UMG’s AI training injunction request shot down by judge in Anthropic lawsuit – but music publishers can now gather more evidence from platform

A federal judge in California has rejected a request from Universal Music Group and other music publishers to block AI company Anthropic from using song lyrics to train its AI models, marking a significant development in the ongoing copyright battle between rights holders and artificial intelligence developers.

In an order issued on Tuesday (March 25), Judge Eumi K. Lee of the U.S. District Court for the Northern District of California denied the publishers’ motion for a preliminary injunction that would have prohibited Anthropic “from using copies of lyrics to compositions owned or controlled by Publishers for future training of Anthropic’s AI models”.

The order is the latest turn in an ongoing legal battle.

UMG, Concord, and ABKCO sued Anthropic in 2023, alleging that the company trained its AI chatbot Claude on lyrics from at least 500 songs by artists, including Beyoncé, the Rolling Stones, and the Beach Boys, without permission.

In her ruling on Tuesday, Judge Lee stated that the publishers had failed to demonstrate “irreparable harm” – a prerequisite for such an injunction.

“If other AI developers are obtaining licenses to use copyrighted material for training purposes, then it follows that… any harm arising from the emerging AI licensing market would be compensable [via damages] rather than irreparable.”

Judge Eumi K. Lee, March 25 order

The court also raised concerns about the scope of the requested injunction, which would have covered not just the 500 works identified in the lawsuit but potentially “hundreds of thousands” of songs.

“[Music] publishers are essentially asking the Court to define the contours of a licensing market for AI training where the threshold question of ‘fair use’ remains unsettled,” Judge Lee wrote in the 13-page order, which has been obtained by MBW (read it in full here).

The court noted that, according to the publishers’ own evidence, “the market for AI training licenses has grown over the course of this lawsuit rather than diminished” – seemingly a reference to companies like OpenAI, which has struck monetized content licensing agreements over the past year with IP owners such as News Corp, Financial Times, and Shutterstock.

The judge said this fact suggests that a commercial licensing framework for AI training already exists (and is growing), meaning that — should UMG and others eventually win their suit against Anthropic — the court would be able to calculate monetary damages rather than requiring injunctive relief at this stage.

The judge emphasized that her latest ruling does not resolve whether training generative AI models with copyrighted materials constitutes infringement or ‘fair use’, which remains “an open question” to be addressed in subsequent proceedings.

COURT MAINTAINS PREVIOUS ‘GUARDRAILS’ RULING WHILE REJECTING TRAINING INJUNCTION

It’s important to note that this ruling only addresses one aspect of the publishers’ legal claims against Anthropic and does not affect a previous favorable ruling for the publishers.

The case effectively has two separate components:

• First, whether Anthropic can reproduce copyrighted lyrics in Claude’s outputs (what users see when interacting with the AI);
• Second, whether Anthropic can use copyrighted lyrics as training data (i.e. inputs) to build its AI models in the first place.

The publishers previously secured a significant victory on the first issue in January 2025, when Anthropic agreed to implement “guardrails” that prevent Claude from reproducing copyrighted lyrics in its responses (i.e. outputs) to users.

That agreement, which was approved by the court, remains intact and wasn’t challenged in this new ruling.

IN BETTER NEWS FOR Universal, COURT GRANTS FAVORABLE DISCOVERY TERMS

In brighter news for UMG and its fellow publishers, the court issued two separate-but-related discovery orders on March 25 – granting Universal and co. significant investigative tools to potentially improve their legal arguments.

Within these orders (read: here), Magistrate Judge Susan van Keulen ordered Anthropic to produce a “statistically significant sample” of Claude prompt and output records from a six-month dataset spanning September 2023 to March 2024.

In simple terms, the court ordered Anthropic to cast a wider net when searching for evidence related to the publishers’ songs.

“the Court ORDERS Anthropic to produce a statistically significant sample of the Claude prompt and output records… At a minimum, the sample must include both pre-suit and post-suit prompts and outputs and must not separate outputs from their prompts.”

Magistrate Judge Susan van Keulen, March 25

The publishers and Anthropic disagreed about how to structure database searches for song titles. For example, if searching for the song “Uptown Funk”, publishers wanted Anthropic to find all instances where the words “lyrics” and “Uptown Funk” appeared in the same conversation with Claude, regardless of how far apart they were.

Anthropic wanted to limit searches to only instances where these terms appeared within five words of each other (which would miss many relevant conversations).

As a compromise, the judge ordered Anthropic to use a “within 20” proximity connector, meaning they must find all instances where the song title and “lyrics” appear within 20 words of each other.

This broader search approach could yield more evidence of potential infringement, likely giving publishers stronger evidence for their case.

CASE PART OF GROWING LEGAL BATTLES OVER AI AND COPYRIGHT

Back in January, UMG welcomed the court’s approval of “guardrails” that require Anthropic to implement protective measures in its Claude AI models to prevent unauthorized reproduction of copyrighted lyrics.

At that time, UMG stated: “We are pleased the Court approved these important ‘guardrails’ — measures which Anthropic must take to protect the Publisher-Plaintiffs’ copyrighted content, by preventing infringing outputs of copyrighted song lyrics – effectively acknowledging the merits of our claims against Claude’s infringing outputs.”

The Anthropic case is one of several ongoing legal disputes regarding the use of copyrighted works to train AI systems. The outcome could have far-reaching implications for both the creative industries and the development of artificial intelligence technology.

Amid the lawsuit, Anthropic secured an additional USD $4 billion investment from Amazon in late 2024, bringing Amazon’s total investment in the AI company to $8 billion.

MBW has contacted Universal Music Group for comment on the latest ruling.

UPDATE (9.10am ET March 26)

The collective of music publisher plaintiffs involved in the case have issued the following joint statement:

“Despite the Court’s narrow ruling, we remain very confident in our case against Anthropic more broadly. In response to our preliminary injunction motion, Anthropic had already conceded the merits of our claims against its infringing outputs of our copyrighted song lyrics, by entering into a stipulation requiring it to maintain “guardrails” to prevent such infringing outputs, thereby resolving a critical aspect of the motion in our favor. The Court noted this factor as an important part of its preliminary decision.

“The Court explicitly declined to decide the fair use question or the merits of our infringement claims at this early stage of the proceeding. Though the Court denied the remainder of our motion, which would have required Anthropic to refrain from training on our lyrics while the case proceeds, it did so on the narrow grounds that our damages for this injury can ultimately be compensated with money damages, which we intend to pursue vigorously.

“This case remains vital to protecting creators from the wholesale theft of their copyrighted works by Anthropic and other AI companies. We expect that, as we proceed with this case and develop a full discovery record, our claims will be validated.”

Music Business Worldwide