Can a Machine be Considered an Author? And Other AI Copyright Issues in the Courts

For years, even before the public release of tools like ChatGPT and Midjourney, many experts have speculated about how copyright law should apply to various aspects of generative artificial intelligence. Unfortunately, there has been little guidance from either the courts or Congress on the many copyright issues that generative AI raises. That is, until now. Recently, two different courts published decisions in cases involving AI tools, and they are both worth our time to investigate. First, the federal court for the district of Delaware looked at whether using copyright-protected works to train an artificial intelligence model constitutes copyright infringement or fair use. And second, the Court of Appeals for the District of Columbia looked at whether a creative work produced autonomously by an AI tool was protected by copyright.

Thomson Reuters v. Ross: Training AI and copyright

The first case, Thomson Reuters v. Ross, inc, involves the preeminent legal research tool in the country — Westlaw — and a competing AI research tool developed by Ross Intelligence. Ross contracted with a third party to develop “bulk memos” that Ross uses to train its AI tool. These bulk memos were based on Westlaw’s headnotes, which are annotations on legal issues and rules of law that editors at Westlaw write and attach to nearly every case published in the United States. Headnotes are an extremely powerful research tool and part of the reason subscribers use Westlaw in the first place. Thomson Reuters, the company that owns Westlaw, sued Ross in the federal district court in Delaware, arguing that Ross infringed the copyrights in these headnotes by using them to train the AI tool.  

To prove copyright infringement, a plaintiff must show two things: first that they own a valid copyright in the disputed work; and second that the defendant copied protectable parts of that work. The court in the Ross case started by looking at whether Thomson Reuters held a valid copyright in Westlaw’s headnotes. There is no debate that case annotations in the abstract may be protected by copyright. The issue here was that Westlaw’s headnotes often quote heavily from the text of the case, and that text is in the public domain. Indeed, headnotes may contain little, if any, new text from the editor.  

Even considering this, the court held that Westlaw’s headnotes qualified for copyright protection merely because the editor selected some text over other text. “All headnotes, even any that quote judicial opinions verbatim, have original value as individual works” because “even a headnote taken verbatim from an opinion is a carefully chosen fraction of the whole.” As such, the court argued that Thomson Reuters owns valid copyrights in all of its headnotes.  

The court then reviewed the contested headnotes and the comparable bulk memos and found that Ross Intelligence copied protected parts of 2,243 headnotes with the bulk memos.  

After looking at copyright infringement, the court then considered whether fair use protected Ross Intelligence for using the bulk memos to train its AI model. Courts determine fair use by using a balancing test, weighing four statutory factors against each other. Those factors are: 1. the purpose and character of the use; 2. the nature of the copied work; 3. the amount and substantiality copied; and 4. the harm to the market caused by the copying. Ultimately, the court held that factors 1 and 4 weighed toward infringement, 2 and 3 weighed toward fair use, and all together fair use did not protect Ross Intelligence in this case.  

I have several issues with this decision. First, the idea that headnotes that are copied verbatim from a public domain text can be protected by copyright just because an editor selected some public domain text over other public domain text is inconsistent with how copyright law defines copyrightability. Copyright law protects “original works of authorship,” and to qualify as such, works must satisfy three conditions: 1. they must be “independently created” by their authors — meaning, they can’t be copies of other works; 2. they must be “creative,” but even a little bit of creativity is enough for copyright; and 3. and they must be “fixed in a tangible medium of expression,” or saved in some way, at least momentarily. Headnotes that copy case text verbatim seem at a minimum to fail the “independent creation” requirement of copyright because that text copies another work and thus was not independently created by the editor. Moreover, it also may fail the “creativity” requirement since simply selecting some public domain text over other public domain text does not seem to have even the “modicum” of creativity that copyright requires.  

Turning to the fair use claim, training AI models on copyrighted works seems to fit with a line of cases, including Authors Guild v. Google and Vanderhye v. iParadigms, that have held that copying works to use them as data for a broader system and not for their expressive value is fair use. The court in the Ross case did not clearly distinguish it from any of those other cases. Indeed, somewhat confusingly, the court cited Authors Guild when discussing factor 3 of its fair use analysis, but not factor 4. Why is this confusing? In Authors Guild, the Second Circuit wrote, “The larger the amount, or the more important the part, of the original that is copied, the greater the likelihood that the secondary work might serve as an effectively competing substitute for the original, and might therefore diminish the original rights holder's sales and profits.” That case held that Google’s use of books in its Google Books system did not offer a competing substitute for those books because Google Books only allowed users to view small “snippets” from the books in the database. In the Ross case, the court admitted that users could not access Westlaw’s headnotes through Ross Intelligence’s AI tool. How, then, is it a market substitute for the headnotes? The tool itself certainly can compete with Westlaw, but it is, by admission from the court itself, not truly substituting for the headnotes themselves — the thing for which copyright infringement is claimed. For these and other reasons, I find the court’s fair use analysis unconvincing.

Thaler v. Perlmutter and the copyrightability of AI outputs

The second case, Thaler v. Perlmutter, looked at whether an ostensibly creative work produced autonomously by artificial intelligence — without any input from a human — can qualify for copyright protection.

In 2018, Stephen Thaler applied for copyright registration for a pictorial work called A recent entrance to paradise. In this application, Thaler listed as the author “the Creativity Machine” — an artificial intelligence tool Thaler developed — and himself as the owner of the work. The Copyright Office denied this registration, the Copyright Review Board affirmed the denial, and Thaler eventually sued the Register of Copyrights, Sharia Perlmutter, asking the court to find that this work was protected by copyright. The trial court denied him again, and he appealed to the Court of Appeals for the District of Columbia Circuit.  

The appeals court upheld the district court, finding that the work was not protected by copyright because it did not have a human author, as the copyright statute requires. The court wrote that “reading the Copyright Act to require human authorship comports with the statute’s text, structure, and design because humans have all the attributes the Copyright Act treats authors as possessing.” Looking at various sections of the copyright statute, the court noted that several different parts assume human authorship and are incompatible with machines being authors. For example, transferring a copyright from one person to another requires a signature, but machines cannot sign documents.  

This decision seems exactly right to me. Authorship is core to copyright protection, and it seems clear that copyright law does not permit nonhumans to be “authors.” Copyright exists to incentivize and promote the creation of creative works and machines do not need legal protection over their works to encourage them to create those works. They create because that is what they are designed to do. The court was correct that many sections of the statute clearly indicate that only a human can be an author and don’t make sense otherwise. Moreover, while the court did not address this, the Constitution itself says that Congress may only grant “authors” copyright protection. It seems unlikely to me that the founders of this country envisioned an “author” as anything other than a human being, and certainly not a machine.

What happens now?

These two decisions are not the end of the discussion over how copyright and AI intersect. Ross Intelligence has already announced that it will appeal the Delaware court’s decision, and it seems like there is a good chance Thaler will appeal the Washington, D.C. Circuit’s opinion as well. Furthermore, these cases only touch on two issues, and the court in the Ross case explicitly said that the case was not really about generative AI because Ross's "bulk memos" tool does not create new text. Even so, it is certainly relevant to the conversation about generative AI and copyright, and whatever happens on appeal at the Third Circuit will be very important for how we understand the law in this space. And with over 30 cases pending in federal courts around the country, we should start seeing other decisions on these and other topics in the relatively near future.