中国正在用更强的 AI 重整全球最大监控网络，把原本偏被动的系统升级为可即时分析行为、搜寻影像并预测潜在不安定事件的工具。FT 依据十多份采购文件与知情人士访谈发现，地方政府正推进所谓预测式警务；公安部长王小洪在 2024 年发出指令，要求整合更先进 AI。这是对约十年前建成、但受限于老旧硬体与碎片化软体的平台的一次重要更新。
数据显示，早期大规模安装曾花费约人民币 3000 亿元（Rmb300bn）；而 FT 审阅的 12 份招标中，单区预算大多落在不足人民币 100 万元到约人民币 1000 万元之间，显示当前做法更像是在既有基础上叠加 AI 功能。四川 Yaodu 镇一份文件列出人民币 90 万元（Rmb900,000），用于 175 个高清摄影机与异常行为分析系统。
升级重点包括支援电脑视觉与大型语言模型的摄影机、可在设备端直接处理资料的新晶片，以及能以文字提示搜寻影像的功能，例如输入「戴红帽的女人」即可自动找片。系统还可预警危险驾驶、人群聚集、未授权进入与自杀风险等行为。部分地方因预算压力延后付款；有些新型 AI 摄影机成本最高可达旧设备的 3 倍，促使地方改以更换中介伺服器与 AI PCs 来降低云端与重建成本。
China is using more powerful AI to overhaul the world’s biggest surveillance network, turning a largely reactive system into one that can analyse behaviour in real time, search video and predict potential instability. FT, based on more than a dozen procurement documents and interviews with people familiar with the contracts, found that local governments are pushing so-called predictive policing after public security minister Wang Xiaohong ordered a 2024 upgrade. It is a major refresh of infrastructure built about a decade ago, but constrained by ageing hardware and fragmented software.
The numbers suggest a scaled-back rebuild. Analysts estimated the first wave of surveillance hardware and installation in the mid-2010s cost about Rmb300bn. By contrast, the 12 tenders reviewed by FT ranged from just under Rmb1mn to about Rmb10mn per district, implying authorities are layering AI onto existing systems. One Yaodu town document in Sichuan budgeted Rmb900,000 for 175 high-definition cameras with an intelligent video analysis system.
The new tools include cameras with computer vision and large language models, on-device semiconductor chips, and text-based video search, such as finding footage of “a woman wearing a red hat.” The systems can flag erratic driving, crowd build-ups, unauthorised entry and suicidal behaviour. Budget pressure is slowing replacement: some local buyers say AI cameras can cost up to three times more than older models, so they are replacing intermediary servers with AI PCs instead, reducing cloud bills and keeping more processing local. (Key numbers: 3)