This week’s Opinion Column argues that, by notably sidestepping key copyright issues, China’s AI Action Plan signals a policy direction that prioritises access over clear IP rights protection.

Chinese Premier Li Qiang unveiled a sweeping Global AI Governance Action Plan at the 2025 World Artificial Intelligence Conference (WAIC) held in Shanghai last month.

Framed as a 13-point blueprint for responsible AI development, the plan aspires to position China at the center of international AI coordination and governance. It arrives on the heels of the US releasing its own deregulation-oriented AI policy, making clear the geopolitical undercurrents of global AI rulemaking.

While much of the plan focuses on cooperation, sustainability and global equity, those watching the IP implications of AI datasets found the subtext especially revealing; copyright and ownership issues remain largely unaddressed, and in some areas, conspicuously absent. (Note that the US AI action plan only referred to intellectual property once, more in the context of national security.)

Here, we analyse the action plan and predict how it may shape IP rules around training datasets.

China’s AI Action Plan: the vision
The Action Plan calls for multilateral collaboration on AI infrastructure, open innovation and global data sharing. Two provisions – on high-quality data supply and capacity-building cooperation – seem to have ramifications for, or are of particular interest to, the IP community.

The Action Plan explicitly advocates for the co-development of large-scale, high-quality datasets for training AI systems, including corpora for language models. It also envisions platforms for the legal and orderly cross-border flow of data, emphasising diversity, privacy and bias prevention.

But what it does not mention is equally important; there is no reference to copyright, ownership or fair use of such datasets in these sections. The omission raises questions about whether and how rights holders – particularly creators of copyrighted datasets, books, images and other expressive works – will be protected or compensated in these collaborative data-sharing regimes.

The only explicit acknowledgment of copyright obligations appears in a separate section on public sector AI applications, which states that authorities must “respect patents, software copyrights, and data protection”. But this feels compartmentalised, leaving the broader question of training dataset legality untouched.

Elsewhere, the Action Plan encourages international cooperation in building data corpora, including joint laboratories, open benchmarking platforms, and data-sharing for AI education and literacy, particularly for developing countries. These initiatives are laudable from a development and access perspective, but again, copyright and IP governance are not part of the conversation, at least not publicly.

This suggests a policy direction that prioritises dataset accessibility and international collaboration, potentially before establishing a comprehensive legal framework for ownership and use.

China's legal reality: Hangzhou AI copyright decision
Despite the ambiguity in policy pronouncements, Chinese courts have begun addressing the issue. In Shanghai Alpha Animation v Hangzhou Jellyfish AI, a landmark 2024 case decided by the Hangzhou Internet Court ((2024) 浙0192民初1587号), the court ruled on the legality of using third-party works – specifically images of “Ultraman” (奥特曼) – in AI training datasets.

The court held that while the AI platform provider failed to adequately monitor and prevent infringing outputs (and was thus liable for those outputs), the use of copyrighted material for model training alone could constitute fair use under certain conditions.

The court stated:
"In the absence of evidence that the purpose of using copyrighted works in training was to reproduce their original expression, or that such use impairs normal use of the work or causes unreasonable harm to the rightsholder, training may be considered fair use."

Notably, the court rejected the plaintiff’s sweeping demand for the deletion of the entire dataset, citing a need to maintain balance between copyright protection and technical innovation.

This reasoning echoes other judicial and academic voices in China that call for a “cautious and inclusive” approach to AI regulation, particularly regarding the mass ingestion of copyrighted works during model pre-training.

A parallel track: Anthropic decision in the US
Interestingly, China’s evolving legal stance parallels or more accurately came ahead of the June 2025 US federal court ruling involving Anthropic, an AI developer. The Northern District of California held that using “lawfully acquired books” for model training was fair use, while training on books “scraped from unauthorized sources” was not and could incur liability.

If not overly premature, these two cases together might hint that at least some courts in the US and China appear willing to tolerate broad training use if the underlying content is lawfully obtained, while frowning on indiscriminate scraping of copyrighted materials.

The missing piece: implementation regulations
China’s Copyright Law, as amended in 2020 (effective in 2021), includes flexible fair use provisions in Article 24. The law recognises 12 enumerated scenarios of permissible use but also allows for "other circumstances" as provided by law – a catch-all that could accommodate future AI-specific exceptions.

However, implementing regulations to guide this provision have yet to be finalised. The last revision of the Copyright Law Implementing Regulations dates back to 2013, leaving a regulatory vacuum in a space now moving at AI speed.

It won’t be surprising that stakeholders – tech firms, rights holders and academics alike – are finding ways to advocate for clarity on whether training datasets qualify for fair use, and if not, what licensing mechanisms will be required. The delay in updating the implementing regulations reflects the difficult balancing act between promoting innovation and protecting creators.

We suspect that the courts will continue to play a leading role in shaping the rules in China. Although the Hangzhou judgment has shown an example, whether the courts elsewhere in China defer to the same policy signals remains to be seen. Favouring innovation and access sounds appealing, but it is going to be very controversial to leave content creators and dataset curators in uncertain terrain.Either way, we believe that the next phase of China’s copyright framework will shape not only its domestic AI development but also global norms on the lawful creation and sharing of training datasets.