Authors v Anthropic: Judge gets tutorial on AI training in copyright battle

Attorneys for Anthropic PBC, an AI software development company, and a group of authors provided a San Francisco federal judge a "tutorial" on the large language model powering the company's AI chatbot, Claude, during a hearing on Thursday.

At the heart of the dispute is the putative class action filed in August that accuses Anthropic of committing "massive" copyright infringement by allegedly including the plaintiffs' literary works in the dataset that powers Claude. Bartz v. Anthropic PBC, 3:24-cv-05417 (N.D. Cal. filed Aug. 19, 2024).

During the hearing before U.S. District Judge William Alsup in San Francisco, attorneys from both sides provided the court a 45-minute explanatory presentation on how Anthropic's technology takes large amounts of information, including written works, and translates the information into "tokens" to train the Claude chatbot to provide users with realistic responses to text prompts.

Alsup originally ordered the tutorial hearing following a discussion among the parties during a case management conference in October. He wanted information about the underlying claims in the complaint. Throughout the hearing Alsup asked both sides questions about their presentations to better inform himself on the legal issues.

The parties' presentations centered on the plaintiffs' claims that Anthropic used a large open-source data set known as Pile, hosted by the non-profit EleutherAI, that allegedly contained copies of the plaintiffs' copyrighted works in a subset of data known as "Book3."

Anthropic's legal team from Arnold & Porter Kaye Scholer LLP and Latham & Watkins concede that the large language model used Pile to train its AI software but deny liability on the grounds of fair use.

Justin Nelson, one of the plaintiffs' attorneys who led Thursday's tutorial and a Houston partner for Susman Godfrey LLP, summarized the technology tutorial and explained the gravity of the case in a phone interview. Bartz v. Anthropic PBC, 3:24-cv-05417 (N.D. Cal. filed Aug. 19, 2024).

"The AI model takes this data, this human expression, and basically encodes it and so the software is able to spit out a predictive model of what's next," Nelson said. "A ruling against the authors, the copyright holders, would be devastating for all types of creators throughout the country."

In an email from Anthropic's media relations office Thursday, the company thanked Alsup for ordering the hearing and summarized its presentation.

"We appreciated the opportunity to share with Judge Alsup some foundational principles on how AI models are built," the company stated. "As we demonstrated in our tutorial, large language models are trained by deriving facts, patterns and relationships from a myriad of information to create entirely new material and we take steps to prevent the use of our model to generate content that already exists. We believe that the use of any copyrighted content for model training is a quintessential example of transformative fair use under copyright law."

The following attorneys for Anthropic could not be reached for comment Thursday: Douglas A. Winthrop, Angel T. Nakamura of Arnold & Porter Kaye Scholer LLP; and Joseph R. Wetzel of Latham & Watkins LLP.

The following attorneys from Leiff Cabraser Heimann & Bernstein LLP serve as Nelson's co-counsel in the matter: Reilly T. Stoler of San Francisco; Rachel Geman, Anna Freymann and Jacob S. Miller of New York. The following attorneys from Susman Godfrey also serve as Nelson's co-counsel: Rohit D. Nath and Alejandra C. Salinas of Houston.

In the company's answer to the complaint, the defense team denied wrongdoing and explained how its use of copyrighted material does not create grounds for liability.

"Anthropic admits that building Claude requires breaking down input text into smaller pieces --words or portions of words, called 'tokens' -- and then training a model on those tokens,' the filing stated. "The quality and the quantity of the training corpus contribute to the quality of the resulting model and that the phrase 'garbage in, garbage out' is one way to describe that idea. Anthropic admits that it has used portions of the Pile in connection with the training of certain Claude models."

Both parties are drafting dueling motions in the case. The plaintiffs' motion for class certification is due in March, and Anthropic has repeatedly told the court it intends to file a motion to dismiss the lawsuit soon.

In recent years Anthropic has been accused of violating intellectual property laws in several lawsuits across the country.

Currently, the company is battling a similar action by a group of music publishers, including Universal Music Group and Capital Music Group, in the Northern District.

The plaintiffs originally sued Anthropic over claims that "cutting edge" AI virtual assistant is built on "widespread" copyright infringement and illegally transcribes song lyrics for purposes of training the large language model behind Claude. While the parties are still litigating the case, Anthropic agreed to institute "guardrails" to prevent future infringement in a Jan. 2 stipulation. Concord Music Group, Inc. et al. v. Anthropic PBC, 5:24-cv-03811 (N.D.Cal. filed June 26, 2024).