Gordon G. Liu: Can AI reduce the price of healthcare?

China's leading health economist thinks artificial intelligence may tame medical inflation.

Gordon G. Liu is BOYA Distinguished Professor of Economics, Dean of the Institute for Global Health and Development (GHD), and Director of the China Centre for Health Economic Research at Peking University.

On 16 August 2025, he delivered a speech at an event organised by the National School of Development (NSD) and GHD; the speech was later published on the NSD’s official WeChat blog on 4 November.

刘国恩：医疗通胀、鲍莫尔成本病与人工智能

Gordon G. Liu: Medical Inflation, Baumol’s Cost Disease and Artificial Intelligence

What major changes will the intersection of healthcare and artificial intelligence bring? This is a topic of broad interest. Let us begin with medical inflation.

Medical Inflation Is a Global Challenge

Medical inflation is a major challenge faced by countries worldwide. It is closely related to the “cost disease” of labour-intensive service industries identified many years ago by economist William Baumol. Can this cost disease be eased or even solved in the age of artificial intelligence?

Statistics across countries show a common long-term pattern. As economies grow and per-capita income rises, people spend more on healthcare, and healthcare outlays take up a larger share of income. In economics, goods for which spending increases as income rises are termed “luxury goods.” Healthcare services meet the basic conditions of that definition.

Based on U.S. nominal price indices from 1948 to 2022, the CPI increased at an average annual rate of 3.5 per cent, roughly a tenfold rise over 74 years. Over the same period, the price of surgical services rose by 5.5 per cent annually, about a 33-fold increase, while hospital services surged at 8.4 per cent per year, equivalent to roughly a 120-fold increase.

From 1980 to 2022, U.S. GDP grew at an average annual rate of 5.4 per cent, compared with 7.4 per cent for healthcare services, meaning healthcare spending consistently outpaced overall economic growth. As a result, healthcare’s share of GDP has risen year by year to around 18 to 19 per cent, the highest in the world. If this trend continues, U.S. economists project that by 2050, healthcare could account for as much as 30 per cent of GDP.

In China, healthcare currently accounts for about 7 per cent of GDP. Absent fundamental changes, that share could rise to around 20 per cent by 2050, approaching today’s U.S. level.

This is the medical inflation challenge that nearly all countries face.

Baumol’s Cost Disease: A Core Economic Explanation for Medical Inflation

In the 1960s, NYU economist William Baumol published a seminal paper that divided economic activity into two sectors: a progressive sector producing tangible goods and a non-progressive sector providing intangible services.

In the progressive sector, demand is met through production. Over time, producers improve efficiency, quality, and safety. Productivity rises mainly through technological progress, allowing more units to be produced per hour and enabling wages to increase. At the same time, quality per unit of time also improves. As a result, prices do not necessarily rise in line with wages and often fall, as seen in products such as televisions, radios, and mobile phones.

By contrast, non-progressive services are not delivered through a production process. Examples include the arts, clinical care provided by doctors and nurses, and classroom teaching. Productivity gains in these areas are extremely limited, which is why they are described as “non-progressive” services. Before 2000, a doctor would typically spend 10 to 20 minutes with a patient for a minor illness, and up to an hour for a more serious condition. Today, the time required remains roughly the same. Teaching is similar: class time is not easily shortened to raise efficiency, nor is this desirable in many cases. The same is true of the arts. These intangible services are, by their very nature, difficult to make more productive through standardised, mechanised, or automated processes.

The difficulty is that demand for these services is essentially an expression of human needs. People generally do not wish to see warm, personal care replaced by cold machines. Who, then, will provide such services? Naturally, they must be provided by people. Although their productivity has not improved, their pay must at least keep pace with that of manufacturing workers to keep them in these roles; otherwise, there will be no one to provide human services. Yet their labour productivity has not risen, which means their services become more and more expensive.

From 2000 to 2020, the service price indices of several major U.S. industries rose sharply. Hospital services recorded the largest increase, with prices climbing by more than 200 per cent, followed by college tuition, which rose by over 165 per cent. Medical and healthcare services increased by more than 110 per cent, average hourly wages by over 80 per cent, and housing prices by more than 60 per cent.

By contrast, over the same period, price indices in several other major sectors fell rather than rose, including automobiles, clothing, mobile phones, computers, and televisions. Their quality and functionality continued to improve, while unit prices declined. This contrast shows that the “cost disease” described by Baumol is concentrated mainly in industries tied to labour-intensive personal services.

