A market for faces, built faster than the law that governs them
A new industry has appeared in China's booming AI-microdrama sector: platforms that pay ordinary people to license their faces for AI-generated video. According to Rest of World, sites like the Chengdu-based New Claw and Shenzhen-based ActID let producers browse catalogs of faces by age, gender, and archetype — "girl-next-door," "rugged," "supermodel" — and license them for $15 to $700 depending on exclusivity and use (Rest of World, 27 Jul 2026). ByteDance's own microdrama platform, Hongguo, has faced a parallel problem from the other direction: it says it has removed more than 85,000 videos containing unauthorized AI reproductions of faces and voices since early 2026, and announced a July crackdown on look-alike, homogenized AI characters. Over 95% of microdramas released in China in Q1 2026 reportedly used AI in production. The scale is real, and so is the legal exposure it's generating.
The strongest case for tighter rules
Regulators and privacy advocates have a genuinely strong argument here, and it deserves to be stated plainly before it's rebutted. Facial geometry is not a fungible asset like a stock photo — under China's Personal Information Protection Law (PIPL), it is classified as "sensitive personal information," alongside biometric data, health records, and financial accounts, precisely because its leak or misuse can cause irreversible harm to a person's dignity, safety, or property (PIPL Art. 28, translation via DigiChina). PIPL Articles 29–30 require a handler to obtain separate, specifically-informed consent before processing such data — not a bundled terms-of-service checkbox — and to disclose the necessity and impact of that processing. Lawyer Yile Deng, quoted by Rest of World, put the practical problem bluntly: platform contracts are often "so vague that it's impossible to know who" ends up using a licensed face, or for what purpose. A person who licenses their face for a romance drama has no real way to stop it resurfacing in an ad, a scam deepfake, or a dataset used to train a future face-swap model — the exact asymmetry sensitive-data protections exist to correct.
Courts are already doing the work regulation would duplicate
What's notable is that this isn't a legal vacuum being exploited — it's an active, functioning enforcement system. The Beijing Internet Court ruled on March 24, 2026, that an AI-generated face-swap superimposing an actress's likeness onto a drama character infringed her rights, even though the production company argued the resemblance was coincidental. Judge Zhao Qi held that an AI-generated face need not be a perfect match to infringe — recognizability by the public is enough — and ordered both the production company and the distributing platform to publicly apologize and pay damages, explicitly holding the platform liable for inadequate content review (Beijing Internet Court, official case summary; China Daily). That single ruling is now doing real deterrent work industry-wide. Meanwhile the Guangzhou Internet Court has heard roughly 700 AI-related face-theft cases over the past three years — evidence of both a real harm and a judiciary willing to hear and resolve it case by case, without needing a blanket ban on face licensing itself.
Where the actual gap sits
The more defensible regulatory gap isn't consent law on paper — it's enforcement infrastructure sitting one step removed from where the risk actually concentrates. China's dedicated facial-recognition rule, the CAC/Ministry of Public Security's Measures for the Security Management of Facial Recognition Technology Applications, took effect June 1, 2025, and requires safeguards like on-device processing and non-biometric login alternatives (CSET summary). But the Measures explicitly exempt facial data used for "research and development or algorithm training" — precisely the category a face-licensing platform can plausibly claim to fall under, since the license is nominally for training or generating AI content rather than real-time recognition. That carve-out, sensible when written for legitimate model R&D, now sits awkwardly next to a commercial market openly training generative models on purchased faces.
The proportionate fix
The instinct to ban or heavily pre-clear this market would be a mistake. A licensing marketplace with clear terms is a legitimate answer to a real problem: performers and ordinary people gain a lawful, compensated channel instead of having their likeness scraped for free, and PIPL's separate-consent requirement combined with active court enforcement already gives them recourse when a platform oversteps. New Claw's own move to set a 500-yuan floor to prevent underpricing, and ActID's added licensing-management and monitoring services, show the market self-correcting toward more legible terms without regulatory compulsion. The better lever isn't banning face licensing — it's closing the training-data exemption's blind spot with mandatory, standardized scope-of-use disclosures at the point of licensing (what content categories, what duration, whether resale or model-training is included), enforced through the same PIPL separate-consent mechanism already on the books. That preserves the market Chinese creators are voluntarily building while giving PIPL's consent guarantee the teeth it needs at the one point — contract formation — where today's vagueness actually lives.