The office lobby that logs every visitor by face. The app that will not verify you without a selfie. In China, that reflex is now being rewritten by a single new rule.
A quiet but sweeping reset
On March 21, 2025, the Cyberspace Administration of China (国家网信办, CAC) and the Ministry of Public Security (公安部) jointly published the Measures for the Security Management of Facial Recognition Technology Application (人脸识别技术应用安全管理办法). Signed as Order No. 19 and dated March 13, the rules took effect on June 1, 2025. They are narrow in name but broad in reach: almost any organization that processes faces at scale now owes a paper trail.
One detail matters for builders. The measures explicitly exempt research and development — the act of developing or training facial-recognition algorithms is left outside the net. The law targets deployment, not the lab.
No longer the only key
The headline constraint is easy to state and hard to dodge: if another method can do the same job, facial recognition cannot be the sole way to verify a person. A gym entrance, a phone unlock, a building badge system — all must offer a non-face alternative. The state reserves the right to mandate face checks in specific cases, but the default flips from "face first" to "face only if necessary."
Consent is no longer a checkbox buried in a privacy policy. Processing someone's face requires separate, informed, voluntary agreement. For a child under 14, that consent has to come from a parent or guardian. And the data itself is supposed to stay put: except where another law allows it or the individual separately agrees, face information must be stored inside the recognition device and not shipped across the public internet.
Where the cameras now face limits
Public-space cameras are allowed only "where necessary to maintain public safety," with a clearly visible notice. The measures draw a hard line around private spaces: no facial-recognition gear in hotel rooms, public baths, changing rooms, or public restrooms. Coercing someone into a face scan under the guise of better service is explicitly forbidden.
For companies, the compliance math changes. Before deploying, a processor must run a personal-information protection impact assessment (个人信息保护影响评估) and keep records. Systems have to carry encryption, access control, auditing, and intrusion defense. These are not suggestions; they are the operating baseline.
The 100,000-person threshold
Here is the number that turns this from guidance into bureaucracy. Once an organization's stored face records reach 100,000 people, it must file with the provincial or higher-level cyberspace authority within 30 working days. That filing obligation pulls facial recognition into the same regulatory register that already governs large-scale personal-data handlers.
Penalties are not spelled out in the measures themselves — they flow from the Personal Information Protection Law (个人信息保护法). Serious breaches there can draw fines of up to ¥50 million (≈ US$7M / HK$54M) or 5% of the prior year's revenue, plus personal liability for decision-makers.
What it feels like day to day
For ordinary users, the change is mostly invisible until it isn't. You may start seeing "we also offer a non-face option" notices, or be asked to tap a code instead of staring at a camera. Campuses and gyms that once bragged about frictionless entry will have to justify it. The bigger shift is psychological: faces stop being a free, silent input and become a thing you must be asked for.
Why this matters for AI builders
Facial recognition is the most deployed biometric AI (生物识别) on the planet, and China was its largest proving ground. This rule signals a posture shift: from "collect by default" to "justify before you collect." Any product — from smart retail to campus attendance to fintech KYC — now has to design around consent, on-device storage, and an exit ramp that does not require your face.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the face-recognition rules mark a move from "collect by default" to "justify before you collect," a posture that will shape how every biometric AI product is built in the country.
Honest limitations
This article rests on the official CAC text and the Personal Information Protection Law; it is not legal advice. The measures say little about enforcement intensity — we do not yet know how aggressively provincial regulators will police the 100,000-person filing line, or how the exemption for algorithm training will be interpreted when a model is trained on scraped faces. Cross-border transfers of face data and the interaction with sector rules (banking, healthcare, education) are only partially addressed. We have not surveyed how local governments will add their own restrictions on top of the national baseline.
What readers can do now
- If you deploy face tech in China, audit where faces are stored and whether a non-face alternative exists; prepare the impact assessment and the filing workflow before you hit 100,000 records.
- If you build consumer apps, treat separate facial consent as a first-class feature, not a privacy-page footnote.
- If you study AI governance, track the first enforcement cases after June 1, 2025 — they will define the real compliance bar.