The “fair use benefits everyone” amicus parade still turns on whether market dilution is cognizable harm

On 30 October 2026, OpenAI published “Fair use benefits everyone.” The core is not a product ship but a litigation narrative: curated amicus excerpts supporting fair use for generative-AI training, against claims by the New York Times and others. OpenAI bluntly says the suit asks everyone to ignore half a century of fair-use progress for an “undeserved payday” — defendant advocacy, not a court holding.

Excerpts span many actors: the U.S. government warns that excluding training from fair use would hinder research breakthroughs and publishers’ own LLM use; the American Council of the Blind, research-library associations, and others stress disability access, education, and open-source dependents; Authors Alliance attacks “market dilution,” arguing copyright does not protect authors from noninfringing competition at the subject/audience level; Engine and the Developers Alliance emphasize that startups adapt foundation models for ends unlike training-set purposes; scholars such as Matthew Sag argue outputs under realistic conditions are not relevantly similar to intermediate training copies; independent creators say AI cuts operational cost rather than substituting their expression. OpenAI points readers to the public docket for full filings.

The evidence form is checkable legal text fragments, not product metrics. Hardness comes from named institutions and scholars, explicit statutory hooks (17 U.S.C. §107 / §107(4)), and the lineage of Google Books, HathiTrust, and Perfect 10 invoked by Public Knowledge and others. The structural weakness is equally clear: selection, order, and combative framing are controlled by one litigant; readers do not see plaintiffs’ full case on substitutive use, memorization, or licensing markets in this blog; “U.S. government support” is spotlighted while limiting conditions in that brief are compressed.

Steelman: if training does not give the public substitute access to works and outputs do not replace protected expression in the relevant market, placing it in the fair-use tradition (search, accessibility, standards publication) is coherent; treating “new works stealing attention” as copyright harm risks turning copyright into an anti-competition tool. Plaintiffs can still argue that large-scale commercial training changes licensing expectations, that some models emit protected text verbatim or near-verbatim, and that news markets can be eroded by summarization products — points softened rather than resolved here. Coalition breadth proves political mobilization; it does not decide factor four.

For labs and clouds: if courts accept that market dilution is not cognizable, training-data compliance costs reprice; if not, licensing markets and filtering duties may institutionalize faster. For publishers: tying the case to substitution rather than dilution raises the evidentiary bar but stays closer to classic copyright. For disability and library users written as beneficiaries: the policy rhetoric is powerful only if open and commercial systems actually treat accessibility and research tools as first-class — not merely litigation props.

Over the next six months, watch how courts handle §107(4): whether aggregate competitive pressure from generative outputs is cognizable harm, or whether harm requires substitution for protected expression. OpenAI’s “benefits everyone” coalition changes the brief and the public summary; it still does not answer for the judge whether market dilution is the kind of injury copyright law polices.