Turning to the U.S. labour market, data from 1998 to 2018 show that employment indices for physician services, nursing services, and hospital services strongly confirm the prominence of Baumol’s cost disease in healthcare. Over these 30 years, these three employment indices rose by 54 per cent, 34 per cent, and 31 per cent, respectively, while the average employment index across all industries increased by only 17 per cent. Moreover, during the 2008 financial crisis, the overall employment index declined, yet employment in physician services, nursing services, and hospital services continued to rise, driving ever-increasing healthcare expenditures.

The continuous increase in healthcare spending is driven primarily by two factors: first, the steady rise in prices; and second, growing demand for medical services. Together, these factors create a sharply upward trend. U.S. studies show that between 2014 and 2018, price increases were the main driver of the rapid growth in healthcare expenditure, while the increase in demand played a less major role.

Baumol’s cost disease was proposed in the 20th century. Has its relevance changed as time and technology have advanced? In 2017, the University of Chicago Booth School of Business surveyed leading U.S. economists on this question. The results showed that 21 per cent strongly agreed that cost disease continues to shape society and the economy in a significant way, 38 per cent agreed with its current relevance, 10 per cent were uncertain, and a small minority disagreed.

The State of Artificial Intelligence: Global Landscape, Investment Trends, and Adoption

Baumol’s cost disease has become a striking feature of modern economic development. Now, in the era of artificial intelligence, what will happen when cost disease meets AI? Could artificial intelligence reshape human civilisation so profoundly that it constitutes a “third revolution”?

Human history has experienced two great revolutions. The agricultural revolution, more than ten thousand years ago, domesticated plants and animals, replaced foraging with cultivation, dramatically increased productivity, and enabled settled communities, reshaping the course of civilisation. The industrial revolution, beginning over two centuries ago, harnessed machines to provide labour, freeing human hands and physical strength—benefits that continue to shape our lives today.

Artificial intelligence relies on machine-based thinking and serves to liberate human intelligence. But if thinking is increasingly delegated to machines, might human intelligence face a “deflationary” crisis due to prolonged underuse? At the turn of the 18th to 19th century, the French pioneer of evolutionary biology, Jean-Baptiste Lamarck, argued that “use and disuse” is a crucial mechanism of human evolution: functions improve with use and atrophy without it. The Industrial Revolution, for example, led to a decline in human physical strength compared with that of our ancestors. This raises the question of whether artificial intelligence may similarly place human cognitive abilities at risk of regression. Of course, this is a digression.

According to Stanford University’s Artificial Intelligence Index Report 2025, the global distribution of major AI large models in 2024 was highly concentrated. The United States led with 40 such models, far ahead of other countries, followed by China with 15 well-known ones. France had 3, while Canada, Israel, Saudi Arabia, and South Korea each had 1.

Viewed over time, model scales were very small in 2003 in the United States, China, and other countries alike. They then grew explosively and reached a peak around 2021, before the pace of expansion began to moderate somewhat in the most recent period.

Another dimension of AI’s progress is which institutions developed these leading large models over the past year, and how these models currently perform relative to human intelligence. In 2024, Google and OpenAI each released seven major models, ranking first and second. Alibaba followed in third place with six. Apple, Meta, and Nvidia each launched four, while DeepSeek, MIT, Tencent, UC Berkeley, and others contributed two apiece.

Comparing these well-known AI models with the human brain shows a steadily narrowing gap. Over time, AI has gradually approached a level attainable by the human brain. For example, in image recognition, by 2013 to 2014, AI could already match physicians’ performance in interpreting medical images.

However, multimodal AI, often referred to as artificial general intelligence, has not yet reached the level attainable by the human brain because it must switch between multiple modalities. Even so, it is advancing very rapidly.

The evolution of AI is also reflected in levels of non-government investment by country. In 2024, the United States remained far in the lead, with over 109 billion dollars in private AI investment. China followed at around 9.3 billion dollars. The United Kingdom exceeded 4.5 billion, Sweden 4.3 billion, Canada 2.8 billion, France 2.6 billion, and Germany 1.9 billion.

Why is non-government investment particularly important? Because it is driven by market forces, which often gives it greater sustainability and a stronger long-term orientation. AI-related investment is spread across many sectors. Beyond foundational research and infrastructure, healthcare stands out as the leading industry in terms of investment.

Artificial intelligence has been around for many years, yet its widespread adoption remains limited. According to a 2024 report by the U.S. Bureau of Economic Analysis, in manufacturing, information, and healthcare, the deployment rate of AI-enabled equipment is no more than 12 per cent. This naturally raises a question: given AI’s many advantages, why has its adoption remained far below expectations after so many years?

