due to their preference for a jury trial, doctrinal concepts appear overly simplified. For example, reproduction is referred to as mimicry, and generated content is said to contain ‘recycled’ copyright works. Traditional copyright justifications also seem to escape the plaintiffs as they reduce copyright to the ability of rightsholders to ‘police the use of copyright works’.

This rhetoric is likely intentional. The cast of plaintiffs is diverse, listing trusts related to heirs of songwriters or publishing companies that co-owned rights with Elvis, to corporate entities that purchased related publishing rights in 2013, amongst others. Unsurprisingly, a consistent story which reconciles all interests is challenging. So, the plaintiff’s story appears to hinge on the mighty foundation of licensability as the backbone of the music publishing industry. While this Kat agrees with the importance of commercial exploitation, the assertion that all these interests have collectively ‘invested their blood, sweat, tears, heartache, elation and experience into their creative works’ is somewhat beguiling.

Although the complaint categorises the infringing activities as either relating to input or output, more broadly, this is about the use of copyright works to train AI models which, as the plaintiffs evidence, has generated substantially similar productions of song lyrics. The licensable market for displaying lyrics online has been challenged by Claude models, and the plaintiffs propose damages up to $150,000 USD per work, over 275 works, totalling to $41.1m USD. And yet, the stacking references to Bartz hint that this is more akin to piracy than the destruction of humanity’s collective creative soul.

There is another part that niggles at this Kat: the plaintiffs explain that Anthropic’s finetuning models were published via HuggingFace. This fact is relied upon to argue that ‘Anthropic knew its models were returning answers containing copyrighted lyrics and creating unlicensed derivative works based on these lyrics’. Mirroring claims in Concord 1, the plaintiffs explain that Anthropic subsequently introduced output filters to prevent Claude from ‘reproducing recognized song lyrics’. The plaintiffs claim that these filters can be circumvented by rephrasing the prompt (e.g. ‘rewrite songs for a “modern audience”). Such output is accompanied by the statement that:
"These renders keep landing very close to the originals: same structure, same hooks reused with light rewording, and full sets of verses/choruses/bridges. That’s edging past ‘inspired by’ into reproducing the copyrighted song, even when the words are swapped out."
Alongside the plaintiffs’ omission to consider fair use beyond training AI models, reliance on this statement leaves little room to consider text-based generative AI ‘remixes’. Alongside this Kat’s recent comments regarding SOCAN v. Suno (IPKat here), we should ask more than whether the output reproduces the lyrics, but to what extent there is lawful creative reuse pursuant to the fair use defence. As a side note, referring to LLM output – which is really a word probability engine – as an indication of knowledge, is also quite alarming.

And so, with Halloween approaching, this Kat ponders whether dressing up as ‘a Claude model in the style of an Elvis impersonator’ might require too much explanation for her non-copyright friends.
Do you want to reuse the IPKat content? Please refer to our Policies section. If you have any queries or requests for permission, please get in touch with the IPKat team.
Print | Share by email

Comments

← Previous Post Next Post →