IP Law and AI Memo

Copyright

Brief Primer on Copyright Law

Copyright law protects original works of authorship that fixed in a tangible medium. While the Copyright Act lists a number of types of works (e.g., literary works, musical works, dramatic works), the list is non-exhaustive and the subject matter of what can be copyrighted has been construed broadly, so long as the work has the minimal amount of creativity that copyright law requires and was not copied from a pre-existing work. The work must also be fixed in some tangible medium (such as on paper/film or stored in digital form on a hard drive) for more than a transitory time.

The Copyright Act enumerates specific rights that copyright owners hold in a copyrighted work, with the main rights being the right to make copies of the work, the right to publicly distribute the work, the right to publicly display/perform the work, and the right to make derivative works from the work.

Copyright Protection for AI-Generated Works

The Copyright Office has taken the position that AI-generated works cannot be copyrighted. Although the Supreme Court and Courts of Appeals have not had a chance to address the issue yet, at least one District Court has agreed with the Copyright Office and held that AI-generated works cannot be copyrighted. See Thaler v. Perlmutter, No. 1:22-cv-01564, 2023 WL 5333236 (D.D.C. Aug. 18, 2023).

Artist Lawsuit against AI

In Andersen v. Stability AI Ltd. (U.S. District Court for the Northern District of California, No. 3:23-cv-00201), the plaintiffs—artists and illustrators—allege that several generative AI programs infringe on the plaintiffs’ copyrighted works in two main ways:

Using the plaintiffs’ works to “train” the AI, specifically, by making and storing copies of the works.
Creating derivative works based on the plaintiffs’ works in response to prompts submitted to the AI by users (e.g., a user can ask for an image “in the style of” a given artist, and the AI then attempts to produce an image that, while not an exact copy of one of the artist’s existing works, is a new image that mimics the style).

The defendants moved to dismiss the claims, and, on October 30, 2023, the court issued its order granting large parts of the motion and trimming the claims considerably—but notably allowing one of the main theories to move forward: that using the plaintiffs’ works to train the AI (specifically by making and storing copies) infringes on the plaintiffs’ copyrights. The court dismissed the claims based on the second theory regarding derivative works, but allowed the plaintiffs to amend, finding that the plaintiffs need to actually allege that the derivative works are “substantially similar” to the plaintiffs’ original works (which is the main test for copyright infringement).

On November 29, 2023, the plaintiffs filed their first amended complaint (“FAC”). The FAC adds more detail on how the AI programs were developed and operate, including how the plaintiffs’ works are used to train the AI. The FAC now alleges copyright infringement on the grounds that:

The defendants used the plaintiffs’ works to “train” the AI, specifically, by making and storing copies of the works.
The AI programs produce similar (or identical) copies of the plaintiffs’ works when prompted.
The AI programs themselves are both infringing copies of the plaintiffs’ works and infringing derivative works.
The defendants publicly distribute the plaintiffs’ works by distributing the AI programs.

In response to the court’s order on the motion to dismiss, the plaintiffs added allegations that the derivative works are substantially similar to the plaintiffs’ works, and expanded on the copyright theories by adding claims of infringement contending that the AI programs produce infringing copies, and are themselves infringing copies, of the plaintiffs’ works. The defendants have not filed a motion to dismiss yet, but likely will.

Author Lawsuits against AI

Book Authors

Following Andersen, several more lawsuits were filed against AI programs (all in the U.S. District Court for the Northern District of California as well): Tremblay v. OpenAI, Inc. (Case No. 3:23-cv-03223); Silverman v. OpenAI, Inc. (Case No. 3:23-cv-03416); Kadrey v. Meta Platforms Inc. (Case No. 3:23-cv-03417); and Chabon v. Meta Platforms Inc. (Case No. 3:23-cv-04663). These lawsuits involve claims made by authors against ChatGPT (OpenAI’s large language model (“LLM”) AI program) and LLaMA/Meta AI (Meta’s LLM AI program) that are essentially the author-counterpart to the claims made by the artists and illustrators in Andersen. The plaintiffs in these lawsuits allege that the AI programs infringe on the plaintiffs’ copyrighted books in the following ways:

The AI programs were trained on the plaintiffs’ books, which involved making and storing copies of the books.
The defendants have made the datasets containing the plaintiffs’ works available to others.
The output produced from the AI programs in response to users’ prompts are derivative works.
The AI programs themselves are infringing derivative works, as the programs cannot function without the information extracted from the plaintiffs’ works and retained inside the programs.

The Tremblay and Silverman cases have been consolidated (with Tremblay as the lead case), and the Kadrey and Chabon cases were consolidated (with Kadrey as the lead case).

The defendants have moved to dismiss in each lawsuit. Interestingly, the motions did not challenge the claims alleging infringement by training the AI programs on the plaintiffs’ works or distributing the datasets containing the plaintiffs’ works. In addition, although the motion to dismiss in Kadrey challenged the claim that the AI program itself is an infringing derivative work, the motion to dismiss in Tremblay did not.

