Generative AI: Avoid Potential Copyright and Licensing Pitfalls

Thank you, everyone, I'm very glad to join today’s roundtable as a representative from China. AI and copyright has been a rapidly evolving topic, and I hope to share some key observations from the Chinese perspective, and also interact with our colleagues from the U.S. side. 

Today, I structure my sharing into two parts: 

the input side — meaning the use of copyrighted materials for AI training, 

and the output side — concerning the copyrightability and infringement issues related to AI-generated content.

Whether the use of copyrighted materials for AI training constitutes infringement (China v. U.S.) 

In China, we currently do not have AI-specific copyright legislation. However, the general principles of copyright law apply. 

Chinese courts have started to explore this issue 3 years ago, and an important case often cited is the “Ultraman” case (奥特曼案). In this case, the court emphasized that at the input stage, using others’ works for AI training serves the purpose of learning and analyzing styles, structures, and trends, rather than reproducing the original expression of those works. Therefore, if the AI only temporarily stores and analyzes works during internal training without displaying them publicly, and if the use does not affect the normal exploitation of the original work or unfairly harm the rights holder’s interests, it can be considered fair use.

Otherwise, as a principle, the use of human-created works for AI training should obtain authorization from copyright holders and provide compensation.

Copyrightability of Prompts (China v. U.S.) Another emerging issue is whether prompts given to AI models can be copyrighted.

In China, if the prompt is simple, like “draw a blue cat,” the AIGC is generally not considered a copyrightable work. However, if the prompt is complex, highly creative, and involves detailed description of style, mood, scene, and expression methods, it may be recognized as a copyrightable expression.

Some Chinese copyright scholars are strongly opposed to granting copyright to AI-generated content (AIGC). Their basic view is that AIGC itself is not a "work" in the legal sense.

They argue that if a human gives an AI very detailed, specific, and original instructions, the outcome might qualify as a literary work — but only as a literary work. Traditionally, when humans create other types of works — like music, paintings, or photography — they don’t start by writing detailed text instructions first. Different types of works can’t simply be converted from one form to anotherby providing prompts. For example, a painter doesn't write a script before painting an oil painting, and a film director’s creative notes are very different from the act of actually shooting a film.

So, no matter how detailed the instructions are, compared to the final AI-generated output, those instructions are still just ideas — not expressions.

I think this just shows a problem that traditional copyright law can't really handle. Back when copyright laws or the Berne Convention defined what counts as a "work" and what copyright protects, no one ever imagined that AI would someday get so advanced it could create things as good as — or even better than — what humans can make.

Potential infringement if outputs are substantially similar to third-party preexisting material (China v. U.S.)

In China, the basic principle for determining infringement is access plus substantial similarity.

If the AI output is substantially similar to a third-party work, and there’s evidence that the AI could have accessed the original work, it may constitute infringement.

For works that are already publicly available, courts usually presume the possibility of access — meaning, they don’t require strict proof that the AI actually "saw" the work.

As for substantial similarity, the focus is mainly on the external expression of the work, not just general ideas or styles.

So, if the AI output copies the original work’s concrete expression in a substantial way, it could be considered infringement.

In China, recent court decisions such as the "Spring Breeze" case (春风案) have clarified that if a human user engages in substantial creative activity — by selecting and designing prompts, adjusting the generation parameters, and exercising control on aesthetic expressions over the output — then the resulting work can be recognized as a human-authored work, and eligible for copyright protection.

So, In China, courts assess whether there is meaningful human creative contribution during the generation process. If the AI merely outputs content based on training without substantial human intervention, just like simply presses a button and accepts whatever the AI produces, then the output is not considered a copyrightable work. 

Industry Attitudes and Developments (China v. U.S.) Finally, let’s discuss how industries and regulators are reacting.

Based on my experience, there’s an interesting phenomenon in China that’s quite different from what I see in the U.S. Many of the traditional copyright giants and platform companiesin China are not just rights holders — they’re also investors and developers behind major AI models. Instead of rushing into lawsuits, they tend to take a wait-and-see approach, closely watching international judicial developments, domestic legislative trends, and how courts are interpreting existing rules.

When it comes to other generative AI models scraping their content pools, they’re more likely to build technical firewalls to block crawlers, rather than going straight to court.

Of course, several major lawsuits are currently pendingin China, and legislative reforms are being discussed. 

Recently, there is increasing support for statutory licensing models to regulate the mass use of copyrighted works in AI training.And collective rights management organizations are also seen as a practical solution: AI developers could pay usage fees to collective organizations, which would then distribute the income to rights holders.

While honestly, based on what I learned at an AI tech forum I attended just last week, I have some rather sad news to share: at this point, pretty much all publicly available text, music and images on the internet — and essentially all human-created works, past and present — have already been used to train the current generation of AI models. The truth is, our discussions now might already be a little late.

Closing that concludes my overview from the Chinese perspective. I’m looking forward to hearing everyone’s views and discussing how our different approaches can perhaps inspire each other and find better solutions moving forward. 

Thank you very much.

Recent Speeches and Publications by Allen Wang in the Field of Artificial Intelligence