EFF to Courts: Rewriting Copyright for AI Training Would Betray Its Constitutional Purpose
The Electronic Frontier Foundation has filed amicus briefs in two active AI copyright cases: Concord Music Group, Inc. v. Anthropic PBC and In re Mosaic LLM Litigation. The core argument in both filings is structural rather than technical: extending copyright liability to cover AI training data does not enforce copyright law, it rewrites it, and rewriting it in this direction would contradict the constitutional purpose copyright was created to serve.
The EFF’s Legal Theory
Copyright exists, under Article I of the US Constitution, to “promote the Progress of Science and useful Arts.” The EFF’s briefs argue that this purpose is served by incentivising the creation of new expressive works, not by granting rights holders indefinitive control over downstream uses of existing works in non-expressive contexts.
AI training involves processing text to extract statistical patterns. It does not reproduce, distribute, or publicly perform copyrighted content. The models that result are not copies of the training material. Courts asked to treat model training as infringement would be extending copyright beyond what the statute says and beyond what the Constitution permits.
The Cases
Concord Music Group v. Anthropic PBC: Music publishers sued Anthropic alleging that training Claude on lyrics and other musical content constitutes copyright infringement. The case focuses on whether ingesting copyrighted text to train a generative model falls under fair use or constitutes direct infringement.
In re Mosaic LLM Litigation: A class action consolidating claims from multiple plaintiffs against AI developers. The consolidated structure makes it a significant vehicle for establishing precedent across the industry, not just for one defendant.
The EFF’s amicus role means it is not representing either party. It is making a brief to the courts on how the law should be interpreted — a common mechanism for public interest organisations to inject constitutional framing into commercial litigation.
Stakes
The litigation landscape around AI training data is now among the most consequential IP questions in the US court system. Anthropic settled with a class of authors for $1.5 billion in July, a case in which training was found to be fair use but long-term storage of training corpora was not. That settlement does not bind the music publishers in Concord, who are pursuing their own theory.
If courts adopt a broad reading that treats any training on copyrighted material as infringement, the impact would extend well beyond Anthropic and Mosaic. Every model trained on internet-scale data — which is every major frontier model — would face retrospective liability under the same theory.
The EFF’s position is that courts should not import that outcome through case law without Congress explicitly choosing it.
The Precedent Gap
US copyright law has not been updated to address AI. The fair use doctrine was designed to handle things like quotation, commentary, and parody — not statistical feature extraction across millions of documents. Courts are currently operating without statutory guidance, and the EFF is arguing they should not fill that gap by expanding liability. They should wait for Congress or hold to existing fair use doctrine.
Whether that argument prevails depends on which courts take it up and how appellate review develops. But the EFF’s filing ensures the constitutional-purpose framing is now formally in the record for both cases.