Китай сталкивается с финансовыми барьерами внедрения искусственного интеллекта в здравоохранение

China Is Trying to Make AI Work for Health Care A state-backed push is raising the question: Who pays? China’s efforts to embed artificial intelligence across industries, which it calls its “AI+ plan,” has generated plenty of fanfare, and much of it is justified. For China, AI is a national development project, not simply a technological innovation, with the ambition to lift everything from traditional industries to emerging sectors. But the next challenges are not technical. China has already tapped much of the low-hanging fruit, such as long-term industrial planning, advanced supply chains, and a vast engineering talent pool. Some of the most important AI+ challenges go back to less glamorous problems requiring system-level and institutional policy change. Nowhere is this clearer than in “AI+ health,” made urgent by demographic pressures and an increasing disease burden. China’s efforts to embed artificial intelligence across industries, which it calls its “AI+ plan,” has generated plenty of fanfare, and much of it is justified. For China, AI is a national development project, not simply a technological innovation, with the ambition to lift everything from traditional industries to emerging sectors. But the next challenges are not technical. China has already tapped much of the low-hanging fruit, such as long-term industrial planning, advanced supply chains, and a vast engineering talent pool. Some of the most important AI+ challenges go back to less glamorous problems requiring system-level and institutional policy change. Nowhere is this clearer than in “AI+ health,” made urgent by demographic pressures and an increasing disease burden. In China and beyond, discussion of China’s AI + health push tends to picture futuristic diagnostic models and smart hospitals. But on the ground, practitioners inside the system focus on less glamorous things: data plumbing, the systems that make health data usable; incentives for institutions to cooperate; and the question of who pays. This July, one of us (Guo) attended AI and health-policy forums in Beijing and Shanghai. Practitioners at these forums returned to the same bottlenecks: usable hospital data, institutional data-sharing, and whether AI saves the health system money. Chinese experts see AI as a potential route to more accessible, efficient care, but the consensus is that there are too few high-quality datasets and too little data-sharing at scale. Since the State Council launched its nationwide AI+ initiative in 2025, health has been identified as a priority area for AI applications, with some successes. Leading hospitals have built formidable health-data assets; health-focused large language models have advanced quickly; and health apps have reached sale. Yet the market struggles to convert this capacity into commercial success: Few AI health products are approved and reimbursed by public medical insurance, leaving them with limited prospects in the domestic market. China’s data policy itself is fluid, with a recent shift to valuing “data for AI.” Previously, the buzzword was “big data”: shuzihua, or digitalization, means turning large volumes of data into assets for the digital economy. In 2023, China created the National Data Administration to build a national data market. Now, the preferred terminology is shifting toward “digital intelligence,” with greater emphasis on intelligence. The idea is that data’s value is realized only when it enables a technology, product, or some other measurable result. Health data fits the bill. Hospital data could inform diagnosis, for example. In fact, Chinese hospitals have amassed enormous amounts of data. But there is no clear roadmap for using it. If hospitals want to leverage this data to build AI models, then hospitals and firms need to work backward to compile the datasets that they need. Currently, patient health records are scattered across departments, stored in incompatible formats, and coded in shorthand. One hospital’s “heart attack” may be another’s “acute MI” or in-house code, and the records sit in silos. There is no way to exchange or use the data, especially because China’s national interoperability standards are recommendations, not requirements. A bigger challenge is data quality: AI training needs longitudinal data, records that follow patients over time, while Chinese medical records are mostly point-in-time snapshots. The “health management” aspect remains elusive. Ant Group’s acquisition of Haodf illustrates the problem: Much of its consultation data lacks the follow-up needed to confirm or refine models. Health data contains sensitive personal information and, if not handled properly, can run afoul of China’s privacy laws. Administrators often default to inaction rather than risk liability for a leak. The National Health Commission has encouraged secure methods such as federated learning, which lets models train across institutions without pooling raw data in one place, but implementing that demands technical capacity that most hospital IT teams lack and often requires outside engineers, adding cost and privacy risks. Training a model across multiple hospitals is therefore extremely rare. Pressure on Chinese doctors to publish papers creates some incentives for cooperation. Our research found that most health data sharing occurs between a hospital department (科室) and an outside research team. Since publications weigh heavily in promotions for Chinese physician-scientists, doctors have an incentive to share data with outside collaborators. But data currently transferred through academic channels is typically covered by ethics approvals and agreements barring