Judge demands more detail in music publishers' copyright suit against Anthropic

A setback for music publishers

A federal judge has signaled that music publishers will need to do more work to prove their copyright infringement claims against Anthropic. While Judge Eumi Lee denied a motion by the AI company to throw the case out entirely, she expressed concern over the lack of specific details in the current complaint. The publishers are accusing Anthropic of using pirated materials to train its large language model, Claude.

The lawsuit involves over 20,517 music works that the publishers claim were copied through torrenting. They allege that Anthropic illegally downloaded full copies of music composition books to build its AI capabilities. Because these claims were discovered through a separate class action by book authors, the publishers had to file this specific complaint in 2024.

The publishers include major industry names such as Concord Music Group, Capitol CMG, and Universal Music entities. They are seeking to hold the company accountable for using these pirated libraries. However, the court is currently questioning whether the plaintiffs have provided enough evidence to make their claims plausible rather than just possible.

Arguments over AI guardrails

Anthropic's legal team argued that the company has implemented 'enhanced guardrails' to protect intellectual property. These safety measures are designed to stop the Claude model from outputting copyrighted lyrics when prompted by users. They contend that these protections should be a primary factor in determining whether the model is actually infringing on copyrights in its daily use.

The music publishers countered this by arguing that the model has already 'memorized' the pirated materials. They claim that even if Claude does not keep a searchable dataset of the books, it can still recall the information when a user asks for it. To the publishers, the act of training the model on the pirated data constitutes infringement regardless of the output filters.

The court is currently weighing these two different perspectives on how AI models process information. The judge noted that the specific number of 20,517 works was 'oddly specific' but lacked a clear basis in the current filing. She requested additional briefings to understand how the plaintiffs arrived at that specific number of works.

Liability for company leadership

The court also addressed the role of Anthropic's co-founder and CEO, Dario Amodei, in the litigation. The judge granted a motion to partially dismiss the claims against him, ruling that the publishers failed to show how he personally committed direct infringement. The court found that the publishers did not provide evidence that Amodei personally approved the use of pirated materials.

The publishers argued that a CEO should be held liable for the actions of the company, even if they did not personally click a download button. They suggested that the leadership is responsible for the company's data acquisition practices. However, Judge Lee was not convinced by this logic, stating that every CEO of a small startup would be liable under that specific analysis.

While the direct infringement claims against Amodei were dismissed, he may still face liability for secondary infringement. This type of liability does not require the same level of direct action as the claims the judge dismissed. Anthropic has not challenged the possibility of secondary infringement in this ongoing legal battle.

A history of AI litigation

This case is part of a growing trend of legal challenges against Anthropic as it navigates the complexities of copyright law. The company has been facing legal scrutiny since it first introduced Claude in 2023. It has already dealt with a lawsuit from Reddit over the use of user data and a massive settlement with book authors.

In January 2025, Judge Lee ordered Anthropic to maintain its existing copyright guardrails for both inputs and outputs. This order applies to future models as well, though it did not broadly stop the company from training on copyrighted lyrics. The court is currently overseeing both the 2024 and 2026 cases involving these copyright disputes.

The settlement with book authors in the Bartz v. Anthropic case reached $1.5 billion. This figure represents the largest settlement in the history of U.S. copyright cases. As the music publishers' case continues, the outcome may provide further clarity on how the legal system will treat the intersection of AI training and protected creative works.

What comes next for Claude

The next step in this litigation involves the music publishers providing the specific details the judge requested. They must clarify how the 20,517 works were selected and provide a clearer roadmap of the infringement. The court will then decide if the claims are sufficient to move forward to the next phase of the trial.

The judge also has not yet issued a decision on summary judgment for the 2024 case. This means the legal battle over how Claude handles copyrighted material is still in a critical transition period. The final ruling will likely hinge on whether 'memorizing' data during training is legally distinct from 'using' that data in a final product.

For the music industry, the result of this case could determine how they license content for AI training. If the publishers win, it could lead to stricter requirements for how AI companies source their training data. If Anthropic wins, it may reinforce the idea that guardrails are sufficient to protect copyright holders from AI-generated outputs.

