I saw the film Hamnet immediately after meeting with a neurosurgeon for a second opinion about removing a meningioma, a noncancerous tumor, nestled beside my brain. Around that time, I was immersed in reading about artificial intelligence as part of my research for this article, which I needed to finish before surgery. For professional and personal reasons, therefore, the human brain was on my mind. As AI becomes impossible to ignore, those of us who value academic and creative endeavors are all thinking more about what brains can do that machines cannot, despite the tech industry’s aspirations.
In the July 2025 report Artificial Intelligence and Academic Professions, an ad hoc AAUP committee issued recommendations addressing members’ concerns about the impact of AI on campus and the role of academic workers in institutional decision-making related to educational technology, including intellectual property (IP) rights over course materials. Similarly, AAUP statements on IP, such as the 1999 Statement on Copyright, affirm the rights of faculty members with respect to their institutions. As an AAUP staff editor and writer, I’m equally concerned about how members encounter AI in their professional lives beyond campus in the realm of academic publishing, where IP defines careers, inspires the exchange of ideas, and shapes disciplines and interdisciplinary discourse. My article argues that, in making decisions about AI, academic authors and publishers should reject tech billionaires’ fantasies of a “transhuman” future and maintain their integrity and autonomy as contributors to the collective output of generations of creative, well-trained human brains.
Chloé Zhao’s film adaptation of Maggie O’Farrell’s novel Hamnet was an aesthetic salve for me after the sterility of a medical office and the rush to navigate traffic. Lush scenes of forests and fields interspersed with domestic scenes of simple, handmade items felt restorative. Yet instead of serving as escapism, Hamnet invited connections to my thoughts about brainpower and the human stakes of AI. The film explores themes so classic that they must be well represented in texts AI companies have used to train large language models (LLMs): the fragility of human life and the yearning for immortality. As a fictional rendering of how William Shakespeare and his wife might have responded to the loss of their son Hamnet, it depicts a late sixteenth-century world of humans at the mercy of nature, without the benefit of life-saving technologies: Hamnet’s mother is a healer, but her herbal remedies are powerless against the plague that kills him. Just over a century after the invention of the printing press, the only evidence in the film of a machine-made product is a playbill brought to Stratford, informing Hamnet’s grieving mother about London performances of Hamlet. Through her eyes, we see how Shakespeare has transformed his grief, not by telling the story of his son’s death but by imbuing a different story, of a son avenging his father’s murder, with the emotional weight of personal experience. Hamlet’s final soliloquy has a cathartic effect on Hamnet’s mother—who sees that the memory of her son can live on through a fictional character—and everyone around her, as the other “groundlings” standing near the stage reach out to touch Hamlet after she does. I brought my preoccupations and experiences to the movie theater, and Hamnet suggests that Elizabethan audience members brought theirs to the Globe Theatre.
It’s plausible that AI trained on works like Shakespeare’s play, O’Farrell’s novel, and Zhao’s film could generate other works that imitate the exploration of such timeless themes as grief and immortality. But the emotionally nuanced creative transformation that made these works possible seems far more elusive. There is a long history, examined by Erik J. Larson in The Myth of Artificial Intelligence: Why Computers Can’t Think the Way We Do, of dreaming big dreams about AI that never live up to the hype, in part because human methods of inference are not “programmable.” As Peter J. Denning and B. Scot Rousse have observed, LLMs lack the deeply rooted cultural context of socially acquired language, “common sense,” and human capacities for caring, sharing concerns and commitments, and experiencing “embodied” emotions and individual and collective moods.
We hear optimistic forecasts about medical uses for AI, and companies like Palantir (displaying the slogan “AI-powered automation for every decision” on its website) are pursuing business opportunities in this potentially profitable arena. But the pattern-based predictive logic of AI cannot transform medical studies and patient data into the kind of sensible and sensitive personalized guidance that empowered me to make an informed decision about my surgery. Although I had spent countless hours researching meningiomas online, the “AI overviews” that dominate search engines added to my confusion and uncertainty. The day after I saw Hamnet, my original neurosurgeon shared with me the findings of a newly published academic study on meningiomas. Our discussion about the article and the perspective of another surgeon who had provided a second opinion convinced me to schedule my tumor removal for the earliest possible date. I needed the transformative knowledge and insight of experts familiar with my medical profile and other patients’ experiences to interpret research that indicated delaying surgery could increase risks of an unfavorable outcome.
