← Back to blog
AiAbout 6 min read

China's Broadcast Regulator Goes After AI Micro-Dramas, Starting With Stolen Faces

Published Nov 14, 2026
China's Broadcast Regulator Goes After AI Micro-Dramas, Starting With Stolen Faces

On November 6, China's National Radio and Television Administration announced a new round of enforcement against micro-dramas, the vertical short series that run a minute or two per episode and are consumed almost entirely on phones. The campaign runs two months, from November 1 through the end of December. It follows earlier rounds against obscene content and piracy, and the announcement is unusually specific about what prompted it: uneven quality in AI-made micro-dramas, and AI tooling helping low-grade content come back after being removed.

The regulator also published its running tally: 75,000 violating micro-dramas taken down this year, more than 1,500 accounts actioned.

Most English coverage of Chinese content campaigns stops at the taste clauses. Those are there. The list covers content involving children, soft-pornographic framing, extreme interpersonal antagonism, money worship, and plots that violate family ethics. Unfiled content in special subject categories is also in scope. If that were the whole announcement, it would be a story about content standards and nothing more.

The part worth reading twice sits in the middle of a sentence: the campaign will address "face theft" and "voice theft" and standardize AI micro-drama production.

Why faces and voices are the operative clause

A micro-drama is cheap by design. A hundred episodes might cost less than one episode of a television drama, and the entire business model rests on producing volume fast enough to find a hit. AI production collapsed the cost further, which is why the category filled with synthetic performers.

Two shortcuts became common enough to attract a regulator. The first is taking a real person's face, typically a minor celebrity or a model pulled from social media, and using it as the protagonist without permission. The second is cloning a voice, often an actor's or a well-known dubbing artist's, and using it for dialogue.

Neither shortcut is unique to China. Both are trivial with current tools and both are hard to detect once the output is compressed and posted. What is unusual is a regulator naming them as a named target inside an industry campaign, because it converts a civil dispute into a compliance question. A performer who finds their face in a series currently has to sue. After this campaign, a platform carrying that series has a reason to take it down before anyone sues.

For anyone producing synthetic video for the Chinese market, the practical consequence is that likeness and voice clearance moves from "nice to have" to "part of the file". That is a documentation problem more than a technology problem, and it is the same problem the European Union is approaching from a different direction with its transparency rules.

The second clause is about volume farming

The distribution side of the campaign targets a set of behaviors that will be familiar to anyone who has watched a short-video feed for a week. Low-grade material in titles, promotional assets, comment overlays and traffic-driving clips. The same series re-released under a new name or with the protagonist renamed. First-screen recommendations and rankings that converge on the same handful of plots.

The re-release clause is the one with teeth. In a market where a series that gets taken down can return under a new title within days, takedowns have limited effect. Naming the rename as a violation closes a loop that platform enforcement had left open.

The homogenization clause is more interesting as a signal. Asking platforms to improve algorithmic diversity in recommendations is a soft instruction, but it points at a real economic problem: when production costs approach zero, the feed fills with the same story, and the audience leaves. The regulator is describing a market failure, not a moral one.

What this does to the AI micro-drama business

Three shifts, in order of how fast they will show up.

The first is a short-term chilling effect on synthetic casting. Teams that were using a lookalike face because it converted will either stop or move to clearly fictional, clearly licensed synthetic performers. The second group is where the market will settle, because a licensed synthetic identity is an asset that can be reused across a hundred episodes and cannot be stolen by a competitor.

The second is a rise in the value of the paper trail. Platform reviewers in China already ask for filing numbers on micro-dramas in special categories. Expect likeness releases and voice licenses to become part of the same submission, at least for content that features realistic human faces. Studios that already keep talent paperwork will barely notice. Studios built entirely on scraped references will have to rebuild.

The third is consolidation. Compliance costs are fixed costs, and fixed costs favor volume players. A two-person operation generating fifty episodes a month with unlicensed faces cannot absorb the same process as a studio that already has a legal reviewer. This is the standard outcome of any content regulation, and it is worth being honest that it reduces the number of independent producers.

The comparison worth making

The EU's Article 50 approach asks providers to mark synthetic output in a machine-readable way and asks deployers to disclose deepfakes. It governs the label on the file.

China's approach here governs the provenance of the person. One says: tell people this is synthetic. The other says: prove you had the right to use that face.

Both are necessary and neither is sufficient. A machine-readable mark does nothing for a performer whose likeness was licensed to one production and reused in another. A likeness release does nothing for an audience that cannot tell whether the person on screen agreed to be there. Platforms operating in both markets will end up implementing both regimes, which means the compliance surface for a single synthetic video is now a metadata layer plus a rights layer plus a filing number.

What to do if you ship AI video into China

Audit your reference library first, not your outputs. Every face and voice asset used in production needs an identifiable source and a scope: which projects, which territories, how long, whether derivative works are permitted.

Then check whether your synthetic performers are recognizable. A model trained on one consenting actor and producing a face that resembles a different, famous person is a liability even with clean training data. Resemblance is what gets noticed.

Then decide how you will document a takedown. If a platform pulls an episode because a face looked like someone's, you need a response within days, not weeks, and the response is a document.

None of this is exciting work. It is the work that determines whether the cheap production model survives contact with a regulator, and in this case the regulator moved first.

Related articles

Google replaced the model behind Search, and the number that moved was a shell benchmark

Gemini 4 Hydrogen became the default on 12 November, taking over from Gemini 3.8 Flash across Search, the Gemini app and the developer API. It runs about 2.5 times faster, scores 57% on Terminal-Bench 4.1 where its predecessor scored 18.2%, and costs $1.50 in and $7.50 out per million tokens.

Three ways a coding agent wastes money, measured on 30 repositories

TokenEconomizer watches what an agent did in a repository you work in often, then stops it from rediscovering the same build command, loading tools it never uses and looping on one error. Across five agent and model combinations it cut mean cost by 18.4% to 30.9%, cheaper in all 30 repository cells, with Pass@1 up in four of five.

earshot gives coding agents 922 sounds, and every one is rendered from code

An agent can write a game, an interface or a film and never hear any of it. earshot answers requests like "three variations of a footstep on gravel" from a catalogue of 922 designs built from public-domain recordings, with no audio files in the repository at all. It runs on CPU, as a CLI, a Python package, an MCP server and a Claude Code plugin.

The agents stealing API keys were not sophisticated. The gateways were open.

Amazon's decoy servers spent October posing as inference APIs and logged agents built on ARTEX pulling provider keys out of LiteLLM, Dify, Langflow and Open WebUI. The way in was a default master key and a half-finished install. ARTEX itself ships with budgets, scope gates and evidence for every finding.