Zhuo Xian on AI's Impact on China's Social Security System
A director at China's State Council think tank on how AI is eroding the three pillars of social security systems — and how it can be fixed
Most of the articles I’ve seen are about AI’s impact on employment, basically “how many jobs will be lost.” For today's episode, I want to go deeper than that and introduce an article about AI's impact on social security systems.
The article is by Zhuo Xian卓贤, Director and Senior Research Fellow at the Department of Social and Cultural Development Research, Development Research Center of the State Council. The DRC is one of China's most influential government think tanks, reporting directly to the State Council.
In most countries around the world, social insurance systems are built on the foundation of stable employment, and China is no exception. Whether it's unemployment insurance or maternity insurance, these programs were designed to protect workers against the risk of career interruptions. Three pillars keep the system financially sustainable: stable employer-employee relationships, wages that grow in step with productivity, and a favourable demographic structure. Zhuo warns that as AI is deployed at scale, companies no longer need to maintain large formal workforces and are shifting toward a more fragmented gig economy. At the same time, AI learns faster than any individual can, threatening to break the "learning by doing" ladder of human capital accumulation. As entry-level positions shrink, so do the pathways to becoming senior experts.
On the policy front, Zhuo proposes solutions at several levels. In the near term, he advocates for a differentiated “robot tax,” offering tax incentives for AI technologies that augment human capabilities, while withholding tax breaks or imposing modest levies on technologies that purely replace labour. Operationally, he suggests learning Japan’s approach of earmarking consumption tax revenue specifically for social security, so that social insurance funding is no longer entirely dependent on payroll taxes. The logic is that as labor’s share of national income continues to decline, there must be new mechanisms to channel the wealth generated by AI back into a social safety net that benefits ordinary people.
Over the longer term, Zhuo argues that since AI computing power will become a foundational infrastructure, the state should invest in and retain ownership of core computing assets to capture the economic rents that AI generates. He made an analogy to Norway's sovereign oil fund, which helps inject an "AI dividend" into the social security system, shifting the model from "taxing labor" to "sharing in AI's returns."
On the human development side, he believes education should pivot toward cultivating metacognitive abilities and interdisciplinary thinking, rather than betting on specific technical skills that will quickly become obsolete. In the short term, the government should subsidize wages or cover social insurance contributions for young people entering the workforce, reducing employers' costs of hiring junior staff and preventing AI from blocking young people's career on-ramps.
His article reflect-in some way- how China's policy advisors are thinking about AI. Zhuo directly cites language from the 15th Five-Year Plan proposal on "building an employment-friendly development model," which may signal that AI's impact on the job market and the fabric of society has reached the highest level on China's policy agenda. Beijing's consideration treats AI first and foremost as a governance problem, including whether pensions can be paid on time, whether public health insurance funds can remain solvent, and whether young people can still find an upward path into the middle class.
AI, Employment, and Social Security
Zhuo Xian — Director and Research Fellow, Department of Social and Cultural Development Research, Development Research Center of the State Council
The Relationship Between Economic Growth and Employment Is Changing
In the long agrarian era before the Industrial Revolution, low technological progress brought low-speed growth, corresponding to low population growth and low employment growth. Economic growth was virtually synonymous with the growth of agricultural employment.
The Industrial Revolution broke through the constraints of energy and power and the existing combinations of factors of production, greatly expanding the frontiers of human production. Industrialization and urbanization mutually reinforced each other through economies of scale. Industrial product prices fell as productivity rose, while wage levels rose along with productivity. Large-scale production and large-scale consumption formed a positive feedback loop, and blue-collar industrial jobs grew rapidly.
The modern corporate system expanded the scope of social division of labor and collaboration. Numerous production processes that were originally completed within a single enterprise (such as logistics, marketing, and legal consulting) spun off into specialised firms, forming a vast network of intermediate inputs and services. While improving economic efficiency, this also gave rise to a surge in knowledge-based white-collar positions within hierarchical organisational structures. After the widespread adoption of personal computers in the 1980s, information-processing positions such as accountants, secretaries, and analysts grew relatively quickly.
The marketisation of household labour was another important engine of job creation. As women entered the labour market on a large scale, work that had previously been performed unpaid within households was transformed into market-based services in national economic accounting. Jobs in lifestyle service industries such as housekeeping, food services, education, and entertainment were continuously created.
For most of the 20th century, “economic prosperity equals full employment” was a form of social cognition shaped by industrial civilisation, and it became the narrative logic and psychological foundation of many current business models and social institutions.
