That strikingly lifelike short video you scrolled past recently may not star a real person at all; it may be generated by artificial intelligence. In the past such content was hard to tell from reality, but on many platforms today it has been quietly marked.
Content already tagged
A label visible to the eye, or hidden inside the file, is becoming the new standard for AI content.
The source of this labeling system is the Artificial Intelligence Generated Content Labeling Measures, effective September 1, 2025. Issued jointly by the Cyberspace Administration of China and three other departments, it requires text, images, audio, and video generated by AI to carry a label telling users "this was made by a machine."
One visible change for ordinary users is this: when you open an AI video, a line of text may appear in the corner noting the content was generated by AI. It does not change the viewing experience, but turns the question of "is this machine-made" from guesswork into a clear notice, giving people a bit more sobriety.
Watermarking and filing, two tracks
Labels come in two kinds. One is an explicit label, placed directly in the image, sound, or text as a visible prompt so users notice at a glance. The other is an implicit label, written into the file's underlying data, usually invisible, yet traceable when needed.
Together, they act like a manual attached to the content and a tracing chip installed inside it.
The measures also bring distribution platforms into regulation. Platforms cannot only distribute; they must check whether files contain implicit labels, and when they detect suspected generated content, they must add a prominent prompt around where it is published.
From generation to distribution, every link is strung together by this labeling thread, trying to answer "which is generated, who generated it, and where it came from."
Explicit labels are easy to delete, while implicit labels are hard to erase afterward; the two complement each other to support credible tracing. That is also why the measures require that no organization or individual maliciously delete, alter, forge, or hide these labels, nailing the bottom line down with rules.
Responsibility does not fall only on companies. The measures require that when users publish generated content through distribution services, they should actively declare it and use the labeling functions the service provides. From generator to platform to user, the net of responsibility is complete only when every ring is linked.
A new clause in the cybersecurity law
Between encouraging innovation and guarding against hazards, legislation is also adding a deeper pillar. On October 28, 2025, the Standing Committee of the National People's Congress passed a decision to amend the Cybersecurity Law, adding a clause that specifically mentions artificial intelligence.
This clause stresses that the state supports AI research and development and infrastructure building, while improving ethical norms and strengthening security regulation. The amended law took effect on January 1, 2026.
The tone of this new clause is to weigh development and security equally. It states both that the state supports foundational AI research and key technological breakthroughs, and that ethical norms must be improved and security monitoring and assessment strengthened. For the industry, this is a clear signal: innovation is encouraged, but boundaries are also drawn.
Seen globally, China has written a dedicated AI clause into a foundational law. It lifts AI from scattered departmental rules to coordination at the legal level, and its signal value is no less than the specific articles themselves, leaving a superior-law basis for subsequent detailed rules.
Nearly a thousand services filed
Filing and registration are like a dynamically updated roster. According to an announcement by the Cyberspace Administration of China, from May to June 2026, 120 generative AI services were newly filed, and 68 applications or functions were newly registered.
As of June 30, 2026, a cumulative 988 generative AI services had been filed, and 598 applications or functions had been registered.
The meaning of these numbers lies not in their size, but in depicting a vast ecosystem being brought under management. Services with opinion-forming or social-mobilization attributes must go through filing or registration;
Applications already online should also display the model name and filing number used in a prominent place, letting users choose with full knowledge.
Filing and registration target two different kinds of objects. Filing applies to the generative AI services themselves, while registration applies to applications or functions that directly call the capabilities of already-filed models through interfaces. Tiered management holds the base while also covering the many applications above it.
Global regulation converges
Looking globally, labeling and regulation are forming a converging understanding. The European Union's Artificial Intelligence Act took effect on August 1, 2024, regarded as an early and systematic comprehensive framework, assigning obligations by tier of hazard.
South Korea's Artificial Intelligence Basic Act took effect on January 22, 2026, another comprehensive AI law after the European Union.
China's approach is summarized as prudent tolerance and classification by tier. It does not want to crush innovation at the start, yet must hold the red line for key-area applications. The three pillars—the labeling measures, the Cybersecurity Law amendment, and filing and registration—are exactly the handles that turn this approach into practice, letting technology grow in a predictable environment.
By contrast, the United States relies more on executive orders and state legislation, without yet a unified federal framework; the European Union and South Korea chose to roll out with comprehensive laws at once. The routes differ, but all head in the same direction: letting AI run in the sunlight and keeping users informed.
To ordinary people, these statutes may feel distant, yet they decide whether every piece of AI content you see online will be honestly labeled.
Labeling is the start of trust
Labeling itself is not the endpoint. It solves the first step of "letting users know"; real trust still depends on sustained tracing ability, clear allocation of rights and duties, and penalties against malicious deletion or forgery of labels. Only when labeling becomes a habit need people no longer doubt everything on screen.
Tagging content is, in truth, returning trust to people. AI will keep getting more lifelike, but with labels as a safety net, viewers at least know they are talking to a machine. This transparency is a plain yet necessary sense of security in the age of human-machine coexistence.
Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.