On September 17, 2026, officials of China's National Radio and Television Administration (NRTA) said micro-drama users had passed 800 million. They said the sector was worth more than 100 billion yuan in 2025, that about 430,000 titles were released in the first eight months of 2026, and that more than 90% of them were AI-generated. The same officials said 68,000 rule-violating micro-dramas had been removed this year, according to Global Times. They also said AI dramas must be labeled and platforms must respect portrait and voice authorization.
Those numbers show a regulator facing a new production model. They are also a useful test of what proportionate rules for synthetic media look like.
What the measures actually do
The NRTA published a draft of the Micro-Short Drama Development Management Measures for comment on June 24, 2026, with feedback open until July 23. The final text was issued as NRTA Order No. 16 and took effect on September 1, 2026.
The measures define a micro-drama as a serial under 20 minutes per episode with a clear storyline and continuous plot. They sort titles into three categories by investment size and subject matter. Higher tiers face filing, public notice and a distribution licence before release. The lowest tier is handled largely through platform-level review.
Platforms must verify licensing credentials, run content-review systems, guard against addiction-driven algorithm design and remove prohibited material. Penalties run from warnings to fines of up to 100,000 yuan and suspension of broadcast licences. AI-generated or AI-produced dramas must carry a visible indicator in every episode.
The strongest case for the regime
The case for tough rules is real. When more than 90% of 430,000 releases are synthetic, human review cannot scale. Fake endorsements, cloned voices and unlicensed likenesses are cheap to produce and hard for a victim to trace. Micro-dramas are consumed on phones by audiences that include minors, and the sector's monetisation pattern of paywalled cliffhangers rewards engagement over care. A regulator that ignored this would be failing its basic duty. The 68,000 removals suggest the problem is not hypothetical.
Where the design is sound
The AI labeling duty is the most defensible element. Disclosure tells viewers what they are watching without banning any technique. It works whether a drama is made with a camera or a model. It also fits the direction of other jurisdictions that have chosen transparency over prohibition for synthetic media, which is a proportionate tool.
Likeness and voice authorization has a similar logic. Using a real person's face or voice without consent is a rights violation, not a creative choice. Enforcing consent protects performers and lets legitimate AI production grow on clean inputs. The NRTA also said live-action dramas remain the mainstay of quality production, but did not suggest that AI production should be suppressed, according to MyDramaList's summary of the announcements.
Where the risks sit
The first risk is the licensing tiers. A pre-release filing and licence regime suits a slow film industry. It fits poorly with a market that shipped roughly 1,800 titles a day on average across the first eight months. If higher tiers depend on investment size and subject matter, producers face incentives to under-report budgets or steer toward lower tiers. That pushes compliance effort into paperwork rather than into content quality.
The second risk is that platforms become the front line. Duties to verify credentials, audit content and remove violations within short windows tend to produce over-removal. Under liability this broad, a platform's rational response is to delete anything ambiguous. That hurts small creators, who cannot afford legal review, more than the large studios.
The third risk is vague content categories. The measures list prohibited content that includes material that endangers national unity or spreads false information. Such categories are difficult to apply consistently to fiction. A drama is not a news report, and treating invented plots as potential misinformation invites uneven enforcement and self-censorship. A liberal approach to speech would keep prohibitions narrow, defined and reviewable, and would separate genuine harms such as piracy, fraud and harms to minors from political judgement.
A better calibration
The evidence so far supports the measures' core, which is disclosure, consent and platform accountability for illegal content. It does not yet support heavy front-end licensing for everything. Three adjustments would help.
- Keep licensing for high-risk categories only. Rely on filing and post-hoc audit for low-budget titles, and publish the tier thresholds so producers can plan.
- Make labeling machine-readable. A visible mark on each episode is useful. Standard metadata would let platforms and viewers filter synthetic content without relying on manual checks.
- Publish enforcement data. The 68,000 removals need a breakdown by reason, such as piracy, likeness misuse, minors or political content. Without it, the public cannot judge whether the rules target real harms.
The NRTA also said it is accelerating a radio and television law, per Global Times. Elevating these rules into statute is a chance to fix ambiguities: define the categories, set proportionate penalties, and give producers a clear appeal route. If that happens, China could have a workable model for labeling and consent in AI-generated video. If it does not, the sector's speed advantage may erode under compliance costs that fall hardest on small producers.
The broader lesson applies well beyond China. Regulators handling AI video do best when they regulate deception and rights violations directly, and avoid rules that treat every new production method as a threat to be licensed.