China and South Korea AI labeling rules: what they require and who they bind

This page covers China and South Korea in depth, and India, Vietnam and a few other places briefly. It is general information, not legal advice. Chinese and Korean sources are in those languages: where we quote a Chinese text we give our own plain-English summary beside it, which is not an official translation. Where we could not reach the official text, or rest on law firm or news summaries, the row says so. For the rest of the map, start at the AI disclosure laws overview.

The rules at a glance

China: the Labeling Measures

The Measures for Labeling AI-Generated Synthetic Content were issued by the Cyberspace Administration of China (CAC), the Ministry of Industry and Information Technology, the Ministry of Public Security and the National Radio and Television Administration. The notice is dated March 7, 2025 and was published on the CAC site on March 14, 2025. Article 14 says they apply from September 1, 2025. The key provisions, in our plain-English summary:

• Who is bound (Article 2). Online information service providers that fall under China's algorithm recommendation, deep synthesis and generative AI rules, when they carry out AI content labeling.
• Two kinds of label (Article 3). An explicit label is added to the content or the interactive interface, presented as text, sound or graphics, and clearly perceivable by users. An implicit label is added by technical means to the content's file data and is not easy for users to notice.
• Explicit labels (Article 4). Providers add them by content type: text prompts at the start, end or middle of text; spoken prompts or audio rhythm cues for audio; a prominent prompt in a suitable place on an image; prompts on the opening frame and around the playback area of a video; and, for downloads, copies and exports, the file must contain the explicit label.
• Implicit labels (Article 5). Providers add them in file metadata, carrying the content's attribute, the provider's name or code and a content number.
• Platforms (Article 6). Content distribution platforms check file metadata for implicit labels. Where metadata says the content is AI-generated, they add a prominent notice near it. Where a user declares it, they add a notice that it may be AI-generated, and where they detect other generation traces they treat it as suspected.

The article that reaches users, and the one that forbids removing labels, is Article 10. The first paragraph reads:

用户使用网络信息内容传播服务发布生成合成内容的，应当主动声明并使用服务提供者提供的标识功能进行标识。

In plain English: users who publish AI-generated content through a content distribution service must proactively declare it and use the label function the service provides. The second paragraph reads:

任何组织和个人不得恶意删除、篡改、伪造、隐匿本办法规定的生成合成内容标识，不得为他人实施上述恶意行为提供工具或者服务，不得通过不正当标识手段损害他人合法权益。

In plain English: no organization or individual may maliciously delete, alter, forge or conceal the labels the Measures require, or provide tools or services for doing so. That is one reason GotAlt's tools never strip or remove markings.

The Measures contain no fine schedule. Article 13 sends violations to the competent authorities to be handled under other laws, regulations and rules:

违反本办法规定的，由网信、电信、公安和广播电视等有关主管部门依据职责，按照有关法律、行政法规、部门规章的规定予以处理。

The Measures also say a provider may supply output without an explicit label when the user agreement sets the user's labeling duty, and then must keep logs for at least six months (Article 9).

Sources: Measures for Labeling AI-Generated Synthetic Content, CAC (Chinese text); Interim Measures for Generative AI Services, CAC (Chinese text). We read the CAC page through a tool that returns verbatim fragments, so the quotes are exact but are not the whole text. Verified October 21, 2026.

GB 45438-2025: the numeric figures are unverified

GB 45438-2025, a cybersecurity standard on the method for labeling AI-generated synthetic content (our translation of its title), is a mandatory national standard on the labeling method for generative service providers and content distribution providers. Per front matter reported in search results, it was issued on February 28, 2025 and took effect on September 1, 2025. We could not reach the official publication, so we have not read the standard from its official source.

Source: Standards reseller excerpt of GB 45438-2025 (not the official publication). Verified October 21, 2026.

Does it reach a foreign seller?

This is our reading, not a quoted source, and no source we read says otherwise.

• A foreign company that offers a generative AI service, such as a tool, app or API, to the Chinese public is within the Interim Measures for Generative AI Services whatever its location. Their Article 20 says the cyberspace authority should notify the relevant bodies to take technical and other measures against a service from outside China that does not conform.
• A foreign seller that only uses AI imagery on its own website is not a "service provider" in the Measures' sense.
• A foreign seller that posts AI content on a Chinese platform has the Article 10 duty to declare it and use the platform's label function, and the platform adds its own labels under Article 6.

No source we read reports a foreign small business being fined under these rules. We found no Taobao, Tmall or Douyin rule that requires labels on merchant product images.

Enforcement in China

On April 28, 2026 the CAC published a notice naming the CapCut (Jianying) app, the Maoxiang app and the Jimeng AI website. It said they had not effectively carried out the labeling rules, found breaches of the Cybersecurity Law, the Interim Measures and the Labeling Measures, and directed local offices to apply interviews, orders to correct, warnings and strict handling of the responsible staff. The notice states no fine amount. A CAC page of September 2, 2026 about the second phase of a campaign to rectify AI application problems names Doubao, Yuanbao, Qianwen and Wenxin Yiyan as required to strengthen labeling. Xinhua reported on May 8, 2026 that inadequate labeling was one of the 14 problems the campaign listed.

This is the only confirmed enforcement of an AI labeling rule that we found anywhere, and it is directed at services and platforms. We also found claims, none sourced, of fines of three times ad spend for unlabeled AI ads in China, of a requirement for foreign brands to embed a Chinese social credit code in watermarks, and of 2024 penalties against ByteDance for labeling. We do not repeat them.

Sources: CAC notice of April 28, 2026 (Chinese text); CAC page of September 2, 2026 (Chinese text); Xinhua, May 8, 2026 (Chinese text). Verified October 21, 2026.