The concept of transformation is central to US copyright law—which privileges “transformative” uses of copyrighted material in evaluating “fair use”—and emerging case law on AI. It’s worth thinking about what kind of transformation we value as human readers and writers and as beneficiaries of published academic research—particularly as we reckon with piracy in the training of LLMs and the unchecked growth of the AI industry. Considerations about how academic publications enable AI’s transformative processes extend beyond concerns about emotional authenticity important in creative writing to those about intellectual integrity and factual accuracy.
Authors, editors, and publishers will need to make consequential IP decisions—including those about settlements in lawsuits over AI piracy, invitations to enter into licensing agreements with AI companies seeking to avoid future lawsuits, and editorial policies and guidelines to prevent the misuse of AI in academic research and writing. Some individuals and organizations, including scholarly publications and presses, will encounter opportunities to “cash in.” However, their relatively modest financial gains facilitate the disproportionate enrichment of AI companies that use copyrighted material for training LLMs. Even if that use is transformative in the strict legal sense, it fails to effect the kind of transformation that depends on the uniquely human capacities for thinking, feeling, and complex analysis. Academic journals and university presses must also protect IP—by upholding ethical standards and principles of copyright law—and commit to publishing human-authored works.
Unfair Use and Bartz v. Anthropic
The US Copyright Office Fair Use Index provides guidance on the four key factors to consider in evaluating whether the use of copyrighted material is fair. The index lists first the “purpose and character,” advising that “‘transformative’ uses are more likely to be considered fair. Transformative uses are those that add something new, with a further purpose or different character, and do not substitute for the original use of the work.” This understanding of transformation has nothing to do with quality, and so far it has legitimized the output of generative AI, which theoretically draws from enough sources to avoid direct plagiarism. Nonetheless, there is evidence that LLMs retain copies of works on which they are trained and sometimes spit out unaltered (thus plagiarized) portions of text.
The Copyright Alliance has tracked dozens of federal lawsuits filed over claims of AI-related copyright infringement, some later consolidated with other cases. Those most relevant to academic publishing involve “literary works,” a category including fiction and nonfiction texts. Whether a US court might award damages for harms other than initial theft of copyrighted material for training purposes remains unclear, but the focus of IP protection is shifting toward licensing. Whatever AI companies spend on acquiring inputs for training or on damages for materials already illegitimately acquired, we can expect them to recoup costs—and reward venture-capitalist investors—by making AI profitable, whether through subscription models, including contracts with colleges and universities, or advertising revenue from intensified data surveillance.
In June 2025, US District Court Judge William Alsup of the Northern District of California issued a summary judgment in Bartz v. Anthropic, a class-action lawsuit brought by a group of authors against the AI company. Anthropic had created a central digital library with millions of books, consisting of both pirated digital copies and digitized copies of print books, for the purpose of training its Claude LLM. Alsup deemed the use of the books for training—but not the unauthorized downloading of the pirated digital copies—as fair use and “exceedingly transformative.” As is all too common with the projection of human qualities onto AI, Alsup describes the training’s purpose and character in anthropomorphic terms: “Like any reader aspiring to be a writer, Anthropic’s LLMs trained upon works not to race ahead and replicate or supplant them—but to turn a hard corner and create something different.” Aspirations for Claude are in its makers’ minds, but such language attributes creative agency to AI. Anthropic’s name is an adjective that Merriam-Webster defines as “relating to human beings or the period of their existence on earth,” pointing implicitly to the prospect of a posthuman era (after AI takes over?), while Alsup’s idiomatic phrase “turn a hard corner” evokes, intentionally or not, hasty negligence in not slowing down before changing direction. It may be fair use for LLMs to draw on abundant training resources to “create something different,” but that doesn’t mean it will be something better than what humans create painstakingly without the tremendous use of electricity and water that data centers require.
