A flood the old rules couldn't catch
On September 1, 2026, China's National Radio and Television Administration (NRTA) brought into force the Measures for the Development and Management of Micro-Short Dramas — Order No. 16, adopted by the agency's office meeting on July 27 and published via Xinhua (news.cn). It is China's first dedicated regulatory framework for the micro-drama format: serialized episodes under 20 minutes, built for vertical-scroll viewing on Douyin, Kuaishou and dedicated apps.
The timing tracks the industry's growth curve. In the first half of 2026 alone, China released roughly 367,000 micro-dramas, and more than 74% of them were AI-generated — a share that has climbed from single digits a year earlier as generative tools cut per-episode costs from hundreds of thousands of yuan to a few thousand (InsideAI News). The market is projected to top 120 billion yuan (~$16.5 billion) in 2026, overtaking China's theatrical box office, and now reaches roughly 660 million users (The Next Web).
What the rules actually require
Three provisions matter most. Article 34 requires that any episode created or produced with AI carry a "prompt identification in a conspicuous location." Article 37 bars platforms from deploying recommendation algorithms that "induce user addiction or excessive consumption," and requires them to periodically evaluate and prioritize quality content over engagement-optimized feeds. And ahead of formal review, platforms must pre-screen AI-generated submissions for repeated synthetic faces, copyright conflicts, and visual defects (NRTA notice; InsideAI News).
The part that gets less attention outside China is the licensing architecture underneath all of it. The Measures sort every micro-drama into three tiers by budget and subject matter. Category One (big-budget, or anything touching politics, military affairs, diplomacy, national security, ethnicity, religion, or law enforcement) and Category Two (mid-budget, general themes) both require content review and distribution approval from provincial authorities before broadcast. Only Category Three — low-budget, general-interest work — is left to platform self-review with an assigned identification code (Caixin Global).
The case for it, stated fairly
Regulators aren't inventing a problem. A market producing 1,500-plus new AI titles a day, three-quarters synthetic, has real failure modes: NRTA's own list of prohibited content — material promoting "money worship, obscenity, gambling, or discrimination" — reads like a response to things that were actually airing. Copyright disputes over AI models trained on lifted footage, and synthetic "actors" indistinguishable from real ones without disclosure, are legitimate consumer-protection concerns, not hypotheticals. And an addictive-recommendation ban targeting a format explicitly engineered for compulsive vertical scrolling is a narrower, more defensible intervention than a blanket content quota — it addresses a specific design pattern rather than picking winners by genre or nationality.
Where the proportionality breaks down
The AI-labeling mandate is the right instrument: cheap to comply with, informative to viewers, agnostic about what gets made. It's the kind of rule most jurisdictions experimenting with synthetic-media disclosure have converged on, because it preserves the market's ability to sort good AI content from bad while fixing the actual information asymmetry — viewers not knowing what they're watching is real.
The classification-and-pre-approval system is a different animal. Requiring provincial government sign-off before broadcast for anything above Category Three isn't a disclosure rule; it's a licensing gate applied to a format whose entire commercial logic is speed — AI compression means a title can go from concept to publish in weeks, sometimes days. A prior-review requirement sized for state television doesn't scale gracefully to a format built around iteration at that tempo, and the vagueness of what counts as "substantial investment" or general "themes" for Category Two gives provincial reviewers wide discretion over which projects get slowed down and which don't. That discretion tends to reward incumbents — established studios and platforms with compliance staff and existing relationships with provincial regulators — over the single-creator AI shops the cost collapse was supposed to enable in the first place.
The addictive-algorithm ban has the same structural problem in miniature: "induce addiction or excessive consumption" is a standard almost impossible to define ex ante, which functionally hands platforms an incentive to over-comply by suppressing engagement broadly rather than targeting the specific dark patterns (autoplay chaining, artificial cliffhanger pacing) that actually cause the harm.
None of this means the underlying concerns are illegitimate — synthetic-face disclosure and copyright screening are reasonable asks of an industry moving this fast. But a regime that pairs a genuinely proportionate transparency rule with a genuinely disproportionate licensing one ends up doing two different things at once: informing viewers, and slowing down exactly the low-cost, high-volume production model that made China's micro-drama boom possible. Only one of those was the stated goal.