Contents7
On 15 July 2026, China began enforcing its rules on AI services that imitate a person to keep users company: a reminder every two hours, virtual relationships banned for minors, conversation data protected. They come on top of the mandatory labelling of AI-generated content, in force since 1 September 2025, and the article on AI that the Cybersecurity Law has contained since 1 January 2026.
This post in our Business AI section takes stock as of 20 October 2026, based on the texts published by the Chinese authorities. It is not legal advice: for a service open to the public in mainland China, a local lawyer has the final say.
In short
Any AI-generated content distributed in mainland China has carried a visible label and an invisible mark in its metadata since 1 September 2025.
Distribution platforms read those marks and display a notice; users who publish generated content must declare it.
Since 15 July 2026, AI “companions” must remind users that they are machines, prompt them after two hours of use and offer no virtual relationship to minors.
A customer service chatbot, a question-and-answer AI or a work assistant falls outside that last text.
Labelling: a notice you see, a mark you don’t
The measures on labelling AI-generated content were published on 14 March 2025 by the Cyberspace Administration of China and three other authorities, and have applied since 1 September 2025, according to their official text. A mandatory national standard on the labelling method has applied since the same date, notes the Deep Lex regulatory tracker.
The text distinguishes two marks. The explicit label can be seen or heard: text, sound or a graphic placed on the content or in the interface. The implicit label is written into the file’s data, invisible to the user: the generated nature of the content, the provider’s name, a content number. When a tool offers to download or export generated content, the file must keep its explicit label, the Xinhua news agency details.
The path of generated content under China’s measures
- The provider
Its tool adds the visible label and writes the mark into the metadata.
- The user
Declares the content as generated when publishing it.
- The platform
Reads the metadata and displays a notice next to the content, confirmed, possible or suspected.
The platform grades its notice by what it knows: “generated” if the metadata says so, “possibly generated” if only the user declared it, “suspected” if it detects other traces. A provider may supply content without a visible label to a user who asks for it, provided the terms of use hand that user the responsibility and the logs are kept for at least six months. Deleting, altering or forging these marks is forbidden for everyone (articles 6, 9 and 10 of the text).
AI companions: the rules of 15 July 2026
The interim measures on anthropomorphic interactive services were published on 10 April 2026 by five authorities, the Cyberspace Administration announced. The full text we quote is a copy published on a website of the Chinese Ministry of Commerce (mofcom.gov.cn), not the Cyberspace Administration’s own page; the passages used here are our translation, not an official one. The text covers services open to the public in mainland China that imitate a person’s personality or way of speaking in an ongoing emotional interaction (article 2). It expressly excludes customer service, question answering, work assistants, education and research when they do not sustain that kind of relationship.
of continuous use trigger a reminder to the user
registered users, or 100,000 monthly active users, require a security assessment
is the maximum fine when the breach endangered life or health
The rest of the text describes a closely watched service. Users must know they are talking to an AI, and a reminder appears if they show signs of dependence (article 18). Intimate virtual relationships, “relatives” or “partners”, are banned for minors, a minor mode is mandatory, and children under 14 need a parent’s consent (article 14). Conversations may not be passed to a third party without explicit consent, users can copy and delete them, and their sensitive data only trains a model with their separate consent (article 16).
These questions are not unique to China: we raised them for Europe in our article on AI companions and teenagers.
The Cybersecurity Law has addressed AI since 2026
China’s Cybersecurity Law was amended on 28 October 2025, with effect from 1 January 2026. Its new article on AI states the government’s support for research, training data and computing power, and provides for ethical rules, risk monitoring and safety oversight, according to Xinhua. It is a general framework: the concrete obligations remain in targeted measures, such as those on labelling and companions.
Europe and China: the same ideas, two methods
European Union (AI Act)
Machine-readable marking by the tool’s provider
Deepfakes disclosed by the business that publishes them
Users told when they talk to an AI, since 2 August 2026
Tools already on the market: deadline of 2 December 2026
Mainland China
Visible label and metadata mark, by the provider
Declaration by the user, notice displayed by the platform
Specific rules for AI companions, since 15 July 2026
In force for everyone since 1 September 2025
Europe gives the marking duty to the provider, under article 50 of the AI Act, and only asks businesses to disclose their deepfakes; China puts the label on the whole chain, down to the platform that displays the content. The European detail is in our guide to marking AI-generated content, and the question of whether to say a text comes from an AI in should you disclose AI-written text.
