China's labelling measures require AI-generated content to carry labels, but a label does not travel with pasted text. The rules took effect on 2025-09-01. A label can be lost the moment someone copies the words out, so detecting from the writing fills the gap.
What the rules require
The Measures for Labeling AI-Generated Synthetic Content (人工智能生成合成内容标识办法) were issued on 2025-03-07 and have applied since 2025-09-01. They call for two kinds of label:
- Explicit labels. Notices visible to users.
- Implicit labels. Marks in the file metadata.
The supporting national standard is GB 45438-2025. The measures were issued jointly by the Cyberspace Administration of China, the Ministry of Industry and Information Technology, the Ministry of Public Security and the National Radio and Television Administration (notice 国信办通字〔2025〕2号, dated 7 March 2025), and the measures and the standard both took effect on 1 September 2025, according to the CAC's official release.
How DeepSeek handles labels
DeepSeek's terms of use say it adds labels within the platform that remind users content is AI-generated, and forbid removing or concealing them. No hidden text watermark or public checker has been published. Details are on the watermark status page.
Why pasted text loses its label
A visible notice at the end of a reply is easy to leave out when you copy. Metadata belongs to a file, and plain text has none. So a text that has been copied, pasted or retyped usually has nothing for a label check to read. That follows from the format, and no vendor source states it.
What style detection adds
DeepSeekDetect reads how a text is written. It reports whether it reads as AI-written, which passages carry the signal, and how close the style is to DeepSeek, Qwen and GLM. Of fully AI-written DeepSeek-family texts, 95.1% were caught, per the fact sheet.
For documents and files, Fileverity's detector covers images, text and documents in one place.
Verified 2026-11-10. More in the fact sheet.