Publication would change nothing about what the law is, and Lowenstein Sandler concedes that in the same breath as it raises the stakes: a final Part 3 would not be binding law, only a document that could carry persuasive weight with courts and Congress while fair use for generative AI training stays unsettled [6]. Persuasive weight is a price rather than a value, and the place it shows up is in what a defendant will pay to keep an appellate panel from reading an agency's analysis alongside a plaintiff's brief [5].
The draft is not new. The Office put out the pre-publication version on May 9, 2025 [3]; the alert dates the stay denial to June 30 while describing that pre-publication version as issued last year [1][2], which places the denial in 2026 and leaves the text sitting unfiled for roughly fourteen months [1][2]. Promptly after the May 9 release the administration moved to remove Perlmutter [4], which is the cleanest available evidence of what somebody thought the document was worth.
The analysis runs through the four statutory factors of Section 107 [15], and it cuts both ways. Read as a defendant, the first factor is usable: training on a large and diverse dataset may be transformative where the model is used for research, analysis, or other non-substitutive purposes [9]. Read as a plaintiff, the prima facie section is better: assembling training datasets, the copies made for training, and models that memorize and reproduce protected expression are all identified as infringement absent a license or defense [7], with outputs substantially similar to protected works or competing in the same market falling outside the transformative framing [10]. The fact-specific posture the Office adopts [8] is what makes both readings honest.
Where the draft does discriminate is by catalogue. Highly creative or unpublished works weigh against fair use while factual or functional works may weigh in favor [11], and copying entire works generally weighs against, subject to a sufficiently justified purpose [12]. On that arithmetic the same defendant is worth more to a music or image library than to a holder of factual material.
The publication question resolves one of three ways: substantially as drafted, softened, or not at all. The third stays live, because the Court denied a stay and not the case, and the question of who may remove the Register is still in litigation [1][13]. The reading that publication helps plaintiffs fails if the factor-one language survives intact and defendants end up quoting it more often than plaintiffs do, which the current text permits [9]. Worth pricing too is what the government spent to get here: an emergency application to the Supreme Court over one office holder [1], with the report the only policy payload the alert identifies [5]. Tens of thousands of comments went into that draft [14], and the only number that moves if it is finally filed is the discount each side applies to a claim it expects to settle.
