Facial information was kept by the tens of thousands inside a single park photo machine, and it took a procuratorate to get it deleted. On October 10, 2026, The Paper learned from the Shanghai Huangpu District People's Procuratorate that a public interest case over an AI interactive photo device unlawfully processing facial information had recently been closed, with administrative and civil tracks proceeding in parallel. The device stood in a park: visitors took a photo of two people, and deep synthesis stitched together a family portrait with a "future baby", to view and download through a WeChat mini program. At no point was there lawful notice, and no separate consent was obtained for facial information, which counts as sensitive personal information.
Over 10,000 photos, stored where their owners never knew
After the Huangpu procuratorate found the lead in April 2026, it first issued a procuratorial suggestion to the competent administrative authority; the company then added notice pop-ups and an AI label, so the front end looked fixed. The storage side ran deeper. In May 2026, a forensic analysis report from the Shanghai procuratorate's technical centre showed that more than 10,000 captured photos were extracted from the device's application folder, and the synthesised images were also uploaded to and stored in the cloud. The device itself offered no download function, so there was no need to retain the photos for long, and among them were misfires, test shots and even clear faces of other visitors in the background, people who never knew they had been stored and had no way to exercise a deletion right. No retention period had been set. Consent inside the mini program was flawed too: the user agreement and privacy policy were bundled into one document that did not say clearly who processed the data or how long it stayed, and it contained clauses on permanent storage, secondary use and unconditional overseas transfer, with no separate pop-up for facial processing.
The scope is China only; do not carry the conclusions elsewhere
This case sits inside mainland China's personal information protection framework: faces are sensitive personal information, and processing them requires separate notice, separate consent and data minimisation. It binds operators providing such services within mainland China: companies running AI photo kiosks, interactive devices and deep synthesis mini programs, plus the parks, scenic sites and malls that install such devices. The company involved runs many similar businesses in Shanghai and across the country, so the remedy is not taking down one machine but fixing the whole class of devices. The EU and the US phrase their facial data rules differently; this case's conclusions cannot be carried directly onto overseas operations, and the reverse holds as well.
Four things operators need to change now
- Separate consent, shown separately. Facial processing needs its own notice and its own opt-in, not something swept into a bundled agreement.
- Delete once generation is done. If there is no reason to keep the originals after the service is delivered, do not store them; where storage is genuinely needed, state a retention period and honour it, and build a specific clearing mechanism for misfires and bystanders' faces.
- Remove the overreaching clauses from the agreement. Permanent storage, secondary use and unconditional overseas transfer conflict directly with minimisation and purpose limitation, and writing them in amounts to keeping evidence against yourself.
- Labels and rights channels must actually work. Deep synthesis content needs an AI label, and deletion and account cancellation must be usable in practice. The negotiated outcome of this case was exactly those three items: deleting the unlawfully stored facial information locally and in the cloud, revising the privacy policy, and adding an account cancellation function.
Labelling duties are not unique to China; the EU likewise imposes hard requirements on marking generated content, as our earlier explainer covered in how ChatGPT text watermarks are mandatory in the EU yet off by default in the API.
The real risk is that no user complaint is needed
This case did not wait for a visitor to sue. The procuratorate found the lead in the course of its duties, filed the civil public interest case on April 30, 2026, completed the public announcement procedure on June 2, and closed the case through negotiation. In other words, the risk of processing facial information does not depend on whether users complain: the stored data itself is the evidence, and the forensic report counted it precisely. For operators, building notice, consent, deletion and retention limits into the product flow before launch costs far less than retrofitting each item later inside a public interest case.
