The Label Mandate
China's Framework for AI Content Transparency
You open Douyin—known internationally as TikTok—on your phone and scroll to a video. In the lower left corner, a semi-transparent label catches your eye: “This work contains AI-generated content.” It is small, unobtrusive, but unmistakably present. For hundreds of millions of Chinese internet users, this label has become an increasingly common sight since September 2025.
This is not a platform feature that creators can opt into or ignore. It is a legal requirement. Under China’s Measures for the Labeling of AI-Generated and Synthetic Content, effective since September 1, 2025, anyone who uses AI to generate content must declare it, and platforms must display that declaration prominently. The regulation covers four categories of responsible parties: AI content generation service providers, internet content distribution services, app store platforms, and individual users. What appears as a simple label is, in fact, the visible endpoint of a comprehensive regulatory framework that governs how AI-generated content moves through the Chinese internet.
Mandatory Labeling: How the System Works
The core logic of China’s approach is straightforward: AI-generated content must be labeled, and non-compliance carries consequences. The national standard GB/T 45438-2025 and the labeling regulation together establish what officials call a “dual-track” system of explicit and implicit identification.
Explicit labels are what users see: text, icons, or audio cues that appear on the content itself. For videos, the standard requires that these labels appear at the start of playback and remain visible or periodically reappear throughout. The technical specification mandates that visual labels occupy at least five percent of the image height, ensuring they cannot be easily overlooked or cropped out. Platforms like Douyin display these as semi-transparent overlays in the corner of the video frame. Kuaishou places additional indicators in the title bar beneath the video. The goal is consistent: users should know they are watching AI-generated content before becoming immersed in it.
Implicit labels operate beneath the surface. AI service providers must embed metadata into generated files, including the provider’s identity, a unique content ID, and a precise timestamp. This information travels with the file, enabling traceability even when content is shared across platforms or downloaded and re-uploaded elsewhere. The technical standard requires that these embedded markers withstand common transformations like compression and transcoding. Some providers, like ByteDance’s Jimeng AI, use frequency-domain watermarking, a technique that embeds information into the video’s spectral data rather than its visible pixels, making it resistant to editing and screen recording.
The regulatory framework assigns obligations across the entire content chain. Platforms that distribute content must detect and preserve these labels. They cannot strip out metadata during server-side processing. Users who post AI-generated content without proper labeling face penalties, and platforms that fail to enforce the rules risk regulatory action. Fines for violations can reach 100,000 yuan, and serious cases may result in criminal liability.
This design represents a clear departure from voluntary approaches. The system does not rely on creators’ good faith or platforms’ self-governance. It mandates compliance and backs that mandate with active enforcement.
Enforcement in Action
The strength of any regulatory framework lies in its implementation. China’s AI content labeling system has been tested through sustained enforcement campaigns that demonstrate the government’s commitment to making the rules meaningful.
In April 2025, the Cyberspace Administration of China launched the “Clear and Bright” campaign targeting AI technology misuse. By June, the first phase had yielded substantial results: over 3,500 non-compliant AI products were dealt with, more than 960,000 pieces of violating content were removed, and approximately 3,700 accounts were penalized. Regional offices developed specialized approaches suited to local conditions. Beijing established a “user marking, platform verification, joint enforcement” model. Shanghai pushed labeling compliance across more than 400 enterprises. Zhejiang alone intercepted over 25 million items of non-compliant content. Jiangsu identified and addressed 163 AI-related risk domains.
Major platforms implemented parallel enforcement measures. Tencent processed over 570 non-compliant programs. Weibo removed more than 4,800 pieces of violating content. Douyin developed what it calls a “red-blue confrontation” mechanism, essentially internal teams competing to identify and defeat AI content detection evasion, to improve the accuracy of its automated systems. Alibaba and Kuaishou advanced the deployment of metadata reading and writing infrastructure to comply with the national standard.
In January 2026, the National Radio and Television Administration launched a targeted campaign against “AI-modified” videos, particularly those that manipulated classic television dramas, historical subjects, and portrayals of revolutionary figures. Within one week, seven major platforms collectively removed approximately 4,871 violating videos: WeChat (1,078), Douyin (1,075), Kuaishou (1,057), Bilibili (over 700), Xiaohongshu (487), Baidu (284), and Weibo (over 190). Accounts responsible for repeated violations faced suspension or permanent closure.