The several episodes of “jobless growth” experienced by advanced economies around the turn of the 21st century began to challenge this consensus. Initial research attributed “jobless growth” to post-crisis periods, seeing it primarily as a cyclical anomaly resulting from newly established firms increasing equipment investment, rather than a structural change in the relationship between employment and growth. However, subsequent research showed that the disappearance of routine cognitive and manual jobs did not occur gradually but was concentrated during economic recessions. Firms used crises as a concentrated “cleansing mechanism” to permanently eliminate mid-skill positions that could be replaced by automation. When the economy recovered, those jobs did not return. Although the service sector eventually absorbed the majority of the workforce, it did so at the cost of sacrificing wage growth and job stability.
Synthesising recent domestic and international literature on AI’s impact on employment, artificial intelligence has not caused large-scale unemployment. Many studies have even found that while the unemployment rate among workers in industries with high AI exposure is indeed rising, the unemployment rate among workers with lower exposure is rising even faster. One possible explanation is that workers with high AI exposure tend to have higher levels of education and stronger re-employment capabilities, and are therefore less affected. The few studies that do demonstrate higher unemployment among those with high AI exposure primarily use large language models to assess the risk of various occupations being replaced by AI — in other words, “AI tells us that AI is worsening unemployment” — and the statistical significance is not high.
Although the impact on overall employment levels is not obvious, in the current era of artificial intelligence, the relationship between employment and growth has already shown some new trends, which can be summarised as “decoupling” in three areas.
First, employment is decoupling from investment. In the industrial and service economy eras, both infrastructure investment and machinery investment generated considerable direct and indirect employment. In the AI era, technology companies are deepening their capital at unprecedented speed, yet the employment-creation effect is declining. Unlike the previous wave of internet investment, the expansion model of the AI era has shifted from “asset-light, people-heavy” to “capital-heavy, computing-heavy,” relying on high-density investment in physical infrastructure such as data centers and energy networks. The combined capital expenditures of Microsoft, Amazon, Google, and Meta in 2025 are projected to reach $400 billion — a figure exceeding the annual GDP of many medium-sized countries. Yet at the same time, tech companies are implementing human capital austerity strategies, cutting hundreds of thousands of jobs and freezing entry-level hiring for graduates. What is unusual is that these actions are occurring against a backdrop of record-high stock prices and robust revenue growth, reflecting a decision-making logic of cutting labor costs to free up funds for investment in computing infrastructure.
Second, technological progress is decoupling from human capital development. In the past, improvements in labor productivity came both from capital and the technology embodied in machinery, and from the contribution of human capital accumulated through “learning by doing.” In the AI era, labor productivity gains are more likely to come from a decline in the denominator of that metric — i.e., the size of the labor force — and the pace of human capital development falls far behind the speed of AI technological progress.
On the one hand, the path of “learning by doing” for human capital accumulation is narrowing. Previously, college graduates accumulated experience through foundational work and gradually developed into senior talent. Now, AI is increasingly competent at tasks performed by junior analysts, junior programmers, and junior copywriters, and hiring demand for fresh graduates in some positions is declining. For example, the traditional law firm model relied on large numbers of junior lawyers to perform document review, legal research, and similar work. AI can now complete these tasks in seconds, but demand for cases such as divorce proceedings does not increase because of AI — leading law firms to sharply reduce hiring of junior lawyers. This not only contributes to rising youth unemployment but may also sever the long-standing ladder for many types of human capital development. If companies no longer hire junior employees, where will future senior experts come from?
On the other hand, in the race between technology and education, the linear pace of human capital accumulation cannot keep up with the exponential speed of technological evolution. A major prescription for the employment challenges of the AI era is lifelong education. But the transformation of educational models is not a panacea in the face of AI technological advances. For the majority of workers, the pace of human capital accumulation can no longer keep up with the evolution of machine intelligence. For example, by the time a university has just launched a course on “prompt engineering,” the latest models may no longer require prompt optimization.
Third, workers’ wages are decoupling from productivity gains. Research on the U.S. labor market shows that the decoupling of labor productivity and real wages has been ongoing since the 1970s, and the accelerated adoption of AI may widen this gap. In the AI era, AI is routinizing non-routine cognitive tasks such as basic code writing, legal document drafting, and foundational financial analysis. The surplus profits of high-efficiency sectors are increasingly converted into capital gains and salary growth for a small number of core talents. Workers remaining in auxiliary roles within high-efficiency sectors are not only declining in number — because their human capital contribution is less than that of AI — but their wage growth will also not keep pace with the sector’s productivity gains.
