Chinese AIGC Platform Found Secondarily Liable for Copyright Infringement

Tag: intellectual property-intellectual property dispute resolution, digital economy, telecommunications media entertainment and technology entertainment, artificial intelligence

I. Introduction

On December 20, 2024, the Hangzhou Intermediate Court affirmed a decision made by the Hangzhou Internet Court, which found the defendant, a generative AI platform provider, contributorily liable for infringing the copyright of the famous Ultraman IP.[1] This is the first Chinese case where an AIGC platform was found contributorily liable for copyright infringement. The first-instance court held that the defendant knew or should have known its users’ direct infringement but failed to take effective action to prevent such infringement, and thus constituting contributory liability.

II. Background

In this case, Shanghai Character License Administrative Co., Ltd. (“Plaintiff”) was granted an exclusive license to the well-known “Ultraman” works in China, which includes the right to reproduce, the right of dissemination over the internet, and the right to enforce its licensed IP, among others.

The defendant, an AI tech company based in Hangzhou (“Defendant”), operates an AI platform, which provides text-to-image and image-to-image AI generating services, as well as online model training services. Defendant’s model training service allows its users to feed their own training data to create a more “customized” AI model, by using Checkpoint base models and LoRA (Low-Rank Adaptation) models.

In the subject case, users of Defendant’s platform created a “customized Ultraman” AI model, by uploading Plaintiff’s copyrighted Ultraman pictures to the base model and adjusting technical parameters. Subsequently, other users also input prompts related to Ultraman, select the base models, and overlay the Ultraman LoRA models to generate more images that are substantially similar to the Ultraman characters. During this process, these AI-generated images and LoRA models can be collected, applied, published and shared by users via links over the internet. 

Pic 1: The screenshot of a user-uploaded Ultraman pictures to train a LoRA model

Pic 2: A picture involving the Ultraman character generated by the Ultraman LoRA Model, posted by a user on Defendant’s platform

Plaintiff asserted that Defendant directly or indirectly infringed its copyright and also engaged in unfair competition behavior, thus requesting for damages in the amount of RMB 300,000. Defendant argued that it was merely using third-party open-source models to provide AIGC services to the public, and the allegedly infringing training data was not provided by it but was uploaded by users instead, and therefore, it should enjoy the "safe harbor" exemption from liability. 

III. Issues

In this case, the first-instance court focused on the following few issues: (1) whether Defendant infringed Plaintiff’s copyright, including the right of dissemination over the internet; and (2) whether Defendant’s acts constituted unfair competition.

1. Whether Defendant infringed Plaintiff’s copyright

The first-instance court found that Defendant was liable for contributory infringement based on the following reasoning.

(1) Whether Defendant constituted direct infringement

The first-instance court ruled that Defendant falls into the category of “AIGC service provider” based on the Interim Measures for the Management of Generative Artificial Intelligence Services issued in July 2023 (the “2023 GAI Measures”). Yet, the court held that because Defendant did not use Plaintiff’s copyrighted works to train the AI models, nor did it collaborate with users in generating or disseminating the infringing Ultraman pictures, Defendant was not directly liable for copyright infringement.

(2) Whether Defendant constituted contributory infringement

When analyzing whether Defendant constituted contributory liability, the first-instance court considered the following factors, including, the nature of Defendant’s AI services, the popularity of Plaintiff’s works and Defendant’s knowledge of users’ direct infringement behaviour, the impact of the infringing AI models, Defendant’s business model, and whether Defendant has adopted effective preventive measures, etc. The court ruled that because Defendant “separately listed” well-known IP works, including Plaintiff’s Ultraman works, under its LoRA model, which facilitated the ease of users’ direct infringement, Defendant knew or should have known users’ infringement behaviour. Also, because Defendant charged users for its AI services, it directly profited from users’ infringement. The court also believed that Defendant could have taken more proactive measures to manage and filter infringing content but failed to do so, thus finding Defendant secondarily liable for infringement.

2. Whether Defendant’s acts constituted unfair competition

The first-instance court dismissed Plaintiff’s claim regarding unfair competition, holding that Defendant’s AI services was trying to provide users with more personalized “content customization” and Defendant did not violate the principles of good faith and business ethics, and thus did not constitute unfair competition.

In summary, the first-instance court found Defendant liable for contributory infringement against Plaintiff’s copyrighted works, but dismissed Plaintiff’s claim on unfair competition. The court ordered Defendant to cease the infringing acts, including deleting infringing images and infringing Ultraman LoRA models, and to compensate Plaintiff for damages and reasonable expenses at RMB 30,000. This decision was later upheld by Hangzhou Intermediate People’s Court. 

IV. Comment

There are two interesting issues in this underlying opinion. First, this is the first AI case where Chinese judges are asked to decide whether the “safe harbor” provision should also apply to AIGC providers as in those ISP infringement cases. The Hangzhou Internet Court eventually applied a similar “contributory liability” test standard, including the “red flag” knowledge test, AIGC provider’s profit model and AIGC provider’s ability to control, etc., to determine whether Defendant’s AI platform should be held secondarily liable for users’ direct infringement.

Second, although this case does not directly involve the AI company’s unauthorized use of copyrighted works to train AI models—in this case, the infringing prompts were uploaded by users and not by Defendant, the judge did write a few interesting paragraphs to indicate its attitude if the AI company were using unauthorized copyrighted works to train AI models and its possible “fair use” argument. The court said, “Gen-AI technology involves four important stages: data input, data training, content generation and content use. To promote the development of Gen-AI technology, in terms of infringement liability relating to data input and data training stages, a relatively relaxed and tolerant standard should be adopted; in terms of infringement liability relating to AIGC content production and content usage stages, a more stringent liability standard should be adopted…” Later in its opinion, the court also briefly discussed fair use although this case did not actually provoke a fair use argument. The court commented that, “[T]he development of Gen-AI technology requires huge amount of training data in its data input stage, which inevitably requires use of others’ [prior] works. The purpose of using others’ works in data training stage, in essence, is to learn and analyse the expressions, language characteristics and styles of such prior works, so as to abstract its corresponding rules, structure, models and trends to transform into new works. The purpose of such use…is different from the purpose of representing the originality of the prior work…” This analysis sounds quite similar to the “transformative use” argument adopted by the U.S. courts in their fair use findings under Warhol[2] and Google[3], among other cases. In the recent case Thomson Reuters v. Ross, the U.S. court rejected Ross’ fair use argument, but the court acknowledged that the case presented before him was not a typical Gen-AI case and the defendant’s use was not “transformative”.[4] It would be interesting to observe how the Chinese courts and/or the U.S. courts would decide on a more typical Gen-AI copyright infringement case involving fair use.

*Thanks to Huang Jiaona and Dou Yinghai for their contributions to this article.

Scan the QR code to subscribe to "King & Wood Mallesons" for more information