Generative AI & Law: Diffusion Models are not Stochastic Parrots

Photos © Aditya Mohan | All Rights Reserved.  These views are not legal advice but business opinion based on reading some English text written by a set of intelligent people. 


The Andersen et al v. Stability AI Ltd. et al (Case 3:23-cv-00201-WHO) lawsuit involved artists Sarah Anderson, Kelly McKernan, and Karla Ortiz suing Stability AI, Midjourney, and DeviantArt for allegedly using their copyrighted works without permission in their generative AI software, specifically Stable Diffusion. Judge William Orrick dismissed most claims but allowed Andersen's direct copyright infringement claim against Stability AI to proceed. He also expressed doubt about the plaintiffs' arguments that AI-generated outputs are simply derivatives of the copyrighted images, highlighting the need for substantial similarity in such claims.

The case of Andersen et al v. Stability AI Ltd. et al (Case 3:23-cv-00201-WHO) involved a significant legal battle regarding the use of copyrighted images in training generative AI platforms. The lawsuit was brought by artists Sarah Anderson, Kelly McKernan, and Karla Ortiz against Stability AI Ltd., Stability AI, Inc., Midjourney, Inc., and DeviantArt, Inc. Here is a detailed overview of the case, incorporating court proceedings quotes and arguments:


The plaintiffs accused the defendants of infringing copyrights in their artwork through the use of generative artificial intelligence software. The central claim was that Stability AI's software, Stable Diffusion, was trained on the plaintiffs’ works to produce output images in their artistic styles. This lawsuit was seen as a pivotal case concerning the legality of using copyrighted images to train AI models.

Photo depicts a frictional scene from the San Francisco Courthouse that features artists in a serious verbal confrontation with cyborg robots, with lawyers assisting them. The atmosphere is intense and charged, and the cute Stochastic Parrot adds a unique touch to the scene. A judge is also present, overseeing the dynamic situation. 

Judge William Orrick's Rulings

U.S. District Judge William Orrick made several key rulings in this case. He largely granted the defendants’ motion to dismiss under Federal Rule of Civil Procedure 12(b)(6) with leave to amend, but denied it only as to the direct copyright infringement claim that Plaintiff Anderson asserted against Stability AI. Orrick recognized that the determination of whether copying in violation of the Copyright Act occurred in the context of training Stable Diffusion could not be resolved at that juncture. He agreed with all three companies that the images the systems created likely did not infringe the artists' copyrights, but he allowed the claims to be amended. Orrick was "not convinced" that allegations based on the systems' output could survive without showing that the images were substantially similar to the artists' work. 

Plaintiffs' Claims and Court's Response

The plaintiffs argued that Stability AI downloaded or acquired copies of billions of copyrighted images without permission. The court allowed the claim of direct copyright infringement against Stability AI to proceed but dismissed the other claims, including vicarious infringement, with leave to amend. The court found the plaintiffs' allegations that Stability AI acquired and used their copyrighted works to train Stable Diffusion, and then stored or incorporated the training images into Stable Diffusion, sufficient to plead direct infringement. However, the claims against DeviantArt and Midjourney were found to be insufficiently detailed, particularly regarding their role in using copyrighted images and the creation of derivative works. The court dismissed these claims with leave to amend.

DMCA, Unfair Competition, and Breach of Contract Claims

The plaintiffs also brought claims under the Digital Millennium Copyright Act (DMCA), alleging that the defendants removed copyright management information from their works. The court ordered the plaintiffs to clarify and allege plausible facts about what specific copyright management information was altered or removed. As for the unfair competition claims, the court found them to be premised on purported copyright violations and thus preempted by the Copyright Act. The breach of contract claim against DeviantArt was also dismissed, as the plaintiffs did not identify specific provisions of the Terms of Service or Privacy Statement that were breached 

Implications and Next Steps

Judge Orrick's rulings and the plaintiffs' subsequent responses highlight the legal complexities surrounding the use of copyrighted materials in AI training. The case's progression will likely provide further insights into how courts view the relationship between AI-generated content and existing copyright laws. The case's outcome could set important precedents for the use of copyrighted works in AI technologies and shape the evolving landscape of digital copyright law.

The Anderson v. Stability AI case represents a crucial juncture in the intersection of AI technology and copyright law, with the potential to influence future legal standards and practices in this rapidly evolving field.

Stochastic Parrots

The term "Stochastic Parrots" was popularized by a research paper titled “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” authored by Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell. Presented in early 2021, the paper critiqued large language models (LLMs) for their potential to propagate biases and misinformation, likening them to parrots that mimic without understanding. The term suggests these models merely replicate patterns found in their training data without true comprehension or originality. However, this characterization is increasingly seen as a simplification, especially for foundation models like LLMs and diffusion models. These advanced AI models, including GPT and Stable Diffusion, demonstrate capabilities beyond simple replication, such as generating novel content, creative outputs, and complex problem-solving, indicating a level of processing and transformation that goes beyond the "Stochastic Parrots" label. This evolving understanding reflects the growing sophistication and potential of these models in various applications.

Diffusion Models are Not Stochastic Parrots

This case supports the notion that diffusion models are not Stochastic Parrots. Judge Orrick's skepticism about the AI-generated outputs being direct derivatives of the original artworks implies an understanding that these models, while trained on existing images, create new, transformative outputs. This suggests that diffusion models like Stable Diffusion do more than just replicate input data; they generate novel creations that are distinct from the source material, showcasing a level of originality and transformation beyond simple replication.

Image depicting a frictional interaction between a robot  with the patent officer, Dr. William Thornton, in the U.S. Patent Office in 1802, submitting a patent application  


Further read