China’s latest campaign against misuse of artificial intelligence has produced a set of enforcement figures that show how quickly content moderation has become part of the country’s AI rollout. Xinhua reported on September 2 that the Cyberspace Administration of China, or CAC, said authorities had removed more than 5.61 million unlawful or rule-violating AI-generated items, acted against more than 49,000 accounts, and dealt with more than 2,400 websites and apps.
The campaign is not presented as a narrow takedown effort. The CAC said its focus includes false information made or amplified with AI, violent or vulgar outputs, synthetic impersonation of real people, and material that harms minors. Those categories connect model providers, social platforms, app stores, and individual accounts in a single enforcement chain. They also make the distinction between producing a model and policing its downstream use increasingly difficult to maintain.
The size of the reported cleanup gives the campaign a different scale from earlier compliance actions. China already had generative-AI rules in force from August 2023, and its content-labeling rules became effective in September 2025. The September 2 figures therefore describe enforcement under an existing regime, rather than the unveiling of a new statute. The latest action follows a period in which the government has also pushed public-sector and commercial adoption, including a 32-city AI data-annotation pilot and a national goal to expand AI service providers.
CAC Sets Out the Campaign’s Four AI Misuse Targets
The CAC’s stated targets are broad enough to encompass familiar social-media problems and newer forms of synthetic-media abuse. Fabricated material can include information made to look authentic through AI tools. Impersonation reaches the increasingly accessible ability to alter faces or clone voices. The regulator also cited content involving violence, pornography, vulgarity, and risks to minors.
The focus matters because Chinese platforms now host a growing number of consumer AI functions, not just conventional user posts. Xinhua identified services including ByteDance’s Doubao, Alibaba’s Qwen, Baidu’s Ernie Bot, and Tencent’s Yuanbao as having stepped up review of training inputs and outputs while working to meet the labeling requirements. The report did not say that the named companies were penalized. Instead, it listed them among services taking compliance measures in response to the regulatory environment.
The CAC’s platform examples ran from Taobao, Douban, Zhihu, Xiaohongshu, and Bilibili to Baidu, Tencent, Weibo, Kuaishou, and Douyin. It said these services had expanded multimodal checks and databases for altered faces, copied voices, and previously identified violations. That list underscores that the challenge extends well beyond standalone AI applications. Video, commerce, search, social networking, and app-distribution businesses all sit within the same policy perimeter.
In practical terms, the campaign brings two different content problems together. One is the policing of public-facing posts or videos after publication. The other is the design of controls before a model’s output reaches a user. The latter is especially significant for platforms racing to offer agents, image generation, or voice features because safety checks have to operate across text, images, audio, and video rather than a single content format.
Labeling Rules Move From Policy Text to Platform Operations
China’s 2025 labeling regulations cover machine-produced copy, images, sound, and video, and call for identifiers that are visible or embedded. The September campaign gives those rules an operational setting: a regulator can now assess whether providers are labeling outputs and whether platforms can trace or detect manipulated content when it is uploaded or shared.
That is a different task from removing ordinary misinformation. A synthetic clip may travel across several services, be cropped or re-encoded, and be separated from the original prompt or account. The CAC’s emphasis on multimodal detection suggests it expects companies to make judgments based on several signals, including known samples and face or voice manipulation. Xinhua did not disclose technical benchmarks, accuracy rates, or a common detection standard, so those should not be inferred from the campaign totals.
Beijing and Shanghai were cited as examples of regional approaches. Xinhua said Beijing authorities directed platforms to improve review systems so they could identify new variants of prohibited material. Shanghai regulators emphasized internal audits and public-facing materials that explain AI risks. The distinction illustrates how a national framework can lead to different implementation priorities at the local level.
The attention to internal audits also connects this campaign to a wider pattern of compliance expectations around AI products. China’s recent customer-service standard similarly focuses on operational safeguards, including disclosure and human handoff, rather than merely setting high-level principles. In both cases, regulators are pressing companies to build governance into day-to-day product systems.
The Enforcement Numbers Leave Important Questions Open
The CAC’s totals are large, but they do not by themselves reveal how many items were removed by automated filters, human reviewers, user reports, or orders from regulators. Nor do they show how violations were distributed among platforms, regions, or categories of abuse. A count of accounts acted against also does not establish how many were individual users, commercial operators, or repeat offenders.
Those gaps are relevant for companies and developers. The core policy signal is clear: AI-generated content is now being treated as a large-scale moderation problem with specific enforcement categories. Yet the available report does not provide a public scorecard for particular models, platforms, or detection tools. It is therefore not possible to rank companies by compliance based on the September 2 announcement.
For users, the campaign raises a more immediate issue. Labels and account actions can make synthetic material easier to identify or remove, but they cannot eliminate the need to evaluate claims, images, and voices with care. For platform operators, the CAC’s figures show that enforcement is reaching every major distribution channel, from chatbot outputs to app-store listings.
The campaign also sits alongside a policy agenda that continues to promote AI applications. China has set a target of at least 3,000 AI application-service providers in a national pool by the end of 2027, as EastFrontier reported in its examination of the MIIT service-provider goal. The September enforcement numbers show the other side of that push: more AI services are encouraged, but providers and platforms are expected to absorb more responsibility for the information those services create and circulate.