Generative AI (GenAI) algorithms require data inputs to analyze, transform, and generate content. But does using copyrighted material without prior authorization for training or operating these algorithms infringe on the rights of the original creators, or is it permitted under the fair use exception?
Fair use is a legal doctrine that permits limited use of copyrighted material without obtaining permission from the copyright holder. Courts will assess whether a use constitutes fair use by balancing the following four factors: (1) the purpose and character of the use, (2) the nature of the copyrighted work, (3) the amount and substantiality of the portion used, and (4) the effect of the use on the market for the copyrighted work.
It’s a fine line that must be assessed on a case-by-case basis, and legal precedent is in flux as courts across the U.S. grapple with these cases of first impression in the GenAI space. Recently, several California federal judges have ruled on cases involving fair use of copyrighted materials in training GenAI algorithms, which may offer insights into litigation outcomes in the future.
Transformative Use or Derivative Work?
When evaluating the first factor of the fair use standard — the purpose and character of the use — whether the use is transformative is an important consideration. Transformative use refers to instances where the use adds something new and alters the original work with new expression, meaning, or message. If the GenAI algorithm’s use of copyrighted material in its inputs or outputs is deemed a “transformative use,” it is not dispositive of a fair use finding, but it strengthens a fair use claim.
However, many copyright holders have argued that use of their works to train GenAI algorithms is directly infringing, or, at best, infringes by creating unauthorized derivative works of the original. A derivative work is an exclusive right reserved for the copyright holder, allowing them to create works substantially similar to the original with the addition of new elements or modifications.
In simple terms, courts must consider whether the use only adds new elements to the original work (derivative) or adds an entirely new expression, meaning, or message (transformative). This evaluation can be subjective and depends on the specific facts and evidence of each case.
Recent California Court Rulings – Key Takeaways
- Using Copyrighted Material to Train GenAI May Constitute Fair Use — But Beware of Using Pirated Content. In June, a California federal judge ruled on summary judgment that the use of copyrighted books to train large language models (LLM) constituted fair use due to being “exceedingly transformative” in training the algorithm to generate new text outputs. The court also ruled that converting purchased print books into digital format in a central library was transformative because it was for the purpose of facilitating storage and searchability, did not create new works, and copies were not disseminated outside the company. However, the court refused to grant summary judgment in favor of the defendants for pirated books uploaded into the central library, as the pirated books displaced demand for legitimate copies. In July, the court has certified a class for actual or beneficial owners of copyrights in books bearing ISBNs or Amazon Standard Identification Numbers downloaded by Anthropic from illegal online libraries, Library Genesis and Pirate Library Mirror, where there was sufficient evidence of metadata identifying the pirated materials. The court declined to certify owners of copyrights in books taken from Books3, a third illegal online library, because those copies came with less metadata, making identification of titles and authors too burdensome. Bartz et al v. Anthropic PBC, 24-05417 WHA, No. 231 (N.D. California).
- Even if a GenAI Algorithm’s Use May Be Transformative, Evidence of Market Harm Can Undermine a Fair Use Finding. In an almost simultaneous ruling on summary judgment, another California judge agreed that it can be highly transformative to use copyrighted materials to train LLM. However, the court noted that even if a use is transformative, there may not be fair use if the defendant’s use significantly harms the market for the original work — with harm to the market being the most critical factor in evaluating fair use. The court ultimately determined the GenAI’s use to be fair use because the plaintiff presented insufficient evidence of market harm. Kadrey et al v. Meta Platforms, Inc., 23-cv-03417-VC, No. 598 (N.D. California).
- Courts May Consider Whether GenAI Models Are Infringing Derivative Works – If Properly Alleged. Last year, a California federal judge suggested that a derivative infringement claim could be plausibly alleged on the basis that a GenAI model is a derivative work of its copyrighted training materials. The plaintiffs’ complaint alleged a GenAI image generator was a derivative work because it contained compressed versions of billions of copyrighted images and made intermediate copies of such images during training. This includes the GenAI allegedly creating a copy of a copyrighted image with slight variations, in multiple iterations.
The court specifically distinguished the case from Kadrey, because the plaintiffs in Kadrey failed to allege the LLM’s “recasting or adaptation” of the copyrighted materials in any output — much like when a book is adapted into a screenplay or translated into another language. Furthermore, the court differentiated between GenAI creating repeated, iterative copies of copyrighted images and GenAI accessing copyrighted digital books without making any changes to the text. This suggests that courts may analyze GenAI’s uses and outputs of text and images differently when evaluating the sufficiency of potential claims for copyright infringement. Andersen et al v. Stability AI Ltd. et al, 23-cv-00201-WHO, No. 223 (N.D. California).
Ultimately, each GenAI model’s operation is unique, whether an LLM or an image generator, meaning case law precedent may apply differently (or not at all) to certain types of GenAI algorithms. Many of these cases also remain ongoing, so time will tell how the courts will ultimately rule in these matters. We will continue to monitor these cases and their impact on copyright holders and GenAI algorithm creators.