The traditional “Baumol-style” productivity-sharing mechanism is breaking down. The “cost disease” theory proposed by Baumol noted that the surplus value created by high-productivity sectors such as manufacturing would spill over — through labor market competition (bidding for scarce labor) or institutional arrangements (union bargaining, minimum wages, etc.) — into sectors with slow productivity growth such as healthcare, caregiving, and entertainment, thereby achieving a general rise in wages across society. This cross-sectoral wage transmission mechanism maintained relative equilibrium in the labor market and served as the primary channel through which workers in low-efficiency sectors shared in prosperity. In the AI era, since high-efficiency sectors no longer need more workers, they do not need to continuously raise wages to maintain their labor force, and therefore cannot pull up society-wide wage levels through a “wage demonstration effect.” When mid-skill workers displaced by AI (such as clerks, translators, and junior coders) flow into service sectors with slower productivity growth (such as ride-hailing, delivery, and basic caregiving), labor supply exceeds demand, and the mechanism by which wages in low-efficiency sectors rise in tandem with those in high-efficiency sectors is severed.
Declining AI costs press down the “hard ceiling” on human wage increases. For a large number of tasks based on rules, logical analysis, information synthesis, and pattern recognition, AI provides a nearly infinite supply, breaking the scarcity of human capital in these areas and pushing down the market price of the relevant skills. AI technology is inherently energy-intensive. If the marginal cost of intelligence ultimately converges to energy costs, and energy costs continue to decline with technological innovations such as controlled nuclear fusion, high-altitude wind power, and space-based solar, then the wage ceiling for humans performing existing tasks faces sustained downward pressure. For example, in a particular task, when the deployment cost of AI drops to $5 per hour, the wage of a worker who performs only that single task can never exceed $5, regardless of how much their productivity has improved.
The Social Insurance System Based on Stable Employment Faces Challenges
Based on different assumptions about the timing, speed, and scope of AI’s employment displacement and creation effects, the “crystal balls” of various institutions diverge widely in their predictions of AI’s impact on future employment. For instance, since 2020, the World Economic Forum has made three consecutive, contradictory judgments on whether AI will increase employment, with a gap of 92 million between its predictions of net job gains and net job losses over the next five years. Compared to changes in overall employment levels, this article is more concerned with the challenges that structural changes in employment in the AI era pose to social security.
The modern social insurance system is a product of the era of large-scale industrialization. Whether public pension and health insurance, or unemployment insurance, work injury insurance, or maternity insurance, their original purpose is the socialized dispersion of “risks of employment interruption for workers.” The design of social security systems is therefore strongly linked to employment contributions, and their continued operation depends on three cornerstones: the growth of employed persons driven by a demographic dividend, the standardization of labor relations formed by large-scale industrial production, and the growth of wage income driven by productivity improvements. It was the historical convergence of these three conditions in the 20th century that made social insurance systems financially viable and politically operable, establishing them as an important institution for states to manage social risk.
The first cornerstone is a favorable demographic structure, which provides the actuarial foundation for social insurance. Under the social insurance system, population growth itself is transformed into a special asset class. Intergenerational transfer payments produce an implicit “biological rate of return” that can even exceed the accumulation of monetary capital. If the sum of an economy’s population growth rate (n) and real wage growth rate (g) exceeds the real market interest rate (r), then introducing a pay-as-you-go social insurance system increases total social welfare. In the decades-long “golden age” after World War II, the baby boom made this “return without capital” a reality. Participating in social insurance was not merely a mandatory burden but an investment superior to private savings. A favourable demographic structure established a social insurance intergenerational contract with social consensus, shifting retirement risk management from dispersed households to centralised social provision.
The second cornerstone is long-term stable employment relationships. Unlike social assistance based on means testing, the modern social security system emphasises the reciprocity of rights and obligations — that is, benefit levels are strictly linked to contribution histories. The original intent of this design is to maintain a dignified life for workers after retirement. Long-term stable employment relationships give workers clear, continuous income streams, ensuring the feasibility of linking “retirement benefits” to “labour contributions.” Highly organised employment relationships not only created a stable middle class but also made workers’ income transparent, calculable, and easy to deduct. This transformed the modern corporate system into an extension of state capacity, turning enterprises into agents of the state’s payroll tax (contribution) collection, improving the administrative efficiency of social security fund collection and expanding its coverage.
