ChinAI #306: Yes Labels for AI-generated Content? A Test of 23 Chinese Platforms
Greetings from a world where…
blooming things are blooming
…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).
Feature Translation: Implementation Status for China’s Regulations on AI Content Labels
Context: The rise of AI-generated misinformation on Chinese platforms has prompted calls for stronger regulations. In one case, scammers profited from fake videos of Jin Dong, known as “the most popular Chinese actor among middle-aged Chinese women.” On March 7, China issued measures for labeling AI-generated content (ChinaLawTranslate’s English translation), with an implementation timeline for September 1. For firms that provide services that “might cause confusion or mislead the public,” these regulations require them to label AI-generated content with explicit and implicit watermarks.
In this week’s feature translation, Southern Metropolis Daily’s Privacy Guard Team [隐私护卫队] reviews the new measures (link to original Chinese). They also conduct their own test of how 23 different Chinese platforms are implementing the requirements to label AI-generated content. The privacy guard team portal is under the Nandu Personal Information Protection Research Center.
Key Takeaways: A few details about these latest measures suggest that these regulations will be strictly enforced.
While the substance of these rules mostly clarify previous regulations on this issue, they are significant in that they are “regulatory documents” [行政规范性文件], which are considered generally binding.
They are also linked with a national standard developed by the important TC260 standardization body: Cybersecurity technology - Method for Identifying Synthetic Content Generated by Artificial Intelligence. Nandu identifies this standard as a mandatory standard, as opposed to a recommended one (for more on the differences between the two, see ChinAI #271).
Nandu’s evaluation of 23 Chinese platforms finds that, in general, the big players have started to implement these labeling rules.
Specifically, Nandu tested 8 social content platforms (e.g., Xiaohongshu, Weibo, WeChat public accounts) and 15 AI service providers. I found the list of AI platforms to be a useful reference point, so I’ll paste that here: DeepSeek, Tencent Yuanbao, Doubao, Kimi Smart Assistant, Nano AI Search (360’s generative AI search product), ErnieBot (Wen Xiaoyan), Zhipu Qingyan, (MiniMax’s) Xingye, (Bytedance’s) Jimeng AI, iFlytek SparkDesk, Tiangong AI, (Alibaba) Tongyi Qianwen, Kling AI, Minimax (Conch Video/Conch Voice), (Baichuan’s) Baixiaoying.
All 8 social media platforms provide functions for users to actively declare whether their post contains AI-generated content (second column in image below), though WeChat Public Accounts, Douban, and Zhihu do not have anything in their user agreements about user obligations for labeling AI-generated content (first column in image below).
All 15 major AI platforms have added clear watermarks to AI-generated content (second column on image below); 13 of the 15 have incorporated AI content labeling requirements in their user agreements (first column on image below). The article highlights Tencent Yuanbao’s efforts here, as their user agreement includes a plank that implicit watermarks will be added into any generated content.
The article also connects these measures to previous AI governance policies such as the algorithm registry: “When AI platforms perform procedures such as algorithm registry filing and security assessments, they should also provide materials related to the identification of generated synthetic content.”
These AI content labeling measures still face limitations.
Aside from AI and social media platforms, application stores also play a role. Nandu’s evaluation found: “In contrast, the verification of AI labeling materials by app store operators has not yet been implemented. After checking the review specifications for application software and inquiring with customer service, we learned that Huawei, Xiaomi, OPPO, vivo and Honor, the five first-tier Android mobile phone companies, currently have not included requirements to verify materials related to AI-generated content labeling in the qualification review details for deep synthesis and AI apps.”
Finally, it’s important to look beyond just the biggest actors. From the article: “Zhiwei Yao, a professor at the law school at Guangdong University of Finance and Economics, previously told reporters that the new regulations have high technical requirements for content dissemination platforms, and it remains to be seen whether small and medium sized content platforms will be able to effectively implement these measures.”
FULL TRANSLATION: Implementation Status for China’s Regulations on AI Content Labels
ChinAI Links (Four to Forward)
Must-read: Artificial Intelligence, China, and America’s Next Industrial Revolution
This op-ed by CSET’s Dewey Murdick and William Hannas highlights the need to balance “promote” with “protect” in tech competition with China. I liked Cole McFaul’s summary and reflections in this summary thread.
On a related note, the retrospectives on the Biden administration’s approach are already in full swing. I’ve participated in two iterations of these in the past week alone. Here’s my two cents, which remains consistent with this take back in September 2022 (ChinAI #196):
When it comes to U.S.-China technology competition, the benefits of the “promote” plank will always outweigh the “protect” plank. And, when we reflect on moves like this one thirty years from now, there’s a decent chance that such “protect” actions were counterproductive.
Should-read: Baidu Executive’s Daughter Exposes Private User Data Online, Sparking Privacy Concerns in China
In CTOL Digital Solutions, Sofia Delgado-Cheng summarizes an incident in which a Baidu VP’s 13-year-old daughter revealed personal information (doxxing or “box opening” [开盒]) about a pregnant women. This sparked concerns that she accessed Baidu internal user data.
Should-read: The Days of Wrestling with Algorithms (in Chinese)
Renwu asked readers to share their concerns about the growing prevalence of algorithms in their lives, from dictating their preferences and entrapment in a vortex of addictive content. They got responses from ordinary people from all different backgrounds.
A few more book plugs:
I joined Kaiser Kuo on the Sinica Podcast in a wide-ranging conversation that touched on the role of culture and public enthusiasm in GPT diffusion. I also enjoyed chatting with Bonny Lin about the book on CSIS’s ChinaPower podcast.
If you’re around Princeton on Monday March 31, I’ll be giving a book talk sponsored by the Center on Contemporary China.
Thank you for reading and engaging.
These are Jeff Ding's (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.
Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).
Also! Listen to narrations of the ChinAI Newsletter in podcast format here.