China’s Plan to Make AI Watermarks Happen
Audio morse codes, encrypted metadata information, or labels in virtual-reality-generated scenes. These are some of the things the Chinese government wants AI companies and social media platforms to use to properly label AI-generated content and crack down against misinformation.
On September 14, China’s Cyberspace Administration drafted a new regulation that aims to inform people of whether something is real or AI. As generative AI tools get increasingly advanced, the difficulty to discern whether content is AI-generated is causing all kinds of serious issues, from nonconsensual porn to political disinformation.
China’s is not the first regime to tackle this issue—the European Union’s AI Act, adopted this March, also requires similar labels; California passed a similar bill this month. And China’s previous AI regulations also briefly mentioned the need for gen-AI labels.
However, this new policy outlines more details of how AI watermarks should be implemented by platforms. For the first time, it also promised to punish social media platforms where AI-generated content is posted and travels far without being properly classified. As a result, there are a lot more financial and legal stakes for AI companies and social platforms if they are tempted to take the shortcut and not instate proper labeling features.
With the speed and proactiveness of its AI legislation, China is hoping to be the domineering regime shaping global AI regulation. “China is definitely ahead of both the EU and the United States in content moderation of AI, partly driven by the government’s demand to ensure political alignment in chatbot services,” says Angela Zhang, a law professor at the University of Southern California studying Chinese tech regulations. And now it has another chance at shaping global industry standards, because “labeling is a promising area for global consensus on a certain technical standard,” she says.
Policing AI Is Harder Than It Looks
For the first part, the new draft regulation asks AI service providers to add explicit labels to AI content. This means watermarks on images, “conspicuous notification labels” when an AI-generated video or virtual reality scene starts, or sounds of the morse code of “AI” (· – · ·) before or after an AI-generated audio clip.
These are, to different degrees, practices that the industry is already employing. But the legislation would change them from voluntary measures into legal liabilities and would force those AI tools with loose labeling mechanisms to catch up or face government penalties.
But the problem with explicit labels is that they are usually easy to alter, like cropping out a watermark or editing out the video ending. So the legislation also requires companies to add implicit labels into the metadata of AI-generated content files, which should include a specific mention of the initialism “AIGC” as well as encrypted information about the companies that produced and spread this file. It also recommends these companies add invisible watermarks in content so users won’t realize they are there.
In reality, the implementation of implicit labels in metadata would require a lot more companies to work together and adhere to common rules.
“Interoperable standards for metadata require that they work across AI models and deployers, tools and platforms—that’s a tall order and does have both technical challenges and costs for the change,” says Sam Gregory, the executive director of Witness, a human rights organization in New York. That would take years, not months, he says.
But perhaps the most challenging part of the Chinese regulation is how it makes social media platforms responsible for finding AI-generated content. The regulation asks “online information content transmission platform services” to examine shared files for implicit labels and AI-generation traces. The platforms have to put on a gen-AI tag/label if the metadata says it, if the user uploading it voluntarily discloses so, or if the platform suspects that it is AI. They also need to add in their own information to the metadata to show the path this content has traveled on the internet.
That creates a lot of new challenges. First of all, it’s not immediately clear which platforms will be considered “online information content transmission platform services” by the law. “Generally speaking, social media platforms like Douyin, WeChat, and Weibo are most likely covered, but it’s unclear whether ecommerce platforms like Taobao and JD and search engines like Baidu also count,” says Jay Si, a Shanghai-based partner at Zhong Lun Law Firm.
Right now, popular vertical video platforms in China allow users to tag a video as AI-generated when posting it. Some also allow users to flag other untagged videos as AI or proactively look for those videos and then put on a label that says, “The content is suspected of being generated by AI.”
But being legally required to screen everything on the platform is a game changer, considering they have hundreds of millions of users inside and outside China. “If WeChat or Douyin needs to examine every single photo uploaded to the platform and check if they are generated by AI, that will become a huge burden in terms of workload and technical capabilities for the company,” Si says. Douyin and Kuaishou, two of the most prominent social video platforms in China, declined to comment for this story.
China Plans to Go Beyond the EU AI Act
The European Union’s AI Act, usually seen as the most comprehensive legal framework for regulating AI so far, also has an article that addresses the content-labeling issue. It requires that “outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated.” It also requires explicit disclosure from companies when the content contains deepfake visuals or text information that involve public interests.
Companies have already started to police content. “A range of Western companies have started to adopt the C2PA standard which is a metadata-based provenance standard that helps surface the recipe of how AI was used in content,” says Gregory. Supporters of Coalition for Content Provenance and Authenticity (C2PA) include Google, Meta, Microsoft, and OpenAI. It’s a step in the right direction, Gregory says, but it’s not yet widely available and many platforms haven’t adopted it.
Chinese regulators likely learned from the EU AI Act, says Jeffrey Ding, an assistant professor of Political Science at George Washington University. “Chinese policymakers and scholars have said that they’ve drawn on the EU’s Acts as inspiration for things in the past.”
But at the same time, some of the measures taken by the Chinese regulators aren’t really replicable in other countries. For example, the Chinese government is asking social platforms to screen the user-uploaded content for AI. “That seems something that is very new and might be unique to the China context,” Ding says. “This would never exist in the US context, because the US is famous for saying that the platform is not responsible for content.”
But What About Freedom of Expression Online?
The draft regulation on AI content labeling is seeking public feedback until October 14, and it may take another several months for it to be modified and passed. But there’s little reason for Chinese companies to delay preparing for when it goes into effect.
Sima Huapeng, founder and CEO of the Chinese AIGC company Silicon Intelligence, which uses deepfake technologies to generate AI agents, influencers, and replicate living and dead people, says his product now allows users to voluntarily choose whether to mark the generated product as AI. But if the law passes, he might have to change it to mandatory.
“If a feature is optional, then most likely companies won’t add it to their products. But if it becomes compulsory by law, then everyone has to implement it,” Sima says. It’s not technically difficult to add watermarks or metadata labels, but it will increase the operating costs for compliant companies.
Policies like this can steer AI away from being used for scamming or privacy invasion, he says, but it could also trigger the growth of an AI service black market where companies try to dodge legal compliance and save on costs.
There’s also a fine line between holding AI content producers accountable and policing individual speech through more sophisticated tracing.
“The big underlying human rights challenge is to be sure that these approaches don’t further compromise privacy or free expression,” says Gregory. While the implicit labels and watermarks can be used to identify sources of misinformation and inappropriate content, the same tools can enable the platforms and government to have stronger control over what users post on the internet. In fact, concerns about how AI tools can go rogue has been one of the main drivers of China’s proactive AI legislation efforts.
At the same time, the Chinese AI industry is pushing back on the government to have more space to experiment and grow since they are already behind their Western peers. An earlier Chinese generative-AI law was watered down considerably between the first public draft and the final bill, removing requirements on identity verification and reducing penalties imposed on companies.
“What we’ve seen is the Chinese government really trying to walk this fine tightrope between ‘making sure we maintain content control’ but also ‘letting these AI labs in a strategic space have the freedom to innovate,’” says Ding. “This is another attempt to do that.”