The load-bearing instruction in the brief is not the adjective but the sequencing: fair use assessed use by use, internal or intermediate training copies held apart from public-facing outputs, and any output problem cured with output-focused remedies, according to Lowenstein Sandler's read of the filing [7]. Treat that as a damages architecture rather than a doctrine and the stakes get concrete. A training-stage claim scales with the corpus, because every ingested work is a copy and the plaintiff's arithmetic is the size of the library; an output-stage claim scales with the generations a plaintiff can actually put into evidence, and the government's proposed rule is that using copyrighted text to train an LLM is not infringement absent a distinct use that substitutes for protected expression [15]. Those two numbers are not close.
The awkwardness, or rather the more interesting version of the awkwardness, is that the Executive Branch is asking the Southern District for a bright line while the federal office that studied the question declined to draw one: the Copyright Office's prepublication Part 3 report, released in May 2025 under Register Shira Perlmutter, treated training-stage fair use as a fact-specific inquiry turning on the works used, the source of the training material, the purpose of the model, its outputs, and the relevant licensing markets [4][5]. The Administration then tried to remove Perlmutter, and lost its bid to stay the D.C. Circuit's order when the Supreme Court denied the application on June 30, 2026, roughly thirteen months after the report appeared [6][14]. So the court has two federal readings available, one from an Administration whose AI Action Plan came out last year [13], one from an office still headed by the official it tried to displace [6].
The policy warning in the brief repays a second reading, because it argues that judicially imposed licensing requirements for training would slow domestic AI development, strengthen foreign competitors, and concentrate the market among companies able to absorb large-scale licensing costs [10]. The companies able to absorb large-scale licensing costs are, of course, the ones currently defending these suits. The alert's account of that warning carries no figures, naming concentration and foreign competition without pricing either [10], which is the gap for anyone marking publisher claims to this news: a statement of interest advises the court [3], so what moved is a probability, not a cash flow.
My read is that the filing lowers the expected value of training-stage claims and raises the relative value of output evidence and of theories that never touch the training corpus, and that a lab which had booked a training licence as the price of peace now has to defend that line item against its own government's position that the law already permits the use [15]. The counter-thesis sits in the same record: Lowenstein Sandler flags four points as most important heading into summary judgment in OpenAI v. New York Times [12], and a court can accept the characterization of training as pattern extraction rather than expressive substitution [8] and still weigh licensing markets the way the Copyright Office did [5]. The thesis fails if the summary judgment opinion folds the training corpus and the licensing market into a single inquiry, or if publishers keep settling at prices that only make sense when ingestion itself is the injury.