Distinguishing between how copyrighted materials were acquired for training purposes and how they were used, Judge Alsup called for a trial to address the downloading of pirated digital copies as unfair use, but in September 2025 Anthropic agreed on a $1.5 billion settlement in the lawsuit, subject to court approval. After Judge Alsup certified in July the rights holders of around seven million works (later reduced to fewer than half a million works that met the eligibility requirements) as the class for the proposed settlement, Authors Alliance Executive Director Dave Hansen flagged problems with defining the class and notifying potential claimants. On the divergent views of rights holders—including about uses of AI and access to scholarly work for academic purposes rather than monetary gain—he commented, “Authors, publishers, literary estates, and many others (many of whom never asked to be part of this fight) now find themselves represented by three writers—without meaningful input, and possibly without shared interests.” Anthropic ultimately did not use the pirated works for training Claude, but the settlement requires Anthropic to remove them from its central digital library. It remains to be seen whether future cases approach damages differently for pirated works that have been used for training LLMs, even if those cases consider the training transformative fair use. Meanwhile, should wronged parties take what they can get or assert their own interests and priorities?
The district court sent legal notices about the settlement in late 2025 to rights holders with contact information available. A website at https://www.anthropiccopyrightsettlement.com includes a searchable list of pirated works in the settlement class and instructional documents for claimants. Notices outlined four options for class members: (1) filing a claim form and foregoing the right to initiate a separate lawsuit against Anthropic, (2) opting out of the claim and reserving the right to file a separate lawsuit, (3) objecting formally to the settlement yet retaining the option to submit a claim form, or (4) taking no action (while potentially benefiting from the settlement if rights coholders file their own claims). By the time this article appears, the deadlines for all options will have passed, but the third option is intriguing because it allows for objecting to the settlement terms—and articulating reasons the court should not approve the settlement—without having to launch a separate lawsuit and while remaining eligible to receive damages. Objectors could submit documents supporting their arguments and appear at the settlement hearing with or without an attorney. If such an option is available in future settlements with AI companies, this is a way for authors or publishers to have a voice in the litigation process and take principled stands that could influence future updates to US copyright law or its interpretation. Filing a separate lawsuit is expensive and worth pursuing only with a substantially different approach, like that of a group of writers—including investigative reporter John Carreyrou, author of Bad Blood: Secrets and Lies in a Silicon Valley Startup—who collectively opted out of the Anthropic settlement to file a lawsuit (not class-action) against multiple AI companies for repeated acts of reproduction of unfairly acquired works. Anyone eligible to make a settlement claim should be wary of predatorial solicitations urging authors to opt out of settlements and seek larger awards.
Estimated damages per work depend on the number of eligible claims filed and could range from less than $3,000 per work, split by default between authors and publishers (except for “education works,” such as textbooks), to $20,000 (if few claims are filed). For many authors, including academic writers, awards may exceed any income previously received from their books, while they might seem trivial for someone like Maggie O’Farrell, who has eight works in the settlement class, including her 2020 novel Hamnet, a bestseller before the film adaptation. Damage awards will be more lucrative for major commercial publishers like Penguin Random House or HarperCollins, which yield hundreds of results if you search by publisher on the settlement website. Some university presses have equally long lists of books in the settlement class. Any author or publisher who receives a significant amount from this settlement or future settlements might consider directing a portion of the unexpected windfall toward affirming the value of human authorship—for example, by supporting writing programs for young people, purchasing books for libraries, contributing to open-access scholarship initiatives, or advocating for AI-related updates to copyright law. Academics should also unite in solidarity on and beyond their campuses to protect intellectual property and hold the AI industry accountable.
