The Copyright Fight Starts Before the AI Starts Learning.

The complaint is about more than the model

On August 28, 2026, Sony Music Publishing, Warner Chappell, and affiliated publishers filed suit in the Northern District of California against Anthropic, Dario Amodei, and Benjamin Mann. The complaint concerns musical compositions, including lyrics and sheet music. Its counts distinguish alleged torrent-based copying and distribution, contributory infringement by the individual defendants, other direct infringement by Anthropic, and removal or alteration of copyright management information. It also alleges that Claude reproduces protected lyrics. Anthropic disputes the publishers' claims and says it will defend the case, according to Reuters' August 31 report.

The pleading supplies allegations, not a verdict. Naming executives does not establish personal liability, and examples offered by a plaintiff still need to withstand the evidentiary and legal tests applicable to the claim. But the structure of the complaint exposes a weakness in the public debate: asking whether AI training is fair use can leave several different acts unexamined.

An argument about the destination does not explain the route

A company can have a serious argument that a particular analytical use of copyrighted material is lawful without having an equally strong argument for every way it obtained, stored, or distributed that material. The reverse is also true. Buying a copy does not answer every question about what the buyer may subsequently do with it. Neither provenance nor purpose should be treated as a universal solvent.

There is a concrete judicial illustration. In its June 23, 2025 order in Bartz v. Anthropic, the district court granted Anthropic summary judgment on the training use before it. The court separately addressed the company's central library, distinguishing purchased books converted to digital replacements from downloaded pirated copies. It denied Anthropic summary judgment on the proposition that the pirated library copies should be treated as training copies. The order emphasized the separate uses rather than accepting a single umbrella description.

That is a district-court decision on a particular record, not a nationwide license for AI training and not a determination of the new publishers' claims. Its analytical discipline is nevertheless useful. The unit of legal analysis cannot always be the product. Sometimes it is the copy, the transaction, or the stage in a process.

For counsel, that changes the first conversation. Asking an engineer what the system ultimately accomplishes is necessary. Asking what happens to the material along the way is equally necessary. A persuasive account of a model's social value does not tell a court why a separate archive was created or who received a copy from it.

Fair use is not a product certification

Section 107 directs courts to consider purpose and character, the nature of the work, the amount and substantiality used, and the effect on the potential market or value. It does not supply a box labeled artificial intelligence that resolves those factors in advance. Nor does it make commercial use automatically unlawful or a research label automatically sufficient.

The strongest version of an AI developer's position deserves to be taken seriously. Learning patterns from existing expression can support activities different from selling copies of that expression. Copyright should not become a general right to prevent others from developing useful analytical tools. A licensing demand, standing alone, cannot do all the work of proving that a use requires a license.

The strongest version of the publishers' position also deserves precision. A system that returns protected expression in place of a licensed source presents a different concern from a system that uses material to produce something genuinely different. Whether particular outputs create that problem depends on evidence, including what was requested, what was returned, and which protected work is at issue. Calling all outputs transformative would evade the inquiry as surely as calling all learning theft.

A useful defense therefore needs boundaries. What conduct is being defended? Which facts make the use different? What evidence addresses substitution? And what conduct remains outside the theory? A lawyer who cannot state the limits of an argument may not yet know what the argument actually protects.

The information around the work is part of the dispute

The complaint's copyright-management-information claim adds another layer. The legal issue is not simply whether a model copied expression. Section 1202(b) addresses unauthorized removal or alteration of specified rights information and certain distribution conduct, subject to knowledge requirements. For civil remedies, the statute also requires knowledge or reasonable grounds to know that the conduct will induce, enable, facilitate, or conceal infringement. A missing credit, by itself, does not establish every element.

The statute's definition can encompass information identifying the work, author, or copyright owner, as well as terms of use and certain identifying links or symbols. That makes it hazardous to assume that information considered unhelpful for a technical task is legally unimportant. Engineering relevance and legal significance are different classifications.

This is not an argument that every formatting change creates liability. It is an argument for preserving the ability to explain a change. If a processing step removes text, the useful questions include what category of text it removes, why, what remains associated with the source, and what the people responsible understood about the consequences. Those questions can be answered without pretending the statute has already resolved every dispute over automated processing.

The evidence should follow the conduct

My practical conclusion is that a legal review organized only around the model's name will be incomplete. Counsel should be able to reconstruct acquisition, permissions, transformations, retention, and relevant distribution. That is not a promise that good documentation makes a contested use lawful. It makes the factual question answerable, which is a prerequisite to evaluating either side's legal position.

The same discipline should shape a plaintiff's investigation. An allegation of unlawful acquisition is not interchangeable with proof of infringing output. Evidence supporting one theory may leave another untouched. A case becomes stronger when counsel separates the theories and identifies the evidence needed for each, rather than treating every troubling feature as cumulative proof of a single technological wrong.

For a law firm evaluating an AI vendor, this does not mean demanding a complete training dataset before using a product. It means asking what the vendor actually represents about its rights, what its contractual commitments cover, and what happens if the firm receives a claim. A promise about outputs may say little about the vendor's upstream conduct. A promise about data security answers a different question altogether.

The new lawsuit has not settled who is right. It does illustrate where serious legal work belongs: between the broad claims that the technology is revolutionary and that the copying is obvious. Each side must identify the conduct it is talking about. The model may be one product. Its legal history is not one event.

General legal analysis, not individualized legal advice. The new action is pending; the publishers' allegations and requested relief are not judicial findings or an award.