commercial use. And it’s not clear where demand lies or how large it is. China’s AI health companies are still testing what the market will pay for. A narrow, single-purpose tool is easier to commercialize than a medical foundation model, so companies often integrate AI into hardware or devices for a niche use. There are also concerns about the business logic of AI. It improves efficiency and saves money by cutting diagnostic, operational, and research costs, but it rarely increases revenue for hospitals or pharmaceutical firms. The situation is made worse because many public hospitals are already under financial strain, squeezed by mounting debt and recent cost-control reforms, and doctors do not earn more for working more efficiently, so efficiency gains do not automatically generate revenue. Even when hospitals look to technology and pharmaceutical companies to pay for data, pharmaceutical companies cannot finance deployment alone and want some cost-sharing scheme—for example, by asking buyers to cover research costs. Patients are unlikely to pay out of pocket for a technology such as a personalized algorithm. In practice, the decisive payer is China’s public medical insurance system, 医保. But public insurance is not an open-ended payer: Regional insurance funds are required to balance income and expenditure. A separate reimbursement pool for AI health devices is unrealistic. Advocates invoke “big-picture accounting” (suandazhang): An AI product must show that it saves the system money, ideally by replacing costs, such as missed diagnoses or unnecessary procedures, so the cost-benefit analysis works out overall. The easiest route to balance the books is to attach AI to something that the system is already set up to pay for. Imaging AI, for example, can be embedded in a vendor’s scanner or reporting system and reimbursed as part of the equipment. Many practitioners envision AI+ health as a “data flywheel”: Better data trains better models, which improve care and generate more data. But the reality can also run in reverse. Insufficient high-quality data limits fine-tuning, adapting a model to a specific medical use, and commercialization, while no clear payment channel keeps hospitals and developers from investing in data. Many of the emerging policy proposals on AI+ health aim to break this death spiral. But it’s hard. Improving interoperability means overcoming vendor lock-in, the difficulty of switching from one IT provider to another. Most hospitals built proprietary IT systems decades ago. Converting accumulated data into compatible formats requires enormous resources and manual effort. Hospitals can’t swap vendors overnight, thanks to technical incompatibility and the high cost of overhauling an IT system. A diagnostic model that scores around 95 percent in the hospital that trained it can drop to around 60 percent elsewhere, making it difficult to validate and value models across hospitals. China’s vertically organized hospital governance further compounds the problem. The National Healthcare Security Administration collects information from the top down, but hospitals lack a horizontal channel to share data or harmonize standards. Public hospitals answer to their own administrative lines and bear liability for leaks: They bear the risk while receiving little reward for interoperability. So, why share? This is the hardest part of China’s AI health push. An entrenched, contradictory web of incentives pulls the initiatives in multiple directions: How do you get multiple actors across the health system to cooperate and coordinate? Some proposals we heard call for a two-pronged strategy, with each prong addressing one part of the fragmentation problem. First, break down the silos. Interoperability standards should become a condition of hospital software certification. Second, overcome the risk-averse instinct. Leading hospitals need a clear legal safe harbor for sharing through secure architectures such as “trusted data spaces”—controlled environments for data usage. This is frankly easier said than done. Permits, common standards, and legal protections for data sharing across institutions are policy choices only health and data authorities can make. The legal and regulatory effort required is nontrivial, and the potential risk is large: Another layer of bureaucratic risk aversion on top of hospitals’ own. There is some momentum in addressing data privacy concerns. Hospitals can build synthetic datasets generated from properly licensed records. The National Data Administration’s Data Element X competition, which includes a dedicated health data track, is one existing program to draw developers toward these alternative methods. Commercial insurance could solve part of the “who pays” problem. The commercial-insurance “category C” catalogue, launched in 2025, is a good start. It still needs a governed actuarial data pool—a shared dataset on health risks and costs—so insurers can price AI products accurately and actually break even. But fragmentation and misaligned incentives run deeper than the data problem. Where does AI actually help health? Ultimately, policymakers must let the market do most of the work in determining what data hospitals want to share, how much pharmaceuticals want to pay, what channels may exist for secure sharing and training, and whether AI really improves the quality and access of care. Ran Guo is a J.D. candidate at Harvard Law School and an affiliated researcher at the Asia Society Policy Institute’s Center for China Analysis. With a background in political theory and cyber policy, he studies the history, politics, and governance of data. Lizzi C. Lee is a fellow on Chinese economy at the Asia Society Policy Institute’s Center for China Analysis.