China's AIGC Labeling and Registration Rules for AI Short Dramas: A 2025–2026 Timeline (for Creators Publishing in China)
Three rule sets, three dates. 1 Sept 2025: the Measures for Labeling AI-Generated and Synthetic Content and mandatory standard GB 45438-2025 took effect; AI video needs a visible label on the opening frame and an implicit label in file metadata. 1 April 2026: unregistered back-catalogue animated micro-dramas were pulled. 1 July 2026: the NRTA's AI Micro-Drama tiering standard applies, splitting projects at RMB 800,000 and RMB 300,000 between national review, provincial filing and platform self-review, with an “AI-assisted production” title card once AI use crosses the threshold. China only; other markets differ.
The short version
Three separate instruments govern AI short dramas in mainland China. They sit at different points in the pipeline and are enforced by different bodies.
The Measures for Labeling AI-Generated and Synthetic Content (jointly issued by the Cyberspace Administration of China, MIIT, the Ministry of Public Security and the NRTA; published 14 March 2025, in force 1 September 2025, 14 articles) govern what the content itself must carry: an explicit, visible label and an implicit, machine-readable label. Obligations fall on three parties: the generation service provider, the distribution platform, and the user.
GB 45438-2025, “Cybersecurity technology — Labeling method for content generated by artificial intelligence”, is the mandatory national standard that operationalises the Measures. It was published 28 February 2025 and applies from the same date, 1 September 2025.
The NRTA’s tiering rules govern whether a drama may be distributed at all. From 1 July 2026, AI micro-dramas are sorted into three tiers by investment and subject matter, each with its own review body. Before that, a special campaign on animated micro-dramas launched in November 2025 required existing titles to complete registration by 31 March 2026; unregistered ones came down on 1 April.
This article is a dated map of what each rule requires. It is not legal advice. The official texts control; consult counsel for a specific project.
Timeline: February 2025 to July 2026
| Date | Event | What it means for a production team |
|---|---|---|
| 28 Feb 2025 | GB 45438-2025 published | The technical form of labels is fixed: explicit plus implicit |
| 14 Mar 2025 | Labeling Measures published | Duties split among generator, platform and user |
| 1 Sept 2025 | Measures and standard take effect | Opening frame must carry a prominent notice; metadata must carry an implicit label |
| Aug–Sept 2025 | Douyin, Kuaishou and Bilibili ship labeling and self-declaration features | An “AI-generated” declaration toggle appears at upload |
| Nov 2025 | NRTA launches a campaign on animated micro-dramas; they are brought under registration | AI dramas adapted from web novels, comics or stickers are in scope |
| Jan 2026 | One-month intensive enforcement period | Platforms purge non-compliant videos |
| 31 Mar 2026 | Deadline for back-catalogue titles to complete registration | Unregistered titles are no longer accepted |
| 1 Apr 2026 | Unregistered existing titles delisted; platforms had removed ~29,000 videos and 40+ accounts by 31 Mar | Audit the registration status of anything already live |
| 25 Jun 2026 | NRTA Network Audio-Visual Department issues the Management Notice (AI Micro-Drama Classification and Tiering Standard) | Tier is set by investment plus subject matter |
| 1 Jul 2026 | Tiering standard in force | Top tier needs a distribution licence; “AI-assisted production” title card once AI use crosses the line |
Dates from the CAC, The Paper, 21st Century Business Herald and China Daily; checked 2 September 2026.
Explicit and implicit labels: where they go
Article 4 of the Measures says video must carry a prominent notice on the opening frame and in an appropriate position around the player; a notice may also be added at the end or mid-way. For an episodic drama, that means each episode shows a visible AI notice from its first frame, and the title, description or badge area shows one too.
Article 5 requires an implicit label in the file’s metadata: the content’s AI-generated attribute, the service provider’s name or code, and a content identifier. Viewers never see it; platforms read it to verify.
GB 45438-2025 defines the method for both label types. It is a “GB” rather than a “GB/T” standard, which makes it mandatory, and it binds both generation service providers and distribution platforms.
The common mistake is treating the explicit label as a corner watermark. The Measures require a prominent notice and fix the opening frame as a required position. Cropping the tool’s watermark and typing “made with AI” into the description does not satisfy Article 4.
Who is liable: generator, platform, user
The Measures assign duties to three roles. A production team usually occupies two of them at once.
