By Jonathan Band
The U.S. Court of Appeals for the Ninth Circuit just issued a decision that on its surface appears to be a straightforward interpretation of the Copyright Act, but likely will have far reaching implications for software interoperability and generative artificial intelligence.
The decision is the latest development in the decade-long litigation between Oracle and Rimini. (See here for a discussion of one of the previous decisions.) Rimini provides support services to customers of Oracle’s PeopleSoft software. When providing these services, Rimini uses Oracle’s products and creates files that can only work with Oracle’s software. In earlier phases of the litigation, Rimini was found to have infringed Oracle’s copyrights. Rimini then changed its procedures and sought a declaratory judgment that its revised practices did not infringe. Oracle counterclaimed for copyright infringement. The district ruled on summary judgment that Rimini infringed Oracle’s right to prepare derivative works by writing files that only interact and are usable with Oracle software, even if the Rimini files did not contain any Oracle expression. Rimini appealed this holding (among others), and the Ninth Circuit has now reversed.
The Ninth Circuit panel found that the district court had adopted an “interoperability” test for derivative works: if a product can only interoperate with a preexisting work, then it must be a derivative. The panel ruled that “neither the text of the Copyright Act nor our precedent supports this interoperability test for derivative works.” The Copyright Act defines derivative work as “a work based upon one or more preexisting works….” The panel acknowledged that “almost every work borrows and must necessarily borrow from other works and uses what was well known and used before.” Accordingly, “focusing only on whether a work is based upon a preexisting work would make the derivative-works definition hopelessly overbroad.”
The panel noted that Congress provided several textual clues limiting the definition, including examples in 17 U.S.C. § 101 “such as a translation, musical arrangement, fictionalization, motion picture version, sound recording, art reproduction, abridgement, [and] condensation….” The panel reasoned that derivative work must have a meaning related to those examples. Although the term “such as” means this list isn’t exhaustive, “to be ‘based upon’ another work requires copying of the kind exhibited in translation, movie adaptations and reproductions. Mere interoperability isn’t enough.” The panel stressed that the examples of derivative work provided by the Act all incorporate the underlying work. Therefore, “Congress’s list of examples suggests that a ‘derivative work’ must be in the subset of works substantially incorporating the preexisting work.” Whether a work is “interoperable with another work doesn’t tell us if it substantially incorporates the other work.”
The panel observed that the incorporation can be “literal,” such as copying substantial portions of copyrighted code. Alternatively, the incorporation can be “nonliteral,” such as the copying of the preexisting work’s total concept and feel. In this discussion, the panel clarified confusing dicta from its decision in Micro Star v. Formgen. The panel explained that new levels of the Duke Nukem videogame were derivative works because they copied the video game’s story, including the plot, theme, dialogue, mood, setting, and characters. The panel compared “the extra game levels to a book version of the game that recasts the central character even though it doesn’t copy pictures or code of the game.”
Turning to the case at hand, the panel stated that the district court determined that Rimini created infringing derivative works because they interact only with PeopleSoft. The panel ruled:
“Without more, mere interoperability isn’t enough to make a work derivative. Both the text of the Copyright Act and our case law teach that derivative status does not turn on interoperability, even exclusive interoperability, if the work doesn’t substantially incorporate the preexisting work’s copyrighted material.”
This holding is not surprising; courts have long required the copying of protected expression to establish the infringement of the derivative work right. The panel’s opinion quotes the Nimmer copyright treatise asserting that “a work is not derivative unless it has substantially copied from a prior work.” But the district court’s interoperability test, had it been affirmed, could have caused enormous damage to the software industry. As EFF and CCIA (and eight other organizations) argued in an amicus brief filed with the Ninth Circuit:
“For decades, software developers have relied, correctly, on the settled view that a work is not derivative unless it is substantially similar to a preexisting work in both ideas and expression. Thanks to that rule, software developers can build innovative new tools that interact with preexisting works, including tools that improve privacy and security, without fear that the companies that hold rights in those preexisting works would have an automatic copyright claim to those innovations.”
For this reason, the brief argued, “the district court’s analysis on this issue, left uncorrected, is dangerous for developers and the millions of consumers that rely on their work.”
The panel opinion has implications beyond preserving interoperability. Copyright owners have claimed that the outputs of generative artificial intelligence (GAI) systems are infringing derivative works, even if they are not substantially similar to works used to train the GAI systems, because they are “based upon” those works. The panel clearly rejects this overly broad interpretation of the meaning of a derivative work. Under such a sweeping interpretation, all new works–not just all GAI outputs–would infringe existing works because they are inevitably based in some manner on existing works. The panel underscored that a work is “based upon” another work for copyright derivative work purposes only if the new work substantially incorporates protected expression from the existing work. Accordingly, the output of a GAI system is not a derivative of the works used to train the system unless the output is substantially similar in protected expression to the training data.