These numbers indicate that the labeling framework functions as more than a paper requirement. Platforms face real pressure to comply, and creators face real consequences for evasion.
Different Paths: China, Europe, and the United States
How does China’s approach compare with regulatory developments elsewhere? The contrast is instructive.
China’s framework operates through mandatory national standards with explicit technical specifications. The regulation took effect in September 2025 and applies to AI service providers, content platforms, app stores, and individual users. The coverage is comprehensive, addressing text, images, audio, and video. The technical standard specifies precisely how labels should appear and how metadata should be structured. Enforcement has already produced measurable results across multiple campaigns.
The European Union’s AI Act includes transparency requirements for AI-generated content, particularly for deepfakes, but the relevant provisions will not take effect until August 2026. The current approach relies primarily on voluntary codes of practice while the detailed implementing rules are developed. The scope is narrower than China’s framework, focusing on AI system providers and deployers, and does not extend obligations to individual users or general content distribution platforms. The EU’s definition of what requires labeling is also more restrictive, centering on content that “resembles existing persons, objects, places, entities or events.”
In the United States, no federal mandatory framework exists for AI content labeling. The primary mechanism is the Coalition for Content Provenance and Authenticity (C2PA), an industry-led initiative that develops open standards for content authentication. Major technology companies including Adobe, Microsoft, and OpenAI participate. However, membership is voluntary, enforcement mechanisms are absent, and adoption among content creators remains limited. There is no timeline for federal legislative action on mandatory labeling requirements.
The structural differences are significant. China has established binding rules with detailed technical requirements and demonstrated enforcement capacity. The European Union has announced intentions but faces an implementation delay. The United States has deferred to industry self-regulation without binding requirements.
Why Mandatory Regulation?
The urgency behind China’s approach becomes clearer when considering the problems it aims to address.
In February 2024, a finance employee at the Hong Kong branch of a British multinational company received a video call. On the screen appeared the company’s CFO and several colleagues, all requesting an urgent wire transfer for a confidential transaction. The faces looked right. The voices sounded right. Over the following days, the employee made fifteen transfers totaling HK$200 million, approximately US$25 million. Five days later, the company discovered that every person in that video call had been a deepfake, created using publicly available footage from the company’s YouTube channel and media appearances. Hong Kong police described it as the first case they had encountered involving the simultaneous fabrication of an entire executive team.
This case was not an isolated incident. According to the Internet Finance Association of China, AI-based face-swapping fraud caused direct financial losses exceeding 1.8 billion yuan in 2025. The technology that enables such fraud has become remarkably accessible, with deepfake services available online for as little as 9.9 yuan. The combination of low barriers to misuse and high potential for harm explains the regulatory response.
The labeling requirement does not eliminate deepfake fraud. Criminals who use AI for deception will not voluntarily label their creations. But the regulation establishes a baseline expectation: legitimate AI-generated content should be identifiable as such. When content lacks proper labeling, both platforms and users have grounds for heightened scrutiny. The system creates friction for those who would pass off AI content as authentic.
A Framework Tested at Scale
China’s approach demonstrates that mandatory AI content labeling can be implemented at scale without halting industry development. The same period that saw enforcement campaigns against non-compliant content also saw continued growth in China’s AI sector.
As of 2025, over 730 large language models have been registered with Chinese authorities at national and provincial levels. The user base for generative AI products has reached 230 million. These figures suggest that regulatory requirements and industry expansion can coexist. The AI labeling framework adds compliance costs, but it has not prevented the sector from growing.
The semi-transparent label that now appears on Douyin videos represents more than a compliance checkbox. It is the visible manifestation of a regulatory framework that spans content generation, platform distribution, and individual posting. The framework operates through technical standards that specify how labels should appear and how metadata should be embedded. It operates through enforcement campaigns that process millions of violations. And it operates through penalties that give platforms and users reasons to comply.
For countries grappling with questions of AI governance, China’s experience offers a reference point. The approach is neither perfect nor universally applicable. Detection systems sometimes flag human-created content incorrectly, frustrating digital artists and creators who work with sophisticated editing tools. Technical standards will need revision as AI capabilities evolve. But the basic proposition has been tested: mandatory labeling of AI-generated content is implementable, enforceable, and compatible with a growing AI industry.
That label in the corner of a Douyin video is easy to overlook. But what it represents, an entire system designed to make AI content visible rather than hidden, deserves attention.