The dates, from 2025 to today
- 14 March 2025
The measures on labelling generated content are published.
- 1 September 2025
Labelling and its mandatory national standard apply.
- 28 October 2025
The Cybersecurity Law is amended and gains an article on AI.
- 1 January 2026
The amended law takes effect.
- 10 April 2026
The measures on AI companions are published.
- 15 July 2026
Those measures apply.
- 17 October 2026
Consultation closes on a regulation on minors’ internet use, which also covers service design.
What it changes for a team of 20 to 500 people
Three cases, from the most common to the rarest:
- You publish generated images, videos or texts on Chinese platforms: declare them when publishing, and do not strip the metadata your tool wrote into them. An agency or contractor publishing on your behalf must do the same.
- Your website or app is visited from China: if you distribute generated content aimed at the Chinese public, labelling may apply, and making the site load properly there is a separate technical project. For language and markets, see selling in China and Southeast Asia.
- You open a generative AI service to the Chinese public: it goes through a filing with the Cyberspace Administration, which counted 796 such services on 28 February 2026 according to Deep Lex, a private regulatory tracker: that is a secondary source, which we have not cross-checked against the lists published by the authority, and through the companion rules if it is one. That is a project in itself, with local counsel.
From the field
In the routing table we keep for our own translation pipelines, Chinese included, the rule has been written down since August 2026 (internal rule, as it stands in our table on 20 October 2026): a public text may go through any provider in the table, and personal data never goes to a provider established outside the European Union without a contractual framework, whatever its nationality.
Every European and American deadline is gathered in our AI regulation calendar, and the overall approach in our guide to integrating AI in your business. To find out which of your content and tools are affected, market by market, our digital and AI maturity audit starts from your actual uses.
Frequently asked questions
Does AI-generated content have to be labelled in China?
Yes, since 1 September 2025. The tool’s provider adds a visible label and a mark in the file’s metadata; the user who publishes the content on a platform must declare it; the platform displays a notice next to the content. Deleting or forging these marks is forbidden, including for the business that publishes the content.
Is a customer service chatbot covered by China’s rules on AI companions?
No, as long as it does not build an ongoing emotional interaction with the user. The measures of 15 July 2026 expressly exclude customer service, question answering, work assistants, education and research. The content that chatbot generates is still subject to labelling, like any AI-generated content distributed in mainland China.
Do China’s rules apply to a European business?
They target services and content offered to the public in mainland China, whatever the company’s country. A European business that publishes on Chinese platforms, or opens an AI service there, is therefore concerned for that activity. A website that is merely accessible from China is a greyer case, to be assessed case by case with local counsel.
How do the AI Act and China’s labelling rules differ?
The AI Act requires providers to apply machine-readable marking and businesses to disclose their deepfakes, with a deadline of 2 December 2026 for existing tools. Since September 2025, China has required a visible label and an invisible mark on all generated content, and also involves the user who publishes and the platform that displays it.
Sources
- Their official text (new tab)cac.gov.cn
- Notes the Deep Lex regulatory tracker (new tab)deep-lex.com
- The Xinhua news agency details (new tab)news.cn
- The Cyberspace Administration announced (new tab)cac.gov.cn
- A copy published on a website of the Chinese Ministry of Commerce (new tab)dzswgf.mofcom.gov.cn
- According to Xinhua (new tab)news.cn
- Article 50 of the AI Act (new tab)ai-act-service-desk.ec.europa.eu
A copy published on a website of the Chinese Ministry of Commerce
Read on dzswgf.mofcom.gov.cn(new tab)Jim · founder of Brain Plus AI
Also in business ai
Understand9 min
AI Act and digital rules calendar: deadlines 2026 to 2028
AI Act, Data Act, Cyber Resilience Act, NIS2, e-invoicing in France and the first US state laws: every date that matters for a business, from autumn 2026 to August 2028, on one page kept up to date. For each deadline, who is covered and what to prepare.
Practical guide7 min
AI-generated content: mandatory marking by 2 December
Since 2 August 2026, the AI Act requires images, video, audio and text generated by AI to be marked, and tools already on the market have until 2 December 2026 to comply. Who marks, who discloses, what the European code of practice recommends and what a detector really proves.
Stepping back8 min
Are AI companions dangerous for teenagers?
An American survey indicates that most teenagers have already talked to an AI companion, but its figures come from a campaigning non-profit and call for caution. The design question, though, is a legitimate one, and Europe has already answered it in part.