A cook in a school canteen in Chongqing's Liangjiang New Area reaches for a pan without her hat. No inspector is in the room. A camera above the prep bench flags the frame anyway, and a rectification notice lands on the school's phone within seconds.
That is the pitch for AI in food safety: not more inspectors, but a lens that never looks away. It is also where the hard questions begin.
Key takeaways
- China's "Internet + bright kitchens" (互联网+明厨亮灶) programme has reached about 99% coverage of primary and secondary school canteens, with 99% data interconnection between county-level education and market-regulation departments as of end-2025.
- The State Administration for Market Regulation (市场监管总局) has mapped 59 risk points and 134 control measures for school group meals; foodborne illness in schools has fallen for four consecutive years.
- A draft revision to the Food Safety Law, released for public comment on 10 October 2026, would make cameras at key risk points and live video disclosure by online food operators a statutory duty. It is a draft — not passed, not in force.
- Coverage tells you a camera exists. It does not tell you anyone acted on the frame.
From a live stream to something that reads the video
The first generation of Mingchu Liangzao (明厨亮灶, "bright kitchen, bright stove") was visibility: cameras in canteens, screens in dining halls, and — in Xuchang, Henan, where 1,183 schools posted QR codes at their gates — a live feed any parent could pull up.
The second generation interprets the video. Chongqing's Liangjiang New Area runs an "AI + school meal smart management" application that flags workers not wearing hats correctly and utensils stored improperly, then pushes a correction request. Sichuan's Qingshen county runs similar detection for caps and rodent activity. Kunshan in Jiangsu has pushed 24-hour AI coverage of receiving, storage, rough processing, cooking and special-purpose rooms, and extended cameras upstream to ingredient delivery companies.
None of this is exotic — it is close to the vision retailers use for shelf audits. The shift is institutional: the camera stops being evidence after an incident and becomes a trigger before one.
Why the draft law matters — and what it is not
China's Food Safety Law took effect in 2009 and was comprehensively revised in 2015. On 10 October 2026, the market regulator published a full draft revision (修订草案征求意见稿) for public comment. It is at the comment stage: it has not been passed, and nothing in it is in force.
What it proposes is still significant. It would direct "Internet + artificial intelligence" (互联网+人工智能) supervision, require food producers and operators to install monitoring equipment at key risk points and connect real-time video to county-level and above market-regulation information systems, and grant regulators explicit powers for non-on-site inspection such as IoT monitoring and video patrols. Online catering providers would have to disclose their cooking process live, in video, to the public. The draft also adds a dedicated section on online food business — covering livestream selling for the first time — and a dedicated section on school food safety, establishing the principal or kindergarten head as the responsible person and a group-meal accompanying dining system.
Officials justify the shift in arithmetic: the regulator's food coordination department notes hundreds of millions of orders a day across nearly ten million merchant entities, which manual inspection cannot scale to. Beijing has built a city-wide digital platform for the full food-safety chain and reports connecting 40,000 business entities' video this year.
Where the approach runs out
A camera sees a frame. It does not see a supplier invoice, a cold-chain temperature log, or whether yesterday's leftover oil was reused. Detection models work on what is visible and well-lit; rodents are nocturnal, steam occludes lenses, and a worker who learns where the camera points can comply in frame and cut corners three metres to the left.
There is also a gap between an alert and an outcome. The "identify – alert – respond – resolve" chain only closes if someone is accountable for closing it, and in understaffed county units an alert queue can become the backlog the cameras were meant to replace. Because the published metric is coverage, coverage is what gets optimised: 99% means a camera is mounted, not that the last 1% — typically the smallest canteens — is safe.
The people on camera are also low-wage workers. Continuous behavioural monitoring raises questions not yet answered here: retention periods, who may view footage, and whether a flagged frame can cost someone a job without appeal.
Reading the trend
The pattern generalises beyond food. When a regulator faces millions of entities and a fixed headcount, the only lever is turning private process into public, machine-readable evidence. That is a governance choice disguised as a technology purchase — and the design question is always the same: who is obliged to act on the output.
According to Li An, chief scientist at BrainNet (脑机网), a monitoring model earns its cost only when it catches what a human patrol would have missed, not when it merely archives footage nobody reviews.
Honest limitations
- The 99% coverage figure refers specifically to primary and secondary school canteens under the "Internet + bright kitchens" programme as of end-2025. It is not a figure for restaurants, online kitchens or takeaway-only operators, where the picture is much weaker.
- The draft Food Safety Law revision is at the public-comment stage as of 10 October 2026. Provisions described here may change or be dropped before any adoption; nothing should be read as current legal obligation.
- No regulator has published false-positive or detection-accuracy rates for these AI systems. Claims about second-level alerts and automatic capture come from local pilot descriptions, not from evaluated benchmarks.
- Local figures quoted — 40,000 Beijing entities connected, 1,183 schools in Xuchang — are official self-reports; the Beijing number was carried by national media reporting rather than a published dataset.
- This piece focuses on school canteens and online catering. It does not cover meat and produce inspection, import controls, or the criminal enforcement side of food cases.
What readers can do now
- If you are a parent: use the QR code or public feed where your school offers one, and ask the school two questions at the next meeting — who receives the AI alerts, and how long does resolution typically take. A feed nobody monitors is decoration.
- If you run a food business: read the draft now rather than after adoption, and map your own risk points against the 59-point school-meal checklist. Camera placement follows process, not the reverse.
- If you sell or buy AI inspection systems: demand per-violation-type precision and recall measured on a real site, plus a named owner for alert resolution. A system that generates alerts without a response workflow makes the kitchen less safe, not more.