Anthropic hits back against Universal’s AI copyright infringement lawsuit

Photo credit: Tada Images/Shutterstock

Generative AI company Anthropic has hit back at a copyright infringement lawsuit filed by Universal Music Publishing Group (UMG), Concord Music Group and ABKCO in October 2023.

The California-based AI company called on the court last week to reject a request by the music publishers for an injunction to stop Anthropic from using their copyrighted materials in its AI technology.

Anthropic offered several arguments in its filing, including that the company has already put in place “safeguards” that would prevent the kind of lyric-copying that Universal and the others allege; and that Universal and the other plaintiffs, and not Anthropic, are themselves responsible for the copyright infringement they cite in their case.

As first reported by Reuters, Anthropic also argued that the case “was filed in the wrong court, and plaintiffs’ assertions of irreparable harm are unpersuasive on their face”.

In what has become one of the most closely-watched legal cases surrounding the use of copyrighted material to train AI models, Universal, Concord and AKBCO sued Anthropic in a Tennessee federal court last October, alleging that Anthropic’s AI chatbot, Claude, “unlawfully copies and disseminates vast amounts of copyrighted works – including the lyrics to myriad musical compositions owned or controlled by [the plaintiffs].”

Attorney Matt Oppenheim, who is leading the case on behalf of the publishers, said in a statement: “We are confident in our Motion for Preliminary Injunction, and believe that Anthropic’s infringement should be stopped”.

Anthropic has recently been the recipient of large investments by major tech companies, including Amazon, which put USD $4 billion into the AI firm, and Google, which has so far invested $2 billion into the company. Hence the interest in the case: It pits the world’s largest music rights holder (UMG) against an AI firm backed by some of the world’s largest tech companies.

Universal et al’s complaint alleged that “Anthropic’s Claude… copies and distributes publishers’ copyrighted lyrics even in instances when it is not asked to do so… Indeed, when Claude is prompted to write a song about a given topic – without any reference to a specific song title, artist, or songwriter – Claude will often respond by generating lyrics that it claims it wrote that, in fact, copy directly from portions of publishers’ copyrighted lyrics.

“For instance, when Anthropic’s Claude is queried, ‘Write me a song about the death of Buddy Holly,’ the AI model responds by generating output that copies directly from the song American Pie written by Don McLean, in violation of Universal’s copyright, despite the fact that the prompt does not identify that composition by title, artist, or songwriter.”

The lawsuit includes a lyric sheet of a song called ‘The Day The Music Died’ (a lyric from American Pie) which Claude allegedly described as “a song I wrote.” Most of the song’s lyrics are lines taken directly from American Pie.

The lawsuit appended a list of 500 songs allegedly infringed by Claude, and requested $150,000 in statutory damages per infringement, meaning that Anthropic could be on the hook for $75 million in copyright infringement, and possibly more, depending on the final number of infringed tracks, and given that the lawsuit seeks additional damages on behalf of music publishers.

In November, Universal and the other recording companies followed up the original complaint with a request for a preliminary injunction that would prevent Anthropic’s AI from using their copyrighted lyrics while the case proceeds.

First, “Anthropic should be ordered to implement effective guardrails to prevent output that reproduces, distributes, and displays” the companies’ works, the request stated.

Second, Anthropic should be prohibited from using “existing unauthorized copies or creating new unauthorized copies” of the publishers’ lyrics to train new AI models.

In a response to this request, filed with the US District Court for the Middle District of Tennessee on Tuesday (January 16), Anthropic argued that the motion “should be denied.”

“Anthropic is confident that using copyrighted content as training data for an LLM is a fair use under the law – meaning that it is not infringement at all.”

Anthropic, in a legal motion filed in US federal court

Since the complaint against it was filed, Anthropic “has implemented additional technological guardrails to prevent a reprise of what plaintiffs did,” the response states – meaning that users of Claude today would allegedly not be able to prompt the chatbot in such a way that it would offer up copyrighted lyrics.

These safeguards “are a part of the model and will remain in place indefinitely. As of now and going forward, no one should be able to replicate what plaintiffs did to engineer the facts on which this lawsuit is based,” Anthropic stated. So one of the key reasons for the preliminary injunction is “moot,” the Anthropic document stated.

It also argued that users, in general, would not use Claude in the way that Universal and the others used it to build a case for copyright infringement.

“Existing song lyrics are not among the outputs that typical Anthropic users request from Claude. There would be no reason to: song lyrics are available from a slew of freely accessible websites,” Anthropic said, noting that the plaintiffs themselves stated that they license a number of different websites to publish song lyrics.

The AI company also asserted that “this case was filed in the wrong court,” because anyone who uses the Claude chatbot – including the plaintiffs in this case – agrees to Anthropic’s terms of service, which “entails agreeing to litigate any resulting claims in the Northern District of California.”

“Plaintiffs themselves, not Anthropic, engaged in the ‘volitional conduct’ that is a prerequisite to direct copyright infringement liability…”

Anthropic, in a legal motion filed in US federal court

It also suggested that Universal and the other plaintiffs violated Anthropic’s terms of service, in which users agree “not to request the production” of copyrighted materials.

“What is clear… is that [the plaintiffs’] own expert refers to their efforts as ‘attacks’ on Claude, specifically designed to see if the model could be coaxed to do things everyone understands it is not supposed to do,” Anthropic asserts.

Anthropic also argued that, by prompting Claude to offer up copyrighted materials, it was Universal and the other plaintiffs who engaged in the behavior that led to copyrighted materials being offered up by Claude.

“Plaintiffs themselves, not Anthropic, engaged in the ‘volitional conduct’ that is a prerequisite to direct copyright infringement liability for the outputs plaintiffs’ attacks extracted,” the Anthropic response asserted.

The document also offers some insights into how Anthropic plans to defend itself against charges of copyright infringement – by arguing that its scraping of copyrighted materials to train its AI amounts to “fair use.”

“Anthropic is confident that using copyrighted content as training data for an LLM is a fair use under the law – meaning that it is not infringement at all,” the document says.

“Anthropic does not seek out song lyrics in particular and does not deliberately assign any greater weight to them than to any other text collected from the web… But like other generative AI platforms, Anthropic does use data broadly assembled from the publicly available Internet, including through datasets compiled by third party non-profits for the research community,” the document states.

It claims that “[i]n practice, there is no other way to amass a training corpus with the scale and diversity necessary to train a complex [large language model] with a broad understanding of human language and the world in general.”

The idea that AI models have to be trained on vast amounts of copyrighted materials, and that it is too cumbersome for AI developers to get permission from all copyright holders, is being challenged by many rights holders’ organizations.

Earlier this month, tech entrepreneur Ed Newton-Rex, who resigned from AI developer Stability AI over its copyright policies, launched a non-profit, Fairly Trained, to certify AI models based on their respect for creators’ rights.

Anthropic notes in its response that there are “nearly two dozen” copyright cases currently underway in the US that pit copyright holders against AI companies.

Among those cases are two brought by comedian and author Sarah Silverman, along with other writers, against ChatGPT maker OpenAI and Meta Platforms, maker of the LLaMA AI application, which Silverman and other authors say violated the copyright on their books.

Another group of writers, which includes Game of Thrones author George R. R. Martin and The Firm author John Grisham, have also sued OpenAI for copyright infringement.

None of the other lawsuits pending in US courts have requested injunctions to stop AI companies from using copyrighted material, Anthropic alleged.Music Business Worldwide

Related Posts