The third cornerstone is the synchronous growth of workers’ wages and productivity. The synchronous growth of wages and productivity ensures the endogenous expansion of the social security contribution base. Given a fixed demographic structure and collection mechanism, the improvement of social security benefit levels and the solvency of the fund fundamentally depend on the growth rate of the contribution base. Even if population aging occurs and n declines or turns negative, as long as the real wage growth rate g maintains relatively high growth, social security benefit levels can naturally rise along with total social wealth. In the 30 years after World War II, Western countries experienced a golden period of productivity growth. High unionisation rates ensured that productivity gains were translated into wage growth, forming a virtuous cycle of broadly shared productivity gains. The compound growth from a demographic dividend overlaid with a productivity dividend meant that each generation needed to contribute only a small share of its income to support the previous generation in a life better than what they had in their youth.
The modern social insurance system is an institutional arrangement through which human society, by rational design, harnesses the risks of industrialisation. It successfully internalised three specific macro-historical conditions into the parameters of institutional operation, enhancing social cohesion and improving economic and social stability. However, since the late 20th century, population aging has shaken the actuarial logic of the first cornerstone, and the second and third cornerstones also face challenges amid the leap in artificial intelligence technology.
The impact of population aging on the first cornerstone has been extensively discussed and will not be elaborated here. However, it should be noted that the impact of aging on the social insurance system is gradual and predictable, whereas the progress of artificial intelligence is nonlinear and exponential, potentially posing faster, broader, and larger-scale challenges to the second and third cornerstones of the existing social security model.
First, artificial intelligence will change the production organisation model and enterprise forms of industrial civilisation, fragmenting existing formal employment relationships and shaking the second cornerstone.
On the one hand, AI reduces market transaction costs and drives the gig-ification of knowledge workers. If the market is an efficient mechanism for resource allocation, why do firms exist? Coase’s answer is that market transactions involve search, bargaining, contracting, and monitoring costs. When the organisational costs within a firm are lower than the transaction costs in the external market, firms emerge and expand. As AI technology is applied to labour market platforms, the transaction costs of “hiring by task” become negligible relative to “hiring by job.” The basic unit of work will gradually shift from a “job” — a bundled, long-term, loosely defined set of tasks — to a “task” — a single, clearly defined, short-term deliverable — potentially reaching what has been called the “Coase Singularity.” Under the Coase Singularity, a large number of tasks that previously belonged to the firm’s core can be outsourced, even giving rise to “one-person companies,” as workers previously employed on a long-term, stable basis become outsourced personnel. Financial reports from global freelancing platforms such as Upwork and Fiverr show that large enterprises are systematically replacing full-time employees with highly skilled freelancers. If the “enterprise” — the core node of social security contribution collection — is replaced by a “transaction network” of knowledge-based tasks, the likelihood of more office white-collar positions shifting from permanent employment to gig work increases.
On the other hand, AI reduces coordination costs within enterprises and may lead to “middle-layer collapse.” In traditional enterprises, middle managers’ core functions are information transmission, task allocation, and process monitoring. AI agents are beginning to execute complex workflows without continuous human intervention, completing these coordination tasks at extremely low cost. This may lead to the flattening of organizational structures, where senior leaders can directly oversee more business units, and middle managers responsible for coordination and information processing become dispensable. Gartner predicts that by 2026, 20% of organizations will use AI to flatten their organizational structures, and more than half of middle management positions will no longer be needed.
Both of these trends will cause the gig economy to expand from its current domains of construction, manufacturing, food delivery, and courier services into knowledge-worker-dominated producer services, resulting in a larger scale of non-long-term employment relationships. This will lead to a decline in employers’ social insurance contribution responsibilities and an increase in individual workers’ contribution obligations and risk exposure.
Furthermore, if AI’s ultra-large-scale capital deepening continues in its current manner, the tilt of national income distribution toward capital owners and a small number of highly skilled individuals will shake the third cornerstone.
AI may make it difficult for the wage income of middle-income groups to keep pace with productivity growth. The primary source of funding for the social insurance system is a large middle-income population. Unlike previous industrial revolutions that mainly replaced blue-collar manual labour, generative AI accelerates the routinization of non-routine cognition, turning mid- to high-level cognitive abilities into industrially replicable services. Its primary impact is on the white-collar class — educated workers engaged in cognitive work — a group that has stable employment, relatively high wages, and high compliance rates for contributions.
A declining share of labour compensation leads to a relative decline in the social security tax base. Data from the OECD and the International Labour Organisation both show that in the most digitised industries, the share of labour income in value added is declining at an accelerating pace. This means that the dividends from technological progress are increasingly flowing to capital owners who possess algorithms, data, and computing power. Since high-income earners face caps on contributions to public basic pension insurance, health insurance, and unemployment insurance, further income growth for this group contributes almost nothing to social security funds. If capital deepening in the AI era leads to a reduction in the labor income share — particularly the income share of middle-income groups — the social security tax base as a proportion of overall economic output will decline, and economic growth will fail to translate into commensurate growth of social security funds.
