CHINA Trends and Developments Contributed by: Qiang Ma, Yi Xu and Lei Wu, Jingtian & Gongcheng
subjective fault, and does not constitute indi - rect infringement. Analysis Courts across three regions have reached a consen - sus regarding AI platforms. (1) No direct infringement – AI platforms are not liable for direct infringement if users upload data and provide prompts that result in infringing content, which they then publish. The platform does not supply infringing data, cannot control the content generation process, and lacks the intent to collaborate with users. (2) No pre-review obligation – AI platforms are not required to pre-screen user inputs, including training data and prompts. Such requirements would impose an undue burden and impede industry growth. (3) Indirect infringement – If an AI platform fails to meet its reasonable duty of care, knowing or having reason to know about infringements, it may be liable for indi - rect infringement. Key components of the reasonable duty of care include the following. (a) Difficulty in detecting potential infringement by ordinary internet service providers, considering factors like the work’s popularity and whether generated content is easily identifiable. (b) Implementation of reasonable technical meas - ures to prevent infringing content, such as establishing complaint mechanisms, encourag - ing IP respect through user agreements, block - ing prompts, and filtering keywords. Providers that gain economic benefits must uphold a higher duty of care. In summary, if a GenAI provider can show it was hard to identify potential infringing content or demonstrate efforts to prevent it, it may be deemed to have fulfilled its duty of care. Conversely, failure to do so indicates fault. (4) Safe harbour rule applicability – Courts stress the need to balance AI industry development with copy - right protection. Onerous obligations on AI providers could stifle growth, while reasonable duties of care are essential for copyright protection. The assessment of
fault should consider industrial capabilities and prec - edents, promoting a cautious judicial attitude towards generative AI services. GenAI’s data input and model training: the applicable elements, judgment standards, and boundaries of the fair use rule Practical operational difficulties in the traditional authorisation paradigm The traditional authorisation model is challenging for GenAI due to its vast data input and model train - ing needs, which involve many copyrighted works. Securing permissions individually presents three main issues: • high communication costs; However, weakening copyright protection due to these challenges contradicts the intent of copyright law. Sorting out existing judicial judgments Among the three adjudicated cases, the Guangzhou Internet Court’s ruling did not address “data input and model training” since the AI painting function was pro - vided by a third party, and the alleged AI platform did not conduct model training. The third case, from the Jinshan District Court, remains undisclosed, creating uncertainty around its discussion. In the second case, from the Hangzhou Internet Court, “data input and data training” was recognised as “fair use”, leading to the rejection of the plaintiff’s request for deletion of data related to the protected work. “Fair use” rule The determination of whether data input and model training for generative AI platforms qualify as “fair use” under Copyright Law depends on the specific circum - stances of each case and the burden of proof on both parties. Fair use can be recognised in the following scenarios. • Transformative nature – If model training is highly • difficulties in verifying ownership; and • complex licensing considerations. transformative, using original works solely as training materials for creating new works without
86 CHAMBERS.COM
Powered by FlippingBook