A look at human history shows that once new technologies emerge, it often takes a long time before they are widely applied in production and daily life. For example, transmission systems already existed by 1890, but it took several decades before they reached full adoption. Household lighting was invented in the late 19th century, yet achieving universal installation also required forty to fifty years. The same pattern holds for household appliances and auxiliary electric motors in factories: from initial invention to widespread use in homes and industrial production, the process typically took many decades.

The diffusion of new technologies is not linear but nonlinear. It is therefore reasonable to anticipate that, although artificial intelligence has not yet achieved widespread adoption over the past decade or so, its use could expand dramatically in the coming decades if it follows this nonlinear pattern.

AI’s Transformative Potential and Practical Advances in Healthcare

A 2025 paper in the Quarterly Journal of Economics titled “Generative AI at Work” examined the impact of generative AI on workplace settings:

First, access to a generative-AI-based conversational assistant increases worker productivity, leading to a 15 per cent rise in the number of customer issues resolved per hour. This gain reflects three components: a reduction in the time it takes an agent to handle an individual chat, an increase in the number of chats handled per hour, and an increase in the share of chats that are successfully resolved. The effects are especially pronounced for less-skilled and less-experienced workers.

Second, agents who adhere more to AI suggestions experience larger productivity gains, and adherence rates increase over time.

Third, generative AI assistance significantly improves the experience of work and customer relations. Customers are less likely to ask to speak with a supervisor. This helped bolster agents’ morale and confidence and is likely to reduce staff turnover over the long term, offering important lessons for healthcare services.

Although AI adoption remains below 20 per cent, its potential in healthcare is enormous. Human decision-making is essentially about forming judgments under uncertainty, and AI’s advantage lies in lowering the information costs of such decisions, which are especially high in healthcare. In a 1963 paper in the American Economic Review, economist Kenneth Arrow discussed several key properties of information in healthcare:

1. Uncertainty. From whether people fall ill and what diseases they develop to whether a given prescription will actually work after a consultation, uncertainty pervades the entire process.

2. Heterogeneity. The same medication or treatment that succeeds for one patient may not succeed for another.

3. Information asymmetry. Because AI can reduce information costs, there is good reason to expect a broad scope for its application in healthcare.

At the same time, healthcare data have notable advantages in terms of scale, granularity, and continuity. This further favours AI development, since artificial intelligence depends on machine learning and model training.

In recent years, health care and related fields have seen unprecedented scientific advances. In 2024, two Nobel Prizes—in Chemistry and in Physics—were awarded for work closely linked to artificial intelligence. The Nobel Chemistry laureates used AI methods, notably AlphaFold2, to tackle the formidable challenge of predicting the folding and combinations of human protein structures. Over the past half-century, researchers experimentally resolved just over 200,000 macromolecular protein structures; with the help of AlphaFold2, more than 200 million possible protein structures were mapped within only one to two years, effectively compressing what would otherwise have taken more than 20,000 years. The Nobel Physics Prize went to two computer scientists who made major contributions to artificial neural networks, the most important foundation of modern machine learning.

AI has also made major contributions to pharmaceutical R&D, including the discovery of new antibiotics. The period from the 1940s to the 1960s is often described as the golden age of antibiotic discovery, with few breakthroughs thereafter. In 2020, however, a team at MIT used AI to identify several new antibiotics, with findings reported in Cell in 2020 and Nature in 2023. These compounds effectively inhibit the growth of “superbugs” such as methicillin-resistant Staphylococcus aureus, marking a significant scientific advance.

Finally, I would like to discuss AI’s potential to transform clinical healthcare services.

In 2022, several economists at Harvard University examined how artificial intelligence might affect clinical health-care services over the subsequent five years. They concluded that, during this period, deploying existing AI technologies could reduce total health-care spending in the United States—the country with the highest health-care share of GDP—by as much as 5 to 10 per cent. They identify three main channels through which these savings could be achieved:

1. Hospital level. Clinical operations can become much more efficient, including optimising surgery and improving inpatient quality and safety, because AI can rapidly, comprehensively, and accurately detect deterioration and adverse events.

2. Physician level. AI can strengthen management of clinical capabilities, including improving intake, diagnosis, and treatment, and enhancing continuity across the diagnostic and treatment process.

3. Payer level. AI can raise the efficiency of claims processing, including review and prior authorisation, and reduce unnecessary waste. Care management can also be made more efficient to deliver more personalised health management and lower readmission rates. AI will also improve provider-network management. All of this helps lift both efficiency and quality in healthcare.

It is not easy to fully capture how rapidly evolving AI will affect healthcare, and I look forward to further discussion.