- Generation service providers (the companies operating the video, image or voice models): add both labels; explain the labeling method and the user’s duties in the terms of service.
- Distribution service providers (Douyin, Kuaishou, Bilibili, Hongguo and similar): under Article 6, verify file metadata, add a prominent notice to content carrying an implicit label, label content suspected of being AI-generated, and offer a labeling feature while reminding users to declare.
- Users (the account publishing the drama): under Article 10, declare proactively and use the platform’s labeling feature when publishing AI-generated content. No one may maliciously delete, alter, forge or conceal a label.
A team publishing in China is a “user” of the model service and a “user” of the platform, and must declare on both sides. It is not itself a “service provider” unless it exposes model capabilities to third parties. Where an overseas team distributes through a Chinese partner, the partner platform performs the verification step; the declaration duty stays with whoever uploads.
What the “remove watermark” toggle actually means
Most tools offer a watermark-off option. Article 9 of the Measures allows a user to request content without an explicit label, on two conditions: the provider must set out the user’s labeling duties and responsibilities in the user agreement, and must retain logs of who received such content for at least six months.
Operationally: turning off the watermark does not mean the clip is exempt. It moves the job of adding the explicit label from the tool to you. The finished episode still needs the Article 4 notice at its head, and the upload still needs a declaration. The implicit metadata label is unaffected by the toggle and should still be written by the tool; do not strip metadata when you edit, stitch or transcode.
Conversely, taking a labeled export and processing it to remove the label before publishing falls squarely under the Article 10 prohibition.
The NRTA tiers: investment plus subject matter
On 25 June 2026 the NRTA’s Network Audio-Visual Department issued the Management Notice (AI Micro-Drama Classification and Tiering Standard), effective 1 July. It is the first rule set aimed specifically at fully AI-generated and AI-assisted micro-dramas. Two dimensions decide the tier:
- Key micro-dramas: investment of RMB 800,000 or more, or a main plot touching politics, the military, foreign affairs, national security, united-front work, ethnicity, religion, the judiciary or public security. Full filing with the NRTA and a network drama distribution licence before release.
- Ordinary micro-dramas: investment between RMB 300,000 and 800,000, general subject matter. Filing and review by the provincial broadcasting authority.
- Other micro-dramas: investment under RMB 300,000, general subject matter. Platform self-review, with periodic reporting to the local regulator.
The three tier names reuse the existing micro-drama framework; the thresholds are set specifically for AI titles. A second hard rule: if AI use in any one of scriptwriting, rendering or voice synthesis reaches the prescribed threshold, the episode must open with an “AI-assisted production” card. This is the NRTA’s own marking requirement and runs alongside, not instead of, the explicit label under the Measures.
Most AI short dramas land in the third tier, but a sensitive theme puts a title in the key tier regardless of budget. Filing is done by the Chinese production or distribution entity; overseas teams typically go through a partner. How “investment” is counted follows the regulator’s interpretation.
How the platforms implement it
Around 1 September 2025 the major platforms updated their rules:
- Douyin: a content labeling feature that helps creators add the notice, plus a metadata read/write feature for AI content.
- Kuaishou: labeling and user self-declaration, with an “AI-generated” explicit label on platform-generated content and an implicit label written to metadata.
- Bilibili: a labeling option at upload for quick self-declaration; undeclared AI content is labeled by the platform under community rules.
- Hongguo: in April 2026 it issued, jointly with Douyin, creation guidelines for animated micro-dramas and has publicly removed non-compliant AI dramas.
The common pattern: an “AI-generated” toggle at upload; leave it off and the platform labels the content for you. What follows (reduced distribution, takedown, account penalties) varies by platform and is governed by each platform’s current rules.
Checklist, and a note on other markets
| Stage | Action | Basis |
|---|---|---|
| Production | Log AI involvement per episode, separately for script, rendering and voice | Tiering standard |
| Master | Prominent notice on the opening frame; “AI-assisted production” card if a threshold is crossed | Measures Art. 4; tiering standard |
| Files | Preserve the metadata label; do not strip it when stitching or transcoding | Measures Art. 5; GB 45438-2025 |
| Tools | Read the user agreement before disabling watermarks; keep your own generation logs | Measures Art. 9 |
| Registration | Determine the tier; file key and ordinary tier titles before release | Tiering standard |
| Upload | Tick the AI declaration on every platform; never remove or alter labels | Measures Art. 10 |
| Back catalogue | Check whether registration was completed by 31 Mar 2026; relist only after filing | NRTA campaign notice |
All of the above is China-specific. The EU’s AI Act Article 50 transparency duties apply from 2 August 2026 with a different mechanism, and the US has no equivalent federal labeling mandate as of the check date; see how China, the US and the EU treat AI video. Glossary: AIGC label, content moderation.
