Sarah Silverman Ruling Shows ChatGPT Results, Not Inputs, Are Key

Authors, journalists, musicians, and others continue to challenge the use of their creative works as training input for the booming generative artificial intelligence industry, and a decision last month provides guidance about the evidence content creators concerned about copyright infringement should produce.

On Feb. 12, Judge Araceli Martínez-Olguín issued a ruling in a case brought by a group of authors, including comedian Sarah Silverman, against OpenAI, the developer of the popular chatbot ChatGPT, dismissing the plaintiffs’ claim of vicarious copyright infringement. The opinion shows that potential plaintiffs should focus on the work created by the chatbot, not just the copying itself.

VIDEO: ChatGPT and Generative AI Are Hits! Can Copyright Law Stop Them?

To maintain a successful vicarious infringement claim, the plaintiff must prove that the defendant created a work that copied at least some aspects of the copyrighted work. This requirement is met by showing that the defendant directly copied from the plaintiff’s work and produced something that looks or sounds the same.

In cases where infringement is less blatant, plaintiffs can focus on the substantial similarities between the allegedly infringing work and the original copyrighted work.

Here, the judge’s dismissal of the vicarious infringement claim turns on this concept of substantial similarity between the copied works and what OpenAI created. Silverman and the other plaintiffs assert that they don’t need to show substantial similarity because they have evidence of direct copying—that is, evidence that OpenAI directly copied the text of their books to train its large language model.

But this misses the mark. While OpenAI may directly have copied these plaintiffs’ books as input for its models, the allegedly infringing content at issue is the ChatGPT outputs.

The plaintiffs allege that these outputs are derivative works of their copyrighted materials (and, thus, still infringing), but the judge held that they haven’t specifically and adequately alleged substantial similarities. In their original complaint, Silverman and the other plaintiffs claim that OpenAI directly copied the text of their original works, which was then “ingested” by ChatGPT. Because of this alleged direct copying, ChatGPT was then able to spit out “very accurate summaries” of their books.

Notably, the plaintiffs never claim that ChatGPT’s outputs contain direct copies of any text from their books. The plaintiffs specifically point to the fact that the outputs “get some details wrong” because the ChatGPT model “mixes together expressive material derived from many sources.”

Unfortunately for Silverman and the other plaintiffs, this mixing together of expressive material is exactly the type of output that copyright law is intended to protect. Copyright law doesn’t protect concepts, ideas, or themes; rather, it protects the specific expression of an idea.

For example, an author writing the biography of a famous person will be influenced, informed, and guided—in other words, trained—by prior books, news articles, and other written accounts of the subject’s life. These prior authors can’t then claim that the biographer has infringed their copyrights by telling the same story in a different way.

This concept is key to the US Constitution’s promotion of the progress of “useful Arts.” Central to this principle of progress is the idea that all new art builds on prior art. While foundational to US copyright law, this idea predates any American artistic ideals. Philosophers have long debated whether art is creation or, rather, merely imitation of the natural world.

Philosophical debate aside, all art continues to build on that which preceded it. Gavin Alexander, professor of Renaissance Literature at the University of Cambridge, explains the impact of John Milton’s unique words and phrases on the countless creators who have built upon his works: “As with Shakespeare, fragments of whose lines are scattered in the titles of novels and films, Miltonic phrases are found—sometimes deliberately alluded to, and sometimes with no idea of their origin—in the works of modern fiction, film, music, and art.”

Here, the summaries of the plaintiffs’ books generated by ChatGPT are akin to the work of the biographer described above. It isn’t copyright infringement for the AI model to ingest the authors’ works, just as it isn’t infringement for the biographer to read prior work about her subject’s life, as long as ChatGPT and the biographer use the copyrighted materials to create a new, original expression.

Thus, Silverman and her fellow plaintiffs’ success will depend on the availability of evidence that shows that ChatGPT outputs contained material that directly copies at least portions of their books. In this case, it isn’t clear if such evidence exists.

Other plaintiffs in similar cases may be better poised for success. For example, the New York Times filed a complaint in the Southern District of New York in December alleging that generative AI tools, including OpenAI and Copilot, create outputs that not only summarize the Times’ content but recite it verbatim. Similarly, a large group of music publishers filed a complaint last October in the Middle District of Tennessee claiming that AI company Anthropic’s chatbot “Claude” generates output with “identical or nearly identical copies” of lyrics owned by plaintiffs.

As authors, artists, musicians, and other creators continue to challenge the ways that AI companies are using and exploiting their copyrighted materials, plaintiffs should carefully consider if and how these AI models are copying from their works and be prepared to present evidence of direct copying or, at a minimum, substantial similarity.

The case is Silverman et al v. Open AI, N.D. Cal., 3:23-cv-03416, decision 2/12/24.

This article does not necessarily reflect the opinion of Bloomberg Industry Group, Inc., the publisher of Bloomberg Law and Bloomberg Tax, or its owners.

Author Information

Thomas W. Brooke is partner in Holland & Knight’s intellectual property practice in Washington.

Allysan Scatterday is an intellectual property associate in Holland & Knight’s Nashville office.

Write for Us: Author Guidelines