Licensing Deals and Steals
The proliferation of lawsuits against AI companies and anticipation of forthcoming settlements have contributed to a collective rush toward the pursuit of licensing deals. The Copyright Alliance has lists of deals searchable by copyright owner, AI company, or organization at https://copyrightalliance.org/artificial-intelligence-copyright/licensing/. Alliance CEO Keith Kupferschmid characterized the negotiation of licensing deals as the “reining in of the wild west,” bringing “law and order” to the AI industry. Discussing recent deals between major media and AI companies, he acknowledges that they are not perfect but notes optimistically, “As more deals are reached between AI companies and copyright owners, we begin to establish a strong foundation built . . . on compensation, consent, control, and sometimes collaboration.” Those key elements of licensing deals enable rights holders to benefit financially from IP, and profits may be large for commercial content providers. Academic authors and publishers stand to benefit also, but they should consider carefully whether cashing in is the right choice.
Kupferschmid’s Wild West metaphor evokes the land-grab mentality of the AI industry. The US West was won not through some disinterested version of law and order but through consolidation of wealth and power derived from accumulation of private property, often stolen or illegitimately acquired. We still live with that legacy, and ties between Stanford University and Silicon Valley exemplify its relevance for higher education. Journalist Malcolm Harris’s Palo Alto: A History of California, Capitalism, and the World examines how they have symbiotically fostered technological innovation for profitable ends. The 1900 dismissal of Stanford economist Edward A. Ross, who criticized immigrant labor in the railroad industry—a source of founder Leland Stanford’s wealth—provided an impetus for the AAUP’s 1915 founding by, among others, Arthur Lovejoy, one of six Stanford faculty members who had resigned in protest. Although it’s disturbing to learn about Ross’s nativist and ultimately eugenicist views, his dismissal highlighted a need for academic freedom protection from the pressure of corporate interests—coming in this case from university trustee Jane Stanford, Leland’s widow. Ross supposedly said in a class, “A railroad deal is a railroad steal.” These words have an unexpected resonance as we consider whether AI deals in the academic publishing sector might be “AI steals”—at least in the sense of the word steal as a bargain that takes advantage of naivete or need.
Academic authors and publishers are pursuing AI deals in part because AI steals through piracy have contributed to a widespread sense of resignation. Being compensated for past theft of intellectual property (especially for those ineligible for legal damages) or for its probable future theft seems better than nothing. Because Judge Alsup’s summary judgment in Bartz v. Anthropic specified that scanning a purchased print book to make a digital copy is a transformative fair use, one could conclude that AI companies are doing more than they are minimally required to do. They can afford to pay licensing fees (while venture capital keeps flowing), and negotiating deals is cheaper than fighting litigation. There are many lawsuits ahead with uncertain outcomes, and the possibility of changes to US copyright law probably makes licensing deals seem like good investments—or even steals.
While deals reached between major media companies and AI companies may lead to enormous profits, amounts involved for academic publishers and especially authors will be smaller. Anyone wondering about the benefits of handing over IP for training LLMs might conclude that the amounts are not worth it. In July 2025, Inside Higher Ed reported that Johns Hopkins University Press (JHUP) had notified many authors of books from their backlist about plans to begin negotiations for licensing deals with several AI companies, allowing authors to opt out by a specific date; estimates of less than a hundred dollars per licensed book suggest that authors who do not opt out might be swayed more by nonfinancial concerns. JHUP Executive Director Barbara Kline Pope said in a 2024 interview,
I recognize the downsides of AI related to the training of our intellectual property on large language models (LLMs). Sometimes I think we should worry more about obscurity than piracy, but we do need to determine exactly how to ensure we are discoverable through these new tools while remaining financially sustainable and protecting our authors’ work. Is allowing training of our content the right thing to do to fulfill our goal to impact people’s lives with high-quality peer-reviewed work? What happens when it’s “all out there”? And when might it be too late to make a deal with the LLM owners to ensure our authors benefit financially from the use of their work? These are all questions that are intellectually interesting to me and that we are contemplating now.
