It’s been a big week for AI video in China.
ByteDance’s Seedance 2.0 and Kuaishou’s Kling 3.0 is rapidly spreading. Seedance probably takes the cake, promoting smoother movement, stronger scene consistency, and watermark-free downloads.
My favorite trend so far is people making videos of themselves cooking LeBron James.
Meanwhile, the Cyberspace Administration (CAC) announced In the pre-Lunar New Year crackdown, it removed more than 540,000 pieces of illegal AI content on Douyin, Wechat, and Weibo, took action against more than 13,000 accounts, and issued a statement saying it was stepping up its crackdown on unlabeled or AI-generated “garbage” content.
China’s AI video ecosystem is moving at lightning speed, while the Chinese Communist Party, typically the world’s most skilled and overbearing content regulator, is scrambling to tighten control in real time. This is a bit puzzling, since China already has one of the world’s most ambitious frameworks for labeling and managing AI-generated content. Even if there are rules in place for moments like this, why do they feel like they aren’t working?
today:
-
China’s new rules on labeling AI-generated content
-
How those rules are not properly enforced
-
Domestic incentives behind this poor enforcement, such as economic growth priorities
-
Geopolitical incentives, including the use of AI to spread pro-China, anti-American propaganda
Part 2 provides a broader industry overview of how the Chinese model compares to the Western model, which companies are leading and targeting the Global South, and why video generation has different training, open source, and copyright implications for the Chinese ecosystem.
China was the first country to introduce rules specific to synthetic media. Deep synthesis regulations (mutual communication information service depth synthesis management 规defined), which took effect in the Dark Ages of 2022.
Fast forward to March 2025, the Cyberspace Administration of China Regarding display measures for AI-synthesized content (Notice of print publication “Artificial intelligence generation synthesis content identification method”). These rules require platforms to detect when content is confirmed, likely, or suspected to be AI-generated, and to attach both visible labels and embedded metadata by September 1, 2025.
Attached national standards GB 45438-2025 Let’s break down what these labels look and sound like.
-
Image: Visible text label covering at least 5% of the length of the shortest side
-
Audio: 120-160 wpm audio label, or Morse code style tone that stands for “AI”
-
Video: Clears the labels that appear at both the beginning and end of the clip.
-
Metadata: Persistent, machine-readable tagging embedded in files
China is essentially the only major country to have enacted the law. preemptivean upstream requirement that all AI-generated videos be labeled, watermarked, and traceable across the stack.
Most other countries still reactivea harm-based model. South Korea and England Criminalizes certain uses of deepfakes if they cause harm (primarily sexual or reputational). Singapore Elections (Online Advertising Integrity) Amendment Act It bans deepfakes in election ads and carries penalties of up to five years in prison. Russian Federal Law No. 32-FZ Increase criminal law penalties for disseminating “false information” about the military, including deepfakes and synthetic media used to “misrepresent” military operations in Ukraine. The United States has little more than a patchwork of state rules regarding the abuse of political advertising and intimate images (e.g. take it down act), but Google and Sora voluntarily watermark their content.
The closest comparison to China is EU AI lawrequires labeling of AI content starting August 2025. Decentralized execution Enforcement across member states with different regulatory frameworks is likely to lead to fragmented enforcement across the EU, limiting consistent enforcement.
China’s rules were scheduled to come into effect on September 1, 2025. But if you go to Chinese social media today, you’ll find tons of AI-generated videos that are completely unlabeled.
You should see a label similar to the following in the corner of all AI-generated content.
You may also see a message at the bottom that says, “This content may be generated by AI.”
Content creators can also self-report when using AI.
The government wants platforms to automatically detect this through video metadata. But in reality, only Chinese AI companies are obligated to include compatible identifying information in their metadata, and many Chinese models tend to avoid being disruptive, so Sora and Veo-generated clips are extremely popular on Chinese social media and typically produce the most entertaining content. Therefore, when you export and upload, the platform often cannot reliably determine whether a clip is a composite. And China’s social media platforms are locked in fierce competition with each other and with Western markets, with no one willing to be the strictest enforcer, while others are allowing content to circulate freely.
As mentioned at the beginning, CAC say We removed over 540,000 pieces of illegal AI content and took action against 13,000 accounts. That sounds big until you roughly compare 34 million videos posted on TikTok or 100 million throughout the meta In just one day! At this amount, 540,000 clips is well under 1% of the daily output. Mathematically, it’s just a drop in the bucket.
