Please wait while the page is loading...


INTA 2024: Debate continues on copyrights for ChatGPT, other generative AI

23 May 2024

INTA 2024: Debate continues on copyrights for ChatGPT, other generative AI

Intellectual property  lawyers shared their insights on the ongoing debate about copyright questions with the advent of ChatGPT and other artificial intelligence products.

The panel in the INTA 2024 session “Decoding Copyright in the Era of ChatGPT: Generative AI and Copyright Law” from the United States, Japan, and the European Union noted similarities and differences in copyright laws across jurisdictions and their application to AI questions in the context of products such as ChatGPT.

Catherine Zaller Rowland of the Copyright Clearance Center and Joseph C. Gratz, a San Francisco-based partner at Morrison & Foerster, noted that the U.S. Congress has not passed specific AI laws regarding copyright. However, legislation focusing on transparency is pending, and a U.S. Copyright Office study is underway related to copyright’s relevance in AI training.

“In the U.S., we currently do not have any specific law addressing AI and copyright. There is no specific text that mainly addresses AI copyright, and there is no specific treatment of AI copyright in U.S. law,” Rowland told the panel in Atlanta.

“We currently rely on existing copyright law (exclusive rights, exceptions, and limitations) and also on reproduction, distribution and fair use arguments.”

From the European Union’s (EU) perspective, Julian Waiblinger, a partner at Nordemann in Berlin, noted that training AI systems – or the “input side” of ChatGPT – requires reproduction of works, but these are covered under the Text and Data Mining exception of Article 4 of the EU Directive on Copyright in the Digital Single Market (2019).

He said that right holders can also opt out in machine-readable form, though works can be licensed for AI training – but noted that this would be “difficult in practice.”

In Japan, Atsushi Okada, a partner at Mori Hamada and Matsumoto in Tokyo, said that training AI systems requires the reproduction of works but is covered under the data analysis exemption of Japan’s copyright law.

He noted that this has “broad application”. The basic rationale is that AI training does not aim at the enjoyment of thoughts or sentiments expressed in the work for data analysis.

“But this exception has limits,” Okada noted. “The exception to this exemption is if the interest of copyright owners would be unreasonably prejudiced, then this exemption does not apply.” Okada said the rightsholders’ right to opt out is “not explicitly granted.”

Meanwhile, in terms of licensing in AI training, he said that Japanese laws allow copyright owners and AI businesses to enter into license agreements to use high-quality data for training.

In the U.S., Rowland said no law currently requires licensing, prohibits licensing or imposes licensing obligations for copyright on AI, and statutory licensing is “much more limited” to music and satellite transmission, among others. “Having good licensing markets is a good way to deal with this, but in the United States, there’s no statutory law on AI products.”

Gratz, who has represented openAI in U.S. litigations, emphasized that if it’s “fair to use,” you don’t need a license.

“But if the uses are beyond fair use, then licensing is an appropriate direction,” he said.

Regarding whether AI content could be protected by copyright, Waiblinger noted that, in the EU perspective, “copyright protection of AI-generated content is not ruled out in general, “but it is necessary that the human author has had sufficient influence on the result.”

In the U.S., Gratz said, “purely AI-generated works would not qualify” for copyright protection. “But protection is available for original expression and potential compilation ownership.”

“There have been several cases and decisions in the US, and the copyright office has stated its position – which is consistent with the courts’ – that there has to be human authorship. That is all copyrightable expression.

“If there are parts of it that don’t reflect human authorship or expressions that did not come from a human author, those parts aren’t protected by copyright.”

Rowland pointed out that AI is just “another category of public domain.”

“It’s important to look at AI not as a weird, unique thing that is so hard to understand, especially if you’re trying to register works but look at it in combination with other public domain. It’s not super special in that way. It’s just another category of public domain,’ she said.

From the Japanese perspective, Okada said creations using AI “as a tool” could be protected by copyright, but works “created autonomously by AI have no copyright.”

In terms of liability for the output of these AI products, Rowland said, “When you look at outputs, you’ll look at them and see if they are considered derivative work. What is there? Are they substantially similar or competing with the original work? If it is, then absolutely, this is copyright infringement.”

Nonetheless, she noted that this could be subject to exceptions and limitations.

Gratz said there are existing rules for analyzing whether AI-generated work is breaching copyright.

“This is actually pretty normal: a machine outputs something substantially similar to preexisting work. How do we analyze that situation? It happens all the time, and we have normal rules for dealing with it,” he told the panel.

“We look at how similar it is and whether it is fair use. As to who is responsible, we look at issues of causation and all these rules that we’ve developed in other technologies for deciding,” Gratz added. If there is infringement, who is causally the person who is a proximate cause of the ‘infringingness’ of that?”

In the U.S., both Gratz and Rowland noted that technology moves faster than legislation could, but existing laws could be used to address these questions.

“We have a long tradition of addressing new technological uses of copyrighted works through the courts rather than legislation,” Rowland said, pointing out that legislation would likely focus on other areas such as digital replicas.

“In the U.S., at least, the certainty that we’re going to get will come from case-by-case decisions by the courts that establish a sort of rules and practices that people can look at rather than getting legislation that answers the question for us or sets a statutory law,” Gratz said.

- Charlee C. Delavin, reporting from Atlanta

Law firms