DOJ asks the court to judge LLM training separately from model outputs

The April 2025 opinion in this litigation already split the pipeline into acquisition, training and output, and said each stage can raise its own copyright question ^[13]. Those stages are not abstractions to anyone who has shipped a model. Acquisition is crawlers, purchase contracts, and whatever a vendor told you about a corpus. Training is a job that runs once and leaves weights behind. Output is the serving path, where requests arrive and tokens leave.

The government's brief describes training mechanically: copy the works, convert the data into numerical representations, learn linguistic patterns from those representations ^[6]. That description is doing the legal work. It puts the resulting artifact on the statistics side of the line, which is how the filing gets to "highly transformative" and to tasks the original articles never served ^[5].

For that framing to survive contact with a plaintiff, one thing has to be true: the weights must not hand the article back. The brief does not pretend otherwise. It concedes that an output can be its own copyright problem when a model reconstructs and distributes protected material, and argues only that such outputs should not decide the training question ^[2]. The Times has alleged that reconstruction since it sued in December 2023 ^[7].

This is the allocation a builder should want regardless of who wins. You cannot un-train a model, and a dataset audit run after the fact produces a document, not a control. A reproduction test is a control. Write an eval that prompts for known passages, measure what comes back, keep the traces, gate releases on the number. If publishers have to prove infringement at the output level ^[12], the evidence sits in a system you already own and can instrument.

Then count the stages: the court named three ^[13], and the filing argues about only two of them ^[17]. This brief does not address acquisition at all, and the transformativeness of training says nothing about how the bytes arrived. The government itself says fair use stays fact-dependent and that it is not asking the court to excuse every use ^[10].

The licensing section is the loudest part. Pay per work, the brief argues, and only the largest technology companies can afford to train, while publishers with the deepest archives collect the largest payments ^[14]; it calls that a route to oligopoly in model development and a subsidy for established media, and warns that U.S. developers would trail foreign competitors under different IP rules ^[15][16]. An anti-oligopoly argument filed in support of OpenAI ^[8] is a shape worth noticing.

None of it binds the judge, who is free to ignore the statement of interest ^[9]. Two things have to be true before your exposure actually shifts: the court has to adopt the separation, and your serving path has to be able to tell you when it reproduces someone else's text. Building that detection capability is within your control, and it is the cheaper of the two.