Maybe I’m being harsh. The announcement is meant to signal that the CAC is not trying to single-handedly weed out laxity in platforms, but rather that it takes regulating AI content seriously. But at this scale, there is a risk that it sends the opposite message, making CAC appear less in control, especially when a company like ByteDance does something that it explicitly promotes. Video without watermark.
There is also a large amount of content online that is completely non-identifying.
AI-generated ancient towns, subtly distorted Chinese locations, and incredible landscapes that fool the untrained eye (all without labels) are common. My mom sometimes sends me clips and asks, “Have you ever been here?”
I haven’t.
Also, a trend I’ve noticed is half-authentic, half-AI-generated content, further blurring the precedent for whether or not that content should be labeled. (This is consistent with Part 2’s discussion of how Chinese companies are succeeding with AI-enhanced editing tools like CapCut, rather than purely text-to-video generative models.) Scrolling through RedNote, you’ll quickly find clips where the hills and buildings in the shots are clearly real footage, but the clouds are moving and distorting in undeniably artificial ways.
sauce.
What gives?
as Eileen Chan we mourned within us China AI in 2025, Rap:
“AI-assisted and AI-generated content is much more prevalent online than it was nine months ago, both on global platforms and on the Chinese internet. Now is the time to ask: What was the point of labeling it as a policy? Is it actually to protect users from misinformation and create trust, or is it a band-aid measure to avoid platform liability? What kinds of AI uses merit what kinds of mandatory disclosures?”
One rationale is that China is tightening regulations on misinformation while simultaneously pushing generated video as a pillar of its AI+ strategy for digital commerce and economic revitalization, but this is a fine line.
National media are constantly extolling the potential of AI-generated content. A new economic stimulant. The Chinese Communist Party is partly betting on AI video as the core of e-commerce, short-form entertainment, and platform growth. Microdramas, digital human product demos, and automatically generated marketing clips are widespread.
The government itself is creating slop. Chinese broadcasters are deploying AI avatar news anchors, and the Chinese Communist Party is easily broadcasting AI content through official CCTV programs. for example, A timeless poem of praise Chiaki episodeconverting classic poetry into ink animation using the “CCTV Listening Media Model”.
Similar tactics are also underpinning China’s AI push in the Global South, which accounts for the bulk of its AI video user base. If the company’s video generation ecosystem restricts content too tightly, it will struggle to attract new users outside China who expect the creative freedom offered by international competitors.
There have been reports of Chinese state institutions experimenting with deepfake presenters to promote pro-China and anti-American narratives in local languages, while others are promoting multilingualism.AI anchorIt can run 24/7 and auto-generate product videos and livestream scripts for cross-border commerce (although we could only find a few documented cases of such practices, and there was no sign of widespread adoption).
The Chinese government also doesn’t care if the content helps cheer them up.
AI-generated clips mocking foreign rivals tend to circulate without too much trouble. This year, we’ve seen a flurry of videos about dissatisfaction with U.S. H-1B visas, military action in Venezuela, and overweight U.S. factory workers struggling to revive manufacturing in the wake of President Trump’s tariffs. Content that flatters China’s self-image and reinforces the well-known national narrative.
^ A college student finally saved up enough money for an H1B visa. sauce.
^ “Live camera footage” of US soldiers arresting Mr. Maduro. sauce.
I have no doubt that the Chinese Communist Party’s system is effective in blocking content it deems potentially destabilizing. You won’t see Mao and Chiang Kai-shek sharing beers or setting a world Olympic pole vault record (as I always do on my Instagram feed). And political dissent of any kind is rarely seen. But as for the remaining slumps, Beijing appears to be willing to tolerate some of them, especially if they can contribute to ecological growth and serve other interests.
A final aesthetic observation.
After spending far too much time on Chinese social media “researching” this article, I’ve come to believe that AI videos are generally viewed in a more optimistic light than the pessimistic perspective that pervades many of the comments on Western platforms.
For example, it’s very common to find cyberpunk-style depictions of the future that blend ancient Chinese civilization with radical futuristic cityscapes. I’m sure this exists somewhere on Western social media beyond my filter bubble, but it doesn’t seem to be all that noticeable.
this chinese futuristic A disposition, a willingness to embrace rapid technological progress in the realm of utopian, or at least neutral, rather than inherently dystopian, imbues this technological trajectory with an even greater sense of wonder. It inspires an eagerness to see what kind of world might be built, rather than just fixating on the reality it might encompass.
Don’t worry, there’s plenty of brain-rot slop as well.
*Bonus clip: ChinaTalk team tries out Kuaishou Kling’s image-to-video capabilities at NeurIPS 2025:
