After previously dismissing infringement claims on the ground that using copyrighted works to train a large language model qualified as fair use, a federal judge in California has now also dismissed a key Digital Millennium Copyright Act claim in the same case against Meta. The ruling clarifies the boundaries of DMCA liability when the underlying use of copyrighted material is considered lawful under the fair use doctrine.

In addition to the copyright infringement claim, the authors also claimed a violation of Section 1202 of the DMCA, which prohibits the removal of copyright management information from copyrighted works.

The court concluded that the claim could not survive considering the prior fair use ruling, reasoning that if the primary use of the works was lawful, then the removal of copyright management information could not be said to further an infringement. Reading the DMCA to apply in cases of lawful fair use would contradict both the structure and the intent of copyright law.

The court drew a firm line stating that a backdoor through the DMCA that reintroduces liability for noninfringing behavior is incongruous with Congress' intent to protect fair use. The judge explicitly rejected the authors’ reliance on a 2015 Third Circuit case, Murphy v. Millennium Radio Group, noting that it conflicts with Ninth Circuit precedent holding that an infringer cannot be acting willfully if they reasonably believe their use qualifies as fair.  In addition, in Murphy, the underlying use was not protected by the fair use defense.

The outcome reflects an early trend involving generative AI, where the courts appear increasingly willing to view training practices as fair use, especially when the inputs were lawfully acquired. It also signals that ancillary claims like DMCA violations will not get far if they are tethered to conduct already deemed lawful.

For creators concerned with the ingestion of their work by AI systems, this ruling underscores the limits of current copyright law. Unless Congress updates the DMCA or redefines fair use boundaries in the AI context, courts seem likely to continue to shield model training from infringement claims.