Key figures and sources
| Figure | Value | Source | Date |
|---|---|---|---|
| Labeling Measures effective date | In force 1 Sept 2025, 14 articles; video needs a prominent notice on the opening frame and around the player (Art. 4) and an implicit metadata label (Art. 5) | www.cac.gov.cn | 2026-09-02 |
| GB 45438-2025 | Mandatory national standard, published 28 Feb 2025, effective 1 Sept 2025; defines explicit and implicit labeling methods | openstd.samr.gov.cn | 2026-09-02 |
| AI micro-drama tier thresholds | ≥RMB 800k or sensitive themes: NRTA; RMB 300k–800k: provincial; <RMB 300k: platform self-review; effective 1 Jul 2026 | m.thepaper.cn | 2026-09-02 |
| Back-catalogue delisting | Registration deadline 31 Mar 2026; unregistered existing titles delisted from 1 Apr 2026; ~29,000 videos removed by 31 Mar | www.21jingji.com | 2026-09-02 |
| Unlabeled output conditions | User agreement must set out responsibilities; provider keeps recipient logs for at least six months (Art. 9) | www.cac.gov.cn | 2026-09-02 |
FAQ
We only used AI for voices or a few shots. Do we still label?
Yes. The Measures apply to “AI-generated and synthetic content” without a percentage exemption; any AI-generated portion in the master triggers the Article 4 notice and the Article 10 declaration. The NRTA’s “AI-assisted production” card is a separate test: it fires when AI use in script, rendering or voice individually crosses the threshold.
Does the label have to say “AI-generated” in those exact words?
The Measures require a “prominent notice”; wording and style follow GB 45438-2025 and platform specifications. Typical practice is an opening card such as “This work contains AI-generated content” plus the platform badge. Position (opening frame) and legibility matter more than a fixed phrase.
Who files the registration?
The producing or distributing entity, according to tier: key tier with the NRTA (with a distribution licence), ordinary tier with the provincial authority, other tier through platform self-review and local reporting. Foreign teams generally file through a Chinese partner. Document lists are published by the regulator and platforms.
Is turning off the tool’s watermark a violation?
Not by itself. Article 9 permits unlabeled output where the agreement sets out responsibilities and logs are kept. The violation comes later: a master without the explicit notice, an upload without a declaration, or removal of an existing label.
Do these rules apply to our releases outside China?
No. They apply to distribution in mainland China. Other markets have their own regimes; the EU’s AI Act Article 50 applies from 2 August 2026 and works differently. A series released in both places must satisfy both. See the three-jurisdiction comparison.
What happened to titles that were live before April 2026?
The campaign required existing compliant titles to complete registration by 31 March 2026; anything unregistered was pulled from 1 April. Delisted titles can be relisted after filing, following each platform’s procedure.
What if we forget to tick the declaration?
Under Article 6 the platform labels suspected AI content itself. Further consequences (reduced reach, takedown, account penalties) differ by platform. Declaring is a statutory duty of the user, not an option.
Sources
- https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
- https://openstd.samr.gov.cn/bzgk/std/newGbInfo?hcno=F32EA2A561F1886CD8D606513512D547
- https://www.tc260.org.cn/portal/article/2/20250828165129
- https://m.thepaper.cn/newsDetail_forward_33643733
- https://qiye.chinadaily.com.cn/a/202607/01/WS6a448790a310d709c2fbb3c2.html
- https://www.21jingji.com/article/20260410/herald/055ad9a327c3eed4dc4a0aa93f0863f2.html
- https://m.thepaper.cn/newsDetail_forward_31523045
- https://www.stdaily.com/web/gdxw/2025-09/02/content_393775.html
- https://digital-strategy.ec.europa.eu/en/factpages/quick-facts-transparency-rules-ai-systems