South Korea: the AI Basic Act, Article 31

The Framework Act on the Development of Artificial Intelligence and Establishment of Trust, the AI Basic Act, has been in force since January 22, 2026. Article 31 on transparency, in the official English translation:

"(1) An AI business operator that intends to provide a product or service using high-impact AI or generative AI shall notify the user in advance that the product or service is operated based on that AI."

"(2) An AI business operator that provides generative AI or a product or service using it shall indicate that the output was generated by generative AI."

"(3) Where an AI business operator provides outputs, such as synthetic sound, images, or video, that are difficult to distinguish from the real ones, by using an AI system, the operator shall notify or indicate in a manner that users can readily recognize the fact that such outputs have been generated by the AI system."

For artistic or creative works, the fact may be notified or indicated in a manner that does not hinder their exhibition or enjoyment. Details go to a Presidential Decree.

Who is bound

An "AI development business operator" is "A person that develops and provides AI", and an "AI use business operator" is "A person that provides AI products or AI services using AI provided by a business operator". Article 4(1) says the Act "shall apply to any conduct outside the Republic of Korea if the conduct impacts the domestic market or users". A business that only uses AI output in its own marketing is not obviously an operator that provides an AI product or service to users. Secondary reports of the Ministry of Science and ICT's support-desk casebook say exactly that: a business that creates content with another company's AI and posts it is not providing an AI service to users and is not subject to the transparency obligation, and the duty covers content generated after January 22, 2026. We did not read the Ministry's own text.

Fines, and the grace period

Article 43(1) says that "Any of the following persons shall be subject to an administrative fine not exceeding 30 million won", and lists a person who "fails to provide the notification, in violation of Article 31(1)", a person who fails to designate a domestic representative where required, and a person who fails to comply with an order to cease or correct a violation under Article 40(3). Failing to indicate generated output under Article 31(2) or (3) is not itself a listed ground. It becomes fineable only after the Minister orders a stop or correction under Article 40(3), where "the Minister may order the AI business operator to take necessary measures to cease or correct the violation", and the operator does not comply.

Law firm and news summaries say the Ministry told businesses that fact-finding and fines are suspended for at least one year from January 22, 2026, while the legal duties still apply. We found no official Ministry text for that. On those reports the earliest fine would fall around January 2027, and only for the Article 31(1) notice or after a corrective order. We found no enforcement action. Bills to relax Article 31, and bills to add labeling duties for platforms and advertisers, were reported pending as of October 4, 2026, and are not law.

Source: AI Basic Act, official English translation (Korea Legislation Research Institute). The grace period, the Enforcement Decree and the Ministry's transparency guideline rest on secondary sources. Verified October 21, 2026.

South Korea: AI virtual persons in ads

The Korea Fair Trade Commission amended its Review Guidelines on Endorsements and Testimonials in Labeling and Advertising, with a release posted on May 31, 2026 for a June 1, 2026 start. We read the release's title, date and agency on the Korean government briefing site; its body is in attachments we could not open. Secondary reports say AI-generated virtual persons became a fifth category of endorser beside consumers, celebrities, experts and organizations. They say text media should state in the title or opening that the post includes an AI-generated virtual person, for example 가상인물 포함, and that photo and video ads should show the label 가상인물 near the person while the person appears. They also say that a label does not cure a misleading claim: presenting a virtual person as having real experience of a product can still be an unfair advertisement. We do not know whether the guideline reaches a foreign advertiser. The penalties come from the Fair Labeling and Advertising Act, and we did not verify the figures a blog gives.

Source: Korea Fair Trade Commission release on the endorsement guideline (Korean text). Verified October 21, 2026.

India and Vietnam, briefly

We could not reach the official texts for either country, so this section rests on secondary sources and says so.

• India. Amendments to the IT (Intermediary Guidelines and Digital Media Ethics Code) Rules, reported as notified on February 10, 2026 and in force from February 20, 2026, define "synthetically generated information" as audio, visual or audiovisual content artificially or algorithmically created or modified that appears real and depicts a person or event so that it is likely to be perceived as indistinguishable from a real one. Routine editing and accessibility features done in good faith are excluded, and AI-generated text is not covered. The duties fall on intermediaries: those enabling creation or dissemination must label lawful synthetic content prominently and embed permanent metadata to the extent technically feasible, and significant social media intermediaries (over 5 million registered Indian users) must require a user declaration. A seller's own website is not an intermediary, so a seller meets these rules as a user of platforms. A draft of October 2025 would have set a 10 percent label size, and the final rules dropped it for a "prominent" standard. An April 2026 draft asking for continuous labels was reported, and we found no report that it was notified.
• Vietnam. Law on Artificial Intelligence 134/2025/QH15 was passed on December 10, 2025 and has been in force since March 1, 2026. Decree 142/2026/ND-CP, issued on April 30, 2026, has applied since May 1, 2026. In an unofficial English translation, Article 3(5) defines a deployer as an organization or individual that "uses an artificial intelligence system under its/his/her control in professional, commercial or service provision activities", and Article 11 asks deployers to add easily recognizable labels to AI-generated or edited audio, images and video that imitate real people or recreate real events, with a lighter treatment for film, artistic and creative works. The decree, as reproduced by a private database, exempts technical edits that improve quality without changing the essence or context, and text tools for spelling, grammar, summarizing or translation. We found no fine amounts in the statute text; a news summary gives figures we could not trace. Whether a foreign seller is covered is untested.

Sources: Shardul Amarchand Mangaldas, legislative update on the India amendment (PDF, secondary); Vietnam Law on Artificial Intelligence, unofficial translation on LuatVietnam. Verified October 21, 2026.

Other places we checked

We found no binding AI labeling duty for ordinary sellers in these places. Several rows rest on secondary sources, and the table says which.

Verified October 21, 2026.