By Laura Quilter, Copyright & Information Policy Librarian
For the past few years, artificial intelligence (AI), or generative artificial intelligence (GAI), has dominated conversations in higher education, as well as everywhere else: stock exchanges, tech news, courtrooms, and legislative halls. There are scores of active AI litigations, and numerous bills in Congress and legislatures around the world.
UMass Amherst and the Libraries are not immune to AI fever. Librarians have been offering workshops on everything from “what is AI” to using AI in research, as well as assessing how publishers and vendors are using AI in the systems we purchase. As the Copyright & Information Policy Librarian, I’ve been helping faculty assess new AI clauses in publishing contracts, understand monetary settlements from the litigations, ponder student use of AI in the classroom, and assess concerns about the privacy, bias, accuracy, and ethics of AI agents.
Much of the litigation to date around AI systems has focused on copyright law and fair use. Librarians tend to be keen supporters of fair use, which helps ensure that facts, opinions, and other constitutionally protected speech flow freely, and supports educational and library uses. AI “training,” like any kind of research and learning, relies on those same broad understandings of fair use.
Librarians are deeply invested in the scholarly and creative communities. We support authors’ rights, are fans of citation and attribution, and like accurate metadata. AI tools are famous for failing to provide proper citations, or for producing fictitious “hallucinatory” citations, and many authors and publishers feel that AI tools are undermining their work and industry. As such, the library community is carefully watching the AI cases and intervening when appropriate to ensure that library perspectives and values are represented.
A key issue that courts have begun grappling with is whether generative AI systems—ChatGPT, Claude, etc.—infringe copyrights when they scan them to “train” their systems. Very briefly, to “train” an AI system, the system copies various works—books, artworks, whatever—and then analyzes different components of them to form a working model of how those kinds of works are created. Analyzing this article, for instance, an AI would notice that the word “instance” often follows the word “for.” By making many such associations, AI agents can generate a text that follows the same rules and looks (more or less) like a human-authored sentence.
Being able to copy works to analyze them is important for researchers, and also for librarians and library systems, which use those same technologies to generate search indexes, metadata, and all sorts of analyses of texts. Under copyright law, these kinds of uses have long been deemed to be fair uses by courts. U.S. copyright law recognizes that many uses of copyrighted works benefit the public and are not the kinds of uses that copyright grants to rightsholders. Fair use protects quotes, parodies, reverse engineering in software, and full-text indexing, among other uses—so it’s critical that we keep a strong fair use doctrine!
Training AI Tools & Fair Use
In three recent cases, courts looked closely at the use of copyrighted works to “train” AI tools. Looking at these three cases, we can make a few points.
First, judges are grappling with the need to ensure that copyright law doesn’t hinder research and analysis—even when done by or for AI companies. In all three cases, the courts stressed the importance of these fair use principles, reaffirming the basic principal that the copying and analysis needed for “training” is fair use.
Second, judges are worried about competition for authors. In each case, courts expressed concerns about facts that suggested a type of unfair competition for authors or publishers. Was the end product competing with the copyrighted work (Thomson Reuters v. Ross)? Were the source materials unlawful “pirate” copies (Bartz v. Anthropic)? Did the AI produce works that competed with or harmed the market for the original works (Kadrey v. Meta)?
In each case, these “unfair competition” rationales affected the judgment—in Bartz v. Anthropic, they led to a major class action settlement!
Third, judges are more than willing to look at unfair or unethical business practices—what is sometimes called “unclean hands” in legal cases. In each of the three cases, the court made it clear that bad behavior on the part of the AI company could lead to negative outcomes for that company.
Your UMass Amherst librarians will continue to keep an eye on AI tools, the law, and the implications for our readers, researchers, teachers, and students.
For more information, and a deeper dive into the AI cases, fair use, and other issues, visit: websites.umass.edu/copyright.
As an attorney and copyright expert, Laura Quilter works with the campus community to ensure that faculty, staff, and students are comfortable with copyright and other legal issues that affect teaching and research.