A federal district court in California is set to hear arguments this week as a group of music publishers sues Anthropic, alleging that the artificial intelligence (AI) firm is stealing their work.
Led by Concord Music Group, the plaintiffs say that Anthropic trained its AI model, Claude, using copyrighted lyrics from their artists. The next hearing is slated for Oct. 21.
In a separate case, Concord alleges that Anthropic used pirated copies of 20,517 music works—accessed from torrent sites—to train Claude.
The lawsuits are just two of dozens of legal actions taken by publishers in the book, news, and art industries, as creatives argue that AI companies have trained their models on existing work with an intent to replace human ingenuity with machine outputs.
When prompted, Claude will create a new set of song lyrics that are similar to existing work, Concord says, or generate a new song using similar chord progressions.
“For example, when asked by Publishers’ investigator to write “a song about the death of Buddy Holly,” Claude generated a verbatim copy of the lyrics to ‘American Pie,’” a court filing says.
The Claude AI tool will also spit out lyrics to copyrighted songs upon request, and Concord argues that this cuts into its profits.
If a web user wants to know the words to “Uptown Funk” or “Hey Jude,” a simple Google search will bring up any number of websites containing the lyrics. But those websites pay a licensing fee to republish the lyrics. Getting the song texts for free through AI makes those websites obsolete, Concord argues, and deprives the publishers of that extra revenue.
For its part, Anthropic is arguing that the Fair Use doctrine allows it to use the song lyrics. That legal principle allows use of copyrighted work under certain circumstances, such as teaching or criticism, or when the copyrighted material is substantially transformed into something new.
Anthropic recently had to settle a case for $1.5 billion after a group of book publishers sued. It used the same fair use argument, but only with partial success.
Although the judge in that case agreed that using copyrighted material to train AI models was fair use, the company had initially used pirated versions of the books taken from the internet. That may end up being relevant for Concord’s second lawsuit.
Anthropic also said that Claude contains “copyright-protective” features and that it isn’t the company’s fault when a user finds a workaround.
“When a system ‘creates the copy only in response to the user’s command,’ its operator does not engage in volitional conduct,” Anthropic said in a court filing. “The user does.”
Concord is asking the U.S. District Court for the Northern District of California in San Jose to award up to $150,000 per copyright violation, along with another $25,000 for instances where it alleges that Claude removed copyright notices from song lyrics. Or else, it at least wants the profits it lost by the infringement—an undisclosed amount Concord said it will later document in court.
At the end of last month, the plaintiffs submitted a short brief noting a recent appeals court decision they hope will galvanize their cases.
On Sep. 30, the Court of Appeals for the Third Circuit ruled against the fair use argument in a similar case where Thomson Reuters sued AI company ROSS Intelligence.
Reuters developed a legal research database called Westlaw, and the entries contain “headnotes,” which are short, hand-crafted annotations that summarize the information. According to a court filing, ROSS hired an outside company to provide material to train its AI tool, and that company used 25,000 Westlaw headnotes for that purpose.
In ruling for Reuters, the Third Circuit said that this was not fair use, partly because the ROSS tool was specifically created to be a competitor to Westlaw.
However, California is covered by the Ninth Circuit, so the court won’t be strictly bound by that precedent when it considers the Anthropic lawsuit.