These thoughtful questions remain, for the most part, unanswered in 2026 after JHUP backlist authors have decided about whether to participate in prospective deals. Scholars considering licensing opportunities may hope to increase interest in their work. However, it’s unclear whether “discoverability” through generative AI platforms will truly benefit authors when their work is “all out there,” with high-quality scholarly research used for training alongside low-quality, outdated, or inaccurate content. Their ideas and language are at risk of being plagiarized, transformed beyond recognition, unidentified by direct citations, or distorted into falsehoods.
The benefits of licensing deals for academic journals seem even less promising. The AAUP recently had an opportunity to participate in deals that JSTOR sought to negotiate with several AI companies, providing for a 60–40 revenue split between publications and JSTOR. We learned that Academe and its predecessor titles had about thirty-nine million words of content archived in JSTOR and that JSTOR, as an archive for many academic journals, would have more negotiating leverage than individual publishers, which might be able to license such content for a mere thousandth of a penny per word. We declined. Although we could have excluded specific volumes or issues from the deal, there was no provision for allowing individual authors to opt out. We also worried that AI misinformation could misrepresent authors’ ideas or AAUP policies, causing harm by having Academe materials “all out there” without adequate consent or control mechanisms.
As academic journals weigh options to license their content, they should consider the interests of the authors whose articles they have published over the years. Depending on the focus of the journals and on disciplinary attitudes toward AI, authors may be swayed by arguments about discoverability or by enthusiasm about the future of AI. Journals could survey authors before making any licensing decisions and determine whether they see AI deals as exciting opportunities or “AI steals.”
Academic Publishing and Human Power
Whether or not academic journals and presses benefit from legal settlements or participate in licensing deals, they should consider updating their editorial policies and guidelines. Editors cannot ignore the increasing likelihood that submitting authors have used AI-powered tools or generative AI in their research or writing, but it’s unsustainable for publishers to bear the burden of detecting and policing inappropriate AI use. The academic publishing sector serves the full range of disciplines, which are defining their own standards for AI use. Academic authors must take responsibility for learning which uses of AI are acceptable in their disciplines (or for specific publications) and for limiting their own use accordingly.
Scholars from many disciplines may agree that there are roles for AI to play in assisting with tedious, well-defined research tasks or number crunching. For example, the medical study that influenced my surgery decision used AI-dependent software (Brainlab and IBM’s SPSS) to analyze data and correlate patient characteristics and tumor traits visible in MRIs with postsurgical classification of meningiomas as benign, atypical, or malignant, but it proposed a scoring model that doctors like mine can apply without AI for calculating risks in clinical settings. Researchers must confirm for themselves that results of data analyses are accurate and then summarize and interpret results for their intended audiences, following the norms of their disciplines.
Now is the time for professional associations and publishers to develop clear policies or guidelines regarding AI and to communicate and implement them effectively. The Committee on Publication Ethics offers resources and membership opportunities for the scholarly publishing community, and associations, journals, and presses with overlapping interests may wish to compare notes about how to address shared concerns about AI.
The AAUP is developing guidelines for its own publications that emphasize the value of the unique voices of members and other contributors, disclosure of any use of generative AI or AI research tools, and authors’ responsibility for their research and writing. Although texts generated by AI might sometimes appear to be well-written—LLMs excel at following the rules of standard English grammar and syntax—featuring them in Academe, Academe Blog, or the Journal of Academic Freedom would defeat one purpose of our publications: to explore higher education issues from the perspectives of faculty members and academic workers.
Reflecting on Hamnet as a model for drawing from personal experience and engagement with others’ ideas to create something new, I urge any academic author, editor, or publisher making AI-related decisions to prioritize the inimitable capacities of human brains. In More Than Words: How to Think About Writing in the Age of AI, reviewed in the winter 2026 issue of Academe, AAUP Center for the Defense of Academic Freedom Fellow John Warner makes an eloquent case for how writing “involves a wonderful kind of difficulty in which our grasp continually falls short of our reach.” Even as AI companies expand their presence in higher education and academic publishing, authors and publishers should ensure that the works we write, publish, and read remain the product of human-powered intellectual and creative transformation.
Kelly Hand is writer/editor in the AAUP’s Department of External Relations and an associate editor for Academe.