Last week, Tim Hwang, with help from Sam Roland and Joshua Levine, filed an amicus brief regarding Thomson Reuters et al. v. Ross.
Introduction and Summary of Argument
This case presents a narrow question with broad consequences: when a developer copies text at a nonpublic, intermediate stage to train an AI model, and the user‑facing outputs are functional content (here, judicial opinions)—not the copied editorial prose—how should fair use apply? Though the lower court attempted to differentiate this case from new generative AI models, stating explicitly “Ross’s AI is not generative AI,” the analysis the court conducts belies the attempt.2 Mem. Op. at 17–18. Nearly all AI models, whether search, text, vision, or other, undertake the same process of intermediate training and public output. Therefore, the reasoning of this Court—particularly as the first appeals circuit to rule on this issue—will establish an important precedent on all AI fair use. Below, we have provided our view on the most faithful reading of the applicable law in this novel context.
The rule emerging from statute and precedent is clear: factor one favors fair use when (1) the use of the copy at issue is an intermediate step to achieve an end distinct from the plaintiff’s original expressive use, and (2) the taking is no more than reasonably necessary to reach the underlying functional ideas or interfaces. The rule has clear limits that protect the interests of the rightsholder. These limits preserve incentives for expressive works while supporting the continued progress in AI methods that is key to our nation's economic vitality. It does not excuse training or deployment that stores or outputs the plaintiff’s prose (or derivative summaries) to end users; it does not excuse copying beyond what is reasonably necessary where less expressive alternatives would achieve comparable function; and it does not bless downstream products that embed headnote text or replicate West’s selection/arrangement for public display. Those uses would defeat the presumption and create expressive substitution cognizable under 17 U.S.C. § 107(4).
The statute draws two lines. Section 107 separates purpose/character from market effect; 17 U.S.C. §§ 102(b), 103(b) limit protection to original expression, not ideas, facts, methods of operation, systems, or law/taxonomy. Any rule must keep factors one and four distinct. Precedent then supplies the standard. Warhol centers the “use at issue” and presumes factor one weighs against fair use when purposes are the same and the use is commercial—unless there is “other justification.” Andy Warhol Found. for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508, 532–33 (2023). That justification is the settled doctrine permitting nonpublic, proportional intermediate copying to reach functional elements. The justification is derived from a line of cases dealing with functional learning rather than public communication, thereby supplying Warhol’s “other justification.” Factor four then asks only about markets for protected expression and traditional derivatives, not methods or access to public-domain law. Because outputs are opinions only, the relevant market is licensing of or replacement headnote text, and a bare nonpublic “license to be trained on” is not a market for expression; no cognizable harm appears on this record.
This Court should adopt the rule above and evaluate the record accordingly: treat nonpublic, reasonably necessary functional training as satisfying factor one and confine factor-four markets to expression and traditional derivatives.
