China regulates AI through a layered set of instruments: the Interim Measures for Generative AI Services, the algorithm recommendation rules, the deep synthesis provisions and, since 2025, the measures on labelling AI-generated content. Each carries a registration or filing duty, and each is enforced through platform inspection and the periodic publication of filed services.
This section sets out those duties as they apply in practice — what must be filed, which documents carry the weight, how long a filing takes, and what follows when a service changes. It is written for compliance and product teams that need a decision rather than a summary of the statute.
Key topics include algorithm filing, generative AI filing and registration, deep synthesis filing, large-model security assessment, content labelling, face recognition filing, on-device AI registration, and the enforcement campaigns that test them.
In July 2026 the Cyberspace Administration of China deployed a four-month 'Qinglang — Minors Online Protection' special action, whose first phase is the 2026 summer campaign on the online environment for minors. The action focuses on AI 'poison animation' and the hyping of child influencers. This article reviews the regulatory red lines drawn by the special action and what they mean for companies providing AI content and interaction services.
The Interim Measures for the Administration of Anthropomorphic Interactive AI Services (Order No. 21 of five departments) took effect on 15 July 2026 — China's first departmental rules dedicated to emotionally companionable AI. The Measures define the scope of anthropomorphic interactive services, set out six categories of prohibited generated content, lay down special rules for minors, and specify when a security assessment is triggered. This article sets out the scope of application, core obligations and compliance arrangements from an enterprise perspective.
A foreign-owned AI company offering a service to the Chinese public faces the same filing obligations as a domestic one, but the practical path.
On 15 July 2026 the Cyberspace Administration of China announced seven newly filed on-device generative AI services for mobile handsets, including Apple Intelligence.
On 13 July 2026 the National Industrial Information Security Development Research Centre under the Ministry of Industry and Information Technology launched a co-construction initiative for large-model safety evaluation benchmarks, built around six dimensions — content safety, value alignment, robustness, fairness, privacy and truthfulness — and 31 specific risks. This article sets out what the benchmark means for AI companies and how to prepare.
On 6 July 2026 the Cyberspace Administration of China reported on the first phase of the 'Qinglang — Rectifying AI Application Disorder' special action.
On 3 July 2026 the Cyberspace Administration of China re-opened public consultation on the revised draft of the Measures for the Administration of Internet Information Services.
On 18 June 2026 the National Financial Regulatory Administration issued guidance on the secure development and use of AI in banking and insurance, requiring financial institutions to build a full-lifecycle management system, strengthen access management for high-risk applications, and reinforce data security and personal information protection.
To support the 'Qinglang — Rectifying AI Application Disorder' special action, the CAC Reporting Centre opened a dedicated reporting channel for 'AI application disorder' from June 2026.
A user agreement is not an annex at the bottom of a registration page. In a generative AI product it carries the service rules, the allocation of.
On 3 June 2026 the Cyberspace Administration of China set out the thinking behind its 'small, fast, flexible' approach to regulating anthropomorphic AI interaction services.
On 29 May 2026 the Cyberspace Administration of China issued a notice launching a two-month 'Qinglang — Optimising the Business Online Environment and Rectifying Malicious Speculation about Enterprises' special action. It targets four categories of prominent problem and, for the first time, expressly lists AI-generated false corporate information. This article reviews the action and what it means for AI content producers and companies.
In May 2026 provincial cyberspace administrations continued to publish filing and registration notices under the Interim Measures for Generative AI Services, with Guangdong, Jiangsu and Zhejiang releasing new batches including vertical video and spatial-intelligence models. This article reviews the difference between filing and registration and the practical points.
On 30 April 2026 the Cyberspace Administration of China issued a notice launching a four-month 'Qinglang — Rectifying AI Application Disorder' special action nationwide, reported in the media from 1 May. The action focuses on fourteen problem categories and proceeds in two phases, the first being the 'Qinglang — Typical Violations in AI Application Services' campaign. This article reviews the governance framework and what the categories mean for companies' self-assessment.
On 28 April 2026 the Cyberspace Administration of China disclosed that CapCut, Maoxiang and Jimeng AI had failed to effectively implement synthetic-content labelling requirements and were penalised by cyberspace authorities. This article reviews the regulatory basis, common violations and self-inspection points.
On the evening of 20 April 2026 Moonshot AI released and open-sourced the Kimi K2.6 model, further improving long-horizon coding and agentic execution, with the model going live on third-party inference platforms the next day. This article reviews the compliance points in commercial use of open-source models — licences, downstream filing and ownership of results.
On 2 April 2026 ten ministries including MIIT jointly issued the Measures on AI Science-and-Technology Ethics Review and Services (Trial) (Gong Xin Bu Lian Ke [2026] No. 75), setting out the scope, responsible bodies, working procedures and supervision of AI ethics review, and distinguishing general, simplified, expert re-review and emergency procedures. This article reviews the key points and their practical implications.
A defamation case in which the defendant argued that 'the video copy was written by AI' has drawn attention. The court held that users of generative AI services bear a duty to verify information when producing and publishing AI-generated content; publishing without citing sources or checking accuracy amounted to infringement. This article reviews the reasoning and the compliance points for content publishing.
On 17 March 2026 the Cyberspace Administration of China announced that, as of 28 February 2026, 796 generative AI services had completed filing and 481 generative AI applications or functions had completed registration, with 48 new filings in January and February. This article reviews the data and the practical points for applicants.
On 12 March 2026 the Cyberspace Administration of China published the sixteenth batch of domestic deep-synthesis service algorithm filings, covering 779 algorithms, including the Doubao enterprise voice-podcast model algorithm, the Qwen image-recognition-generation algorithm and the Turing Zhixin text-generation algorithm. This article reviews the filing rhythm and self-inspection points.
In February 2026 the Cyberspace Administration of China carried out a campaign against accounts that publish AI-generated synthetic information without adding an AI label and use false content to deceive and mislead the public. Some 13,421 accounts were disposed of in accordance with law and agreement, over 543,000 pieces of unlawful information were cleared, and typical cases were reported. This article reviews the enforcement points and directions for corporate self-inspection.
On 1 February 2026 the Measures for the Supervision and Administration of Live-Streaming E-Commerce took effect.
On 31 January 2026 the Cybercrime Prevention Law (Draft for Comment) was released for public comment through the cybersecurity bureau of the Ministry of Public Security.
On 27 January 2026 the Beijing AI company Moonshot AI released and open-sourced its new Kimi K2.5 model, built on a trillion-parameter mixture-of-experts architecture and natively trained on 15 trillion mixed vision-and-text tokens. This article reviews the compliance points of open-source models from the angles of licence, downstream filing and ownership of results.
Need legal advice on this topic?
We help foreign AI and tech companies navigate China's regulatory landscape. Get a clear, actionable assessment from bilingual lawyers who know both the law and the industry.
Request a Free Consultation →No sales pitch. Just clear advice from lawyers who have handled cases like yours.