On November 20, 2023, the court in Tremblay granted the motion to dismiss, dismissing the claims that the AI program itself is an infringing derivative work and that all output from the AI program is an infringing derivative work. The court rejected the arguments that, just because the plaintiffs’ works were used to train the AI program, means that the AI program itself and all output are derivative works, and required the plaintiffs to specifically describe any output that they contend is infringing and show that the output is similar to the plaintiffs’ works. As mentioned above, the defendants did not move to dismiss the claims that creating and using copies of the plaintiffs’ works to train the AI constitutes infringement, and so those claims will proceed.

On December 11, 2023, the Kadrey plaintiffs filed their first consolidated amended complaint which contains a single cause of action for copyright infringement based only on the creation and distribution of copies of the plaintiffs’ works when the defendants made copies for the datasets used to train the AI program (which datasets were also made available to others).

The plaintiffs in Tremblay, Silverman, and Kadrey are represented by the Joseph Saveri Law Firm and Butterick Law. The plaintiffs in Chabon are represented by Ventura Hersey & Muller, LLP.

The New York Times

The New York Times filed a lawsuit against OpenAI and Microsoft on December 27, 2023, in the U.S. District Court for the Southern District of New York (Case No. 1:23-cv-11195). The claims are similar to those in the Book Author lawsuits, and contend that the defendants infringe on the plaintiff’s copyrighted articles by:

Making and storing copies of the works for training datasets for the AI programs.
Storing and reproducing the AI programs themselves (since the programs “memorized” the works).
Distributing output from the AI programs that include copies and derivatives of the works.

The infringement theories are similar to those raised in the Tremblay and Kadrey lawsuits. The defendants have not yet filed a response, but will likely file a motion to dismiss.

The plaintiff is represented by Susman Godfrey L.L.P.

Music Publisher Lawsuit against AI

In Concord Music Group, Inc. v. Anthropic PBC, a group of major music publishers filed a lawsuit against Anthropic’s AI program in the U.S. District Court for the Middle District of Tennessee (Case No. 3:23-cv-01092). The publishers allege that Anthropic’s AI (“Claude”) infringes on the music publishers’ copyrighted songs because:

The AI is trained on the lyrics of the publishers’ songs, meaning copies of the lyrics are made and stored.
The AI will respond to users’ prompts by generating identical, or nearly identical, copies of the publishers’ lyrics, even if the AI is not specifically asked to do so (e.g., when a user asks for the AI to write a song about a certain topic or in the style of a certain artist).

The plaintiffs have moved for a preliminary injunction to stop the alleged infringement, while the defendants filed a motion to dismiss based on a lack of personal jurisdiction and improper venue, or alternatively to transfer venue to the Northern District of California; the motion to dismiss does not seek to dismiss any claims on the merits/substantive grounds. Both motions remain pending.

The plaintiffs are represented by Riley & Jacobson PLC, Oppenheim + Zebrak LLP, and Cowan, Liebowitz & Latman PC.

Trademark

In Andersen v. Stability AI Ltd., the first amended complaint added claims under the Lanham Act for:

False endorsement under the Lanham Act based on allegations that some of the defendants advertised that users could include the plaintiffs’ names in prompts to produce works in the plaintiffs’ styles.
Vicarious trade dress violation based on allegations that the plaintiffs’ artistic styles constitute trade dress, and the AI programs infringe on that trade dress by producing works imitating those styles when the plaintiffs’ names are used in prompts.

In The New York Times Company v. Microsoft Corporation, the plaintiff asserts a claim for trademark dilution based on allegations that the AI programs fabricate articles purportedly written by the plaintiff and misattributes articles, quotes, excerpts, etc. to the plaintiff.

Patent

The Federal Circuit Court of Appeals has held that an AI program cannot be an “inventor” for purposes of patent law, and so a patent will not issue if the sole inventor identified is an AI program. Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022). However, the court clarified that it was not providing any opinion on whether an invention created with the assistance of AI is patentable.

Right of Publicity

In Andersen v. Stability AI Ltd., the plaintiffs (artists and illustrators) also asserted a claim for infringement of their right of publicity. The plaintiffs originally alleged that the AI uses their names by allowing users to request art in the style of an artist, but, during oral argument, changed the theory to using the artists’ names in advertising and promotion of the AI products. This claim was dismissed because the plaintiffs failed to adequately allege how the named plaintiffs’ names were used by the defendants in connection with advertising, but the court granted leave to amend. The amended complaint made similar allegations, but asserted them as claims under the Lanham Act instead.

Separately, actress Scarlett Johansson has taken issue with Lisa AI’s use of AI to replicate her voice in an advertisement for its AI product. Although it does not appear that any lawsuit has been filed, the use of AI to mimic a person’s voice would implicate the right of publicity.