Copyhype

The primary objective of fair use is not to confer a private benefit on those who copy the works of others, but to promote the progress of science and useful arts.¹Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569, 575 (1994) (“From the infancy of copyright protection, some opportunity for fair use of copyrighted materials has been thought necessary to fulfill copyright’s very purpose, ‘to promote the Progress of Science and useful Arts’”). The Copyright Act directs courts to specifically consider “the effect of the use upon the potential market for or value of the copyrighted work,” and more broadly, courts weigh all four factors together, “in light of the purposes of copyright.”²Id. at 578.

These principles are especially important in the debate over generative AI. Frontier model developers consistently invoke the language of public benefits when advocating for expansive applications of fair use to training.³See, e.g., Google, A Pragmatic Approach to AI Governance in America at 18 (2026) (“Generative AI is a powerful tool for human creativity”); Kadrey v. Meta, 3:23-cv-3417, Declaration of Jesse Panuccio in support of plaintiffs’ letter to the court, Exhibit A, pg. 10-11 (N.D. Cal., Mar. 2, 2026) (“Aside from its highly transformative nature, Meta’s alleged use also has substantial noncommercial, nonprofit, and educational (including research) purposes … More broadly, Meta’s investment and open release is contributing to the U.S. economy, the emergence of a new and important industry, and the U.S.’s global leadership of that industry over geopolitical competitors.”); Concord Music Group v. Anthropic, 5:24-cv-3811, Anthropic PBC’s motion for summary judgment and opposition to plaintiffs’ motion for partial summary judgment, pg. 19 (N.D. Cal., Apr. 22, 2026) (“Finding fair use here would therefore foster innovation, expand access to knowledge, and create powerful tools for human creativity—precisely the outcomes copyright law was designed to encourage.”). The language in Google v. Oracle, where the Supreme Court said it “must take into account the public benefits the copying will likely produce” is cited.⁴Google LLC v. Oracle Am., Inc., 593 U.S. 1, 35 (2021); see Concord Music Group v. Anthropic, 5:24-cv-3811, Anthropic PBC’s motion for summary judgment and opposition to plaintiffs’ motion for partial summary judgment at 18. The argument is simple: if AI models generate social benefits, then those benefits should weigh in favor of permitting developers to train their models on copyrighted works without authorization.

This reasoning, however, is incomplete, because it considers only one side of the public benefit inquiry. Copyright’s objective is not to maximize the immediate utility of copied works regardless of the consequences. It is to encourage the continued production and dissemination of knowledge through exclusive rights. Courts therefore cannot evaluate only the benefits that may be generated by AI models while ignoring the costs imposed on the authors and institutions that make those benefits possible. An analysis that measures only potential gains without considering what society loses when the underlying information ecosystem is weakened fails to advance the purposes that justify fair use in the first place.

Wikipedia illustrates this point.

The New York Times recently reported the site faces an “urgent” concern in the AI era.⁵Tiffany Hsu, Wikipedia is battling for the soul of the internet, NY Times (July 5, 2026). “Wikipedia is being exploited to train its own competitors, as A.I. systems hoover up its content to inform chatbots like Gemini, Claude and ChatGPT,” writes technology reporter Tiffany Hsu. “The bots then regurgitate the information, often imperfectly, polluting the information ecosystem that feeds into the encyclopedia.” According to the Wikimedia Foundation, the non-profit organization that supports the site, automated requests from bots associated with AI training now account for nearly one-third of its most bandwidth-intensive traffic. Bernadette Meehan, the organization’s CEO, is quoted as saying, “Our infrastructure is not free, and when scrapers come in and bulk-download, it really takes a toll… There is a literal dollar cost to that behavior.” Compounding the injury, A.I. generated summaries of Wikipedia topics “siphon away potential visitors” from the site, with human page views of the English edition down eight percent late last year compared with the year prior.

Wikipedia is not alone in this regard.

The Confederation of Open Access Repositories (COAR) reported in April [2025] that more than 90% of 66 members it surveyed had experienced AI bots scraping content from their sites — of which roughly two-thirds had experienced service disruptions as a result. “Repositories are open access, so in a sense, we welcome the reuse of the contents,” says Kathleen Shearer, COAR’s executive director. “But some of these bots are super aggressive, and it’s leading to service outages and significant operational problems.”⁶Diana Kwon, Web-scraping AI bots cause disruption for scientific databases and journals, Nature (June 2, 2025), doi: https://doi.org/10.1038/d41586-025-01661-4.

Every publicly available publisher, archive, and repository of valuable content and information must devote resources to maintaining servers, monitoring network traffic, preventing abuse, and ensuring that ordinary users can continue accessing the site. If AI developers are free to expropriate this content and information, they are able to largely retain the economic gains produced by their models, while many of the operational costs associated with large-scale extraction remain with the institutions that produce and disseminate the underlying content. 

This allocation of costs should matter to the fair use analysis. If a rule, by permitting them to use copyrighted work to train generative AI models or by privileging copying when copyrighted works are made publicly available, shifts costs from commercial users onto the institutions that create and disseminate knowledge, those institutions will have fewer resources available to carry out their missions. For nonprofit organizations like the Wikimedia Foundation, every dollar spent addressing industrial-scale scraping is a dollar that cannot be spent improving educational resources, preserving historical materials, supporting volunteer contributors, expanding digital collections, or making information available to new audiences. The cumulative effect is to reduce investment in the very activities copyright seeks to encourage.

This is true whether we’re talking about authors and publishers operating in the commercial space or nonprofit knowledge organizations. Educational institutions, archives, and open-access publishers all operate within financial constraints, and their ability to provide free public access depends, among other things, upon maintaining technical infrastructure that is both reliable and affordable. As AI companies increasingly rely upon these resources to train proprietary commercial models, the burden of financing that infrastructure increases, yet the burden remains on the very organizations responsible for creating the valuable resource.

The irony is that many advocates of expansive fair use for AI training also advocate for the continued viability of the open web and public access. Making information available for free does not remove the costs. Even free and public resources require sustained investments in technology, personnel, maintenance, security, and administration. When commercial AI developers exploit those resources while shifting a substantial portion of the resulting costs back onto the institutions making those resources available, the long-term consequence may be fewer publicly accessible resources rather than more.⁷That’s not to suggest this is the only consequence of expansive fair use on open access. See also Stephanie Decker, The Open Access – AI Conundrum: Does Free to Read Mean Free to Train?, Scholarly Kitchen (Apr. 15, 2025 (“As AI becomes more deeply embedded in academic research practices, it may significantly disrupt knowledge creation and attribution standards, disrupting careers and decontextualising research insights.”).

Copyright provides exclusive rights that enable authors and publishers to tailor how, when, and where they make their work available to the public to accomplish their goals, whether those goals are commercial, non-profit, or some mix of the two. It protects works that are made publicly available on the open web, and it serves as the foundation for open access licenses like Creative Commons, which permit reproduction and reuse for some, but not all, uses under certain conditions.

The fair use inquiry should therefore ask a broader question than whether unauthorized AI training produces useful technology. It should ask whether the legal rule governing that training strengthens or weakens the institutions that sustain our knowledge ecosystem over time. If unrestricted scraping requires organizations like Wikipedia to devote increasing resources toward subsidizing the development of commercial AI systems, courts should recognize that consequence as part of the fair use analysis. Copyright’s objective is not simply to encourage new technologies, but to sustain the conditions under which knowledge continues to be created, maintained, and made available to the public. A conception of fair use that systematically erodes those conditions ultimately works against copyright’s fundamental purpose.