Exclusive interview with State Council Counselor Liu Yuanli: In certain fields, AI general practitioners are on par with traditionally trained doctors.

Every reporter｜Zhang Hong Every editor｜Wei Guanhong

In 2025, a 58-year-old patient from Yichang, Hubei, was diagnosed with a hidden “high-grade mucinous adenocarcinoma” in the stomach with the assistance of artificial intelligence. As similar cases become more common, the role of artificial intelligence in healthcare is gradually becoming apparent.

This year, “smart economy” was included in the “Government Work Report” for the first time. The 14th Five-Year Plan outlines that the major project of “Artificial Intelligence+” for people’s well-being is part of the core tasks of digital China. In August 2025, the State Council issued the “Opinions on Deepening the Implementation of the ‘Artificial Intelligence+’ Action,” which requires the exploration and promotion of high-level health assistants available to everyone, orderly advancing the application of artificial intelligence in scenarios such as assisted diagnosis and treatment, health management, and medical insurance services, significantly improving the capacity and efficiency of grassroots healthcare services, and the relevant content is included in the special column of the 14th Five-Year Plan.

From March 25 to 29, the “2026 Zhongguancun Forum Annual Conference” was held in Beijing. During the conference, regarding the current penetration rate of AI (artificial intelligence) in healthcare, the bottlenecks in application, and the roles AI can play in various stages, a reporter from the Daily Economic News (hereinafter referred to as NBD) interviewed Liu Yuanli, a counselor of the State Council and a tenured professor at the School of Health Management and Policy at Peking Union Medical College.

“”

Liu Yuanli participates in a roundtable discussion at the forum. Photo source: Daily Economic News reporter Zhang Hong.

Clarifying who pays and how to pay is the core issue in the development of the AI healthcare industry

NBD: What is the current penetration rate of AI in healthcare?

Liu Yuanli: The current market still shows a diversified competitive situation. AI healthcare products are mainly divided into two categories: to C (consumer-facing) and to B (medical institution-facing), and the core issue is who pays and how they pay.

On the surface, the market acceptance of quality products depends on the willingness of consumers or medical institutions to adopt them, but the key driving factor behind this is still medical insurance. Medical insurance has a decisive impact on the decisions of patients and medical institutions, especially the latter.

China’s medical institutions have a public attribute, but their survival and development heavily rely on business income, with more than 60% of business income coming from medical insurance payments. Therefore, if AI products can be included in the medical insurance reimbursement catalog, they will have a market foundation; if not, medical institutions will assess cautiously.

Even if a product is included in the medical insurance catalog, the pricing level and compensation standards are still key variables. The application of any quality medical technology is accompanied by cost expenditures, and medical institutions must conduct rigorous cost-benefit analyses.

Currently, public hospitals in China face challenges of imbalance between revenue and expenditure, and against this backdrop, they face significant financial pressure. Therefore, artificial intelligence products must prove that they can achieve “value-based healthcare”—that is, improving the quality of care, reducing complications, and improving patient outcomes while lowering the overall operating costs of medical institutions. Although quality of care, cost control, and public welfare are often viewed as an “impossible triangle,” if it cannot be proven that technology can break through this dilemma, medical institutions will find it difficult to make adoption decisions.

Currently, some AI products’ market promotion emphasizes technological advancement and research and development investment too much. However, more attention should be paid to the sustainability of their business models. If there is a lack of timely inclusion in medical insurance and reasonable cost compensation mechanisms, the implementation of technology will face the “last mile” obstacle.

Therefore, under market economic conditions, the promotion and application of quality medical technology depend on establishing a sound payment mechanism, clarifying who pays and how they pay. This is a core issue that needs to be addressed in the current development of the AI healthcare industry.

In certain fields, AI general practitioners are comparable to formally trained doctors

NBD: There have been previous reports of doctors using artificial intelligence to assist in diagnosing and identifying lesions in patients’ imaging data. In your opinion, in which areas can AI currently play a better role in healthcare?

Liu Yuanli: There is no doubt about that. The continuous iteration and optimization of medical artificial intelligence currently rely on two main foundations: knowledge base and database. In any area involving knowledge, especially in publicly published medical knowledge, AI’s capabilities have significantly surpassed those of individual doctors. Whether it is health knowledge dissemination, publicity, or the diagnosis and treatment of common and frequently occurring diseases—after inputting symptoms and basic testing indicators, AI can quickly provide judgments. Simply put, the current capabilities of AI general practitioners in common and frequently occurring diseases are already comparable to those of formally trained general practitioners.

However, when it comes to multimodal data, such as combining clinical test results, imaging data, and comprehensive patient history to diagnose difficult and complicated diseases, AI still has obvious shortcomings.

Nonetheless, two points are worth noting. First, despite many flaws, the greatest characteristic of artificial intelligence development is its extremely rapid iteration speed. As long as there is sufficient computational power support, combined with a high-quality knowledge base and real-world data for training, technology will continue to improve, and the potential is quite considerable. Second, this potential cannot be realized passively; it requires our active participation.

Therefore, I believe that medical artificial intelligence products should to some extent be regarded as public goods—the quality technology created can quickly benefit people around the world. The quality of the product, in turn, is closely related to whether each patient, each doctor, and each expert is willing to contribute clinical experience and real-world data. The more experienced the expert, the greater the responsibility to share experiences and contribute data, which can promote the optimization and iteration of this public product. Potential, responsibility, and mission should be integrated.

Based on this, in the development of large models for healthcare artificial intelligence, a data governance alliance in the healthcare field should be established to promote data sharing and collaborative innovation. The UK’s UK Biobank has contributed over thirty new targets for humanity, and China is equally capable of establishing its own China Biobank.

In addition, efforts should be made to build a global alliance. On one hand, joint model development is essential; on the other hand, it is more important to establish a scientific and authoritative evaluation system—not only listening to the opinions of Chinese experts but also gathering professional judgments from experts in various countries to ensure the quality of the produced products.

Promote the use of data from completed original diagnostic tasks for large model training

NBD: First use is needed to generate data; to use, the value must first be proven; and to make a product valuable, data must first exist. Is there a contradiction in this?

Liu Yuanli: This is called the “trustworthy data space” in healthcare, and the National Data Bureau has relevant documents available for reference. Its core mechanism includes several aspects: first, clarifying data ownership; second, motivating all parties to share the value of the data; fundamentally, it is about data governance work. Each data holder needs to properly govern the data, ensuring that it is real, reliable, and usable while being willing to participate in sharing. For the value created by sharing, they should also be able to receive corresponding returns to stimulate the motivation for sharing.

Specifically, a large amount of data will be generated during clinical diagnosis and treatment, and this data has already completed its original diagnostic tasks. It can be further governed to improve its quality and enhance its level of structuring, thereby being directly used for large model training. Therefore, data governance is the first step, usage is the second step, and value distribution is the subsequent stage.