The legal landscape

The court's request for more detail highlights a significant hurdle for copyright holders in the AI era. Proving exactly how a model was trained on specific works can be technically difficult for companies that do not have internal access to the AI's training logs. The publishers are currently tasked with bridging that information gap to keep their lawsuit alive.

This litigation is being watched closely by other rightsholders who are concerned about the use of pirated content in AI development. The fact that the publishers learned of the pirated activity through a separate book author lawsuit suggests a broader pattern of behavior they are trying to address. They are seeking to establish a precedent that prevents the use of pirated libraries for model development.

As the proceedings continue in the San Jose district, the court will serve as a primary venue for defining these new legal boundaries. The outcome will likely influence how other tech companies approach data acquisition and how creators can protect their work from being ingested by large language models. The case remains a pivotal moment for the balance between technological progress and intellectual property rights.

Next steps in the case

The court is currently awaiting additional briefings from the music publishers to address the inadequacy of the current allegations. These briefings must specifically explain the basis for the 20,517 works cited in the complaint. Without this detail, the case may face further hurdles or potential dismissal of specific claims.

Simultaneously, the court is still considering the summary judgment for the 2024 case. This decision will be a major milestone in determining whether the publishers have a viable path to victory. The judge's previous order to maintain guardrails suggests a cautious approach to balancing AI development with copyright protections.

Observers will be looking for whether Anthropic will offer new evidence of their data cleaning processes. If the company can prove that the pirated libraries were successfully purged from the training set, it could significantly weaken the publishers' position. The interplay between training data, model memory, and output filters remains the central tension of this legal battle.

Impact on AI development

For AI developers, this case underscores the importance of transparent data sourcing. If the court finds that training on pirated data is a clear violation, companies may be forced to invest more heavily in licensed datasets. This could increase the cost of developing high-quality models that are trained on diverse and legal materials.

For the music industry, a victory would provide a mechanism to seek damages for the unauthorized use of their catalogs. It would also create a framework for how they negotiate licensing deals with AI companies. The current litigation is essentially a test of whether the 'fair use' doctrine extends to the massive-scale ingestion of creative works for machine learning.

The legal outcome will likely set a standard for how other industries, such as news and literature, approach AI copyright. As the court continues to demand more specific evidence, it is signaling that broad allegations of infringement will not be enough to stop AI companies from developing their products. The focus is shifting toward the specific mechanics of how these models are built and what data they actually retain.

Summary of current status

The music publishers' case is currently in a state of refinement as the court demands more specific evidence. While the case was not dismissed, the publishers must now provide a more detailed account of the 20,517 works they claim were infringed. This puts the burden of proof back on the plaintiffs to show a plausible link between pirated downloads and the model's capabilities.

Anthropic remains on the defensive, maintaining that its guardrails are sufficient to prevent copyright issues. They are positioning their technology as a safe tool that respects intellectual property, even if the training data origins are being questioned. The company's previous multi-billion dollar settlement with authors suggests they are willing to pay significant sums to resolve these types of disputes.

The legal journey is ongoing, with multiple cases moving through the court system simultaneously. The decisions made by Judge Lee in these various proceedings will eventually form the basis of copyright law in the AI age. For now, the music publishers are working to provide the specific facts needed to move their claims from 'possible' to 'plausible' in the eyes of the court.

Key takeaways for the reader

The core issue here is the difference between an AI 'knowing' something and an AI 'outputting' something. The publishers argue that the training process itself is the infringement. Anthropic argues that as long as the user doesn't see the copyrighted material in the final result, the company is following the rules. The court is currently trying to decide which of these arguments holds more weight under current law.

Another key takeaway is the high cost of these legal battles. The $1.5 billion settlement in the book author case shows that copyright holders are increasingly successful in seeking large payouts. This may lead to more aggressive litigation from other creative industries, including music and film, as they look to protect their assets from being used to train competing technologies.

Finally, the court's focus on 'direct infringement' regarding the CEO shows that personal liability for tech executives is a high bar to clear. The publishers must prove that leadership was directly involved in the illegal acts, rather than just overseeing a company that performed them. This distinction will be important for how future corporate AI lawsuits are structured and prosecuted.

The path forward

Moving forward, the music publishers will likely submit a second amended complaint with the specific details requested by Judge Lee. This document will need to clearly explain the origin of the 20,517 works and how they were accessed via torrenting. The quality of this filing will determine if the case proceeds to a full hearing on the merits.

