The first decision in the US dealing with the use of copyrighted works to train generative AI-systems

Vincenzo Iaia

On 30 October 2023, the US District Court for the Northern District of California issued the first decision in the US (and probably worldwide) ruling on the legitimacy of the reproduction of copyrighted works for training generative AI-systems.

In particular, the case No. 23-cv-00201-WHO originated from the lawsuit filed by the artists Sarah Anderson, Kelly McKernan, and Karla Oritz against the AI-companies Stability AI Ltd. and Midjourney Inc. for several grounds, namely copyright infringement, violation of their rights of publicity, unfair competition and breach of contract.

As regards copyright, the plaintiffs alleged that Stability AI scraped over five billion images from the Internet, including their copyright-protected works, to nourish their algorithms and offer creative assistants, like Midjourney, a service that produces images in response to text prompts. However, the plaintiffs did not identify which of their specific works covered by a registration before the US Copyright Office were used as training images. They just relied on the results of a search of their name on the “” website and they referred to it to support the plausibility and reasonableness of their beliefs. The lack of further elements to corroborate plaintiffs’ theory of direct copyright infringement led the court to dismiss the claim.

The artists also alleged that the AI-companies were liable for vicarious copyright infringement. According to settled case law1, this form of liability presupposes that the defendant has i) the right and ability to supervise the infringing conduct and ii) a direct financial interest in the infringing activity. However, the plaintiffs failed to demonstrate the point i) since the AI-generated outputs are unlikely to be a close match for any specific image in the training dataset thanks to the use of conditioning data to interpolate multiple latent images. Hence the AI company could not be in the condition to detect the contents that have been probably used for text and data mining.

For what it concerns the right of publicity, the court acknowledged the absence of allegations regarding how the use of the plaintiffs’ names in the products’ text prompts would produce an AI-generated image “similar enough that people familiar with plaintiff’s artistic style could believe that plaintiffs created the image”. Nor the plaintiffs alleged plausible facts about the use of their names in connection with advertising and any other commercial interests.

More in general, the court dismissed the whole action for lack of evidence but allowed the plaintiffs to amend their claims to clarify how each defendant infringed upon their copyrights, removed or altered their copyright management information, or violated their rights of publicity. It would be quite surprising if the artists would not take this tempting opportunity to better explain their theories and persuade the court that their works have been reproduced without seeking any permission. If and when the procedural grounds for refusal will be fixed the court should enter into the merits of the question by assessing whether text and mining of copyrighted works constitutes fair use or not. Therefore, this is certainly not the last say on the commercial use of copyrighted works to nourish and train AI models.

[1] U.S. Court of Appeals, 9th Circuit, decision of 3 July 2007, No. 494 F.3d 788, Perfect 10, Inc., Plaintiff-Appellant, v. Visa International Service Association.


Vuoi richiedere informazioni?