Anthropic will likely continue to emphasize its guardrails and its commitment to safety. They may also seek to argue that the publishers' claims are based on a misunderstanding of how large language models function. The company's goal is to maintain its position as a leader in the AI space while minimizing the legal risks associated with its training data.

The broader legal landscape will continue to evolve as more cases like these reach the discovery phase. The court's decisions will provide a roadmap for how copyright law adapts to the realities of generative AI. For now, the music publishers are in a critical phase of trying to turn their general grievances into specific, actionable legal claims that the court can adjudicate.

Why this matters for AI

This case is a litmus test for the 'fair use' doctrine in the age of generative AI. If the court finds that training on pirated data is a violation, it could fundamentally change how AI companies source their information. It would move the industry toward a model where licensing is the only legal way to train large-scale models on copyrighted works.

For the average user, the outcome could affect the quality and variety of content that AI models can produce. If companies are restricted to only using licensed data, the diversity of the training sets might change. It could also lead to more robust filters that prevent the AI from generating anything that looks like a copyrighted work, even if it was trained on it.

The legal tension between creators and AI developers is reaching a tipping point. This case highlights the difficulty of creating laws for a technology that moves faster than the legislative process. The court's demand for specific details shows that the judiciary is looking for concrete evidence of harm rather than theoretical concerns about the future of creativity.

The role of data in AI

The heart of this dispute is the data used to build Claude. The publishers allege that Anthropic used pirated music composition books to train the model. This highlights a major concern for creators: the idea that their work can be ingested into a system that then competes with them or provides a service that devalues their original content.

Anthropic's defense relies on the idea that the model is a new creation, not a database. They argue that because the model doesn't store the books in a way that users can search, it isn't a substitute for the original works. This distinction between 'storage' and 'training' is a central theme in many current AI copyright lawsuits.

The court's interest in the specific number of works (20,517) shows that the legal system is looking for a clear scope of the infringement. By asking for more detail, the judge is ensuring that the publishers can prove a specific harm rather than a general grievance. This approach will likely be mirrored in other copyright cases involving AI and various types of media.

Corporate liability in AI

The partial dismissal of claims against Anthropic's CEO, Dario Amodei, is a significant legal nuance. It suggests that simply being the head of a company is not enough to make an executive personally liable for every action the company takes. The publishers must prove a direct link between the CEO's actions and the illegal downloading of pirated materials.

This is a common hurdle in corporate litigation. Plaintiffs often try to hold high-level executives accountable to create more pressure on the company. However, the court's ruling indicates that the legal system still requires a high level of specificity to establish personal liability for direct infringement. This may make it harder for plaintiffs to target individual leaders in future AI lawsuits.

While Amodei may still be liable for secondary infringement, the distinction is important. Secondary infringement usually involves helping someone else commit an act of infringement. By dismissing the direct claims, the court is narrowing the scope of what the publishers can achieve against the company's leadership in this specific lawsuit.

The future of music copyright

The music industry is currently in a high-stakes battle to protect its intellectual property from AI. These publishers are trying to set a precedent that will affect how every music catalog is treated in the future. If they succeed, it could lead to a mandatory licensing system for all AI training data.

The fact that these publishers are joining forces suggests a unified front against the current methods of AI development. They are leveraging the findings from the book author class action to build a stronger case against Anthropic. This collaborative approach could be a model for how other creative industries organize to fight for their rights.

As the case moves forward, the music industry will be watching closely to see how the court defines 'infringement' in the context of AI training. The outcome will determine whether the current AI boom is built on a foundation of legal violations or if the current practices fall under protected use. The next few months of litigation will be crucial for the future of music copyright.

Technical guardrails and law

Anthropic's use of 'enhanced guardrails' is a key technical and legal defense. These are software layers designed to monitor the AI's outputs and block anything that violates copyright policies. The company argues that these layers are sufficient to prevent the model from infringing on the publishers' works.

The legal question is whether the existence of these guardrails matters if the training data was obtained illegally in the first place. The publishers argue that the infringement happens at the training stage, regardless of what happens at the output stage. They believe that the model's ability to recall information is enough to constitute a violation of their rights.