Building an Employment-Friendly Development Model in the Age of Artificial Intelligence
Technology itself is neutral, but technological innovation does not inherently orient toward human well-being. If the purpose of artificial intelligence is to enhance human potential and improve the quality of life rather than “how to replace people with machines,” all the challenges described above could be readily addressed, and the technological dividend could compensate for the disappearance of the demographic dividend. For example, the European medical technology industry association estimates that the widespread application of AI in healthcare could save European healthcare systems between €170 billion and €210 billion annually, with wearable AI devices alone potentially saving approximately €50 billion per year, directly alleviating the pressure on health insurance funds in drug procurement. As another example, an important approach to solving the pension crisis is to extend contribution years. AI technology can eliminate the physiological and cognitive barriers that prevent older people from participating in the labor market, allowing older employees to focus on high-value work requiring judgment, empathy, and complex decision-making, reducing work fatigue, and enabling older workers to opt for a “phased retirement” model — transitioning from full-time to part-time work rather than abruptly cutting off their income source.
However, at least four factors currently steer the direction of AI innovation in ways unfavourable to employment and social security. First is the capital-driven “Turing Trap.” Stanford University’s Erik Brynjolfsson proposed the concept of the “Turing Trap,” pointing out that current AI research and development is excessively focused on “thinking and acting like humans,” developing “human-like intelligence” rather than augmenting human capabilities. This is a result of capital-driven innovation responding to scarcity. Prices, as signals of scarcity, direct the course of technological change, steering innovation toward replacing factors of production that are large in scale and high in price. In advanced economies, this directs innovation toward replacing high-cost labour. Second, geoeconomics promotes a labour-saving innovation pathway. In recent years, under the influence of geoeconomics, advanced economies have pushed for industrial reshoring but face severe shortages of skilled labour. To avoid uncertainties in cross-border investment, immigration policy, and tariff policy, companies are redirecting their technological investment toward “labour-saving” directions. Third, the infinite demand of the digital world exacerbates scarcity in the physical world. AI innovation cannot directly break through the scarcity of atoms. Physical constraints on land, freshwater, lithium, cobalt, and other critical minerals persist, and the scarcity of economic growth shifts to energy, environmental capacity, and key raw materials. From an employment perspective, these are all areas with thin labour demand; accelerating their development may even create a problem of AI competing with human well-being for scarce resources. Fourth, the innovation limitations of AI4Science. A study analysing 67 million papers across six major fields — biology, chemistry, geology, materials science, medicine, and physics — found that while AI tools have improved individual scientists’ output, they have led to a convergence in research topics. Scientists tend to study data-rich areas that AI can easily process, while data-scarce or marginal fields that are difficult for AI to model are neglected. This tendency may narrow the breadth of scientific discovery and reduce the potential for breakthrough innovations that open up new areas of human demand and employment.
Technological progress is path-dependent. Once a certain technological paradigm achieves dominance, society’s engineering capabilities, infrastructure, and cognitive habits are all built around it and become self-reinforcing, “locking in” the development model onto a specific trajectory. The proposal for the 15th Five-Year Plan calls for “building an employment-friendly development model” and explicitly states the need to “improve employment impact assessment and monitoring and early warning” to address “the impact of new technological developments on employment.” This represents the unity of high-quality development and high-quality full employment, and carries great significance for guiding the development direction of AI technology.
Unlike the United States, which bets the bulk of its incremental innovation resources on the training and inference layers of AI, China’s “AI+” action plan emphasises large-scale technological application, distributing innovation resources more evenly across the training, inference, and application layers of AI. This not only shortens the investment return cycle of technological innovation but also facilitates job creation through the development of AI application scenarios across production, consumption, and distribution. Moreover, China’s labour costs are far lower than those of the United States, making the gains from AI replacing labour less substantial and leaving more room for public policy to steer AI development “toward the good.” Beyond the conventional policies already in place, this article proposes several policy directions for discussion.
On the “robot tax.” Because some countries provide tax credits or accelerated depreciation for automation equipment while levying high payroll taxes (including social security contributions) on labour, this effectively subsidises the replacement of human workers with AI technology. Although many studies have proposed a robot tax, no country has yet implemented one. The Korean government, often mistakenly cited as having introduced the “world’s first robot tax,” did not directly tax robots but rather reduced tax credits for corporate investment in automation equipment. In theory, a robot tax could internalise the social costs of AI development (such as unemployment) and slow excessive employment displacement. In practice, however, it faces definitional challenges — for instance, what constitutes a “robot,” and should an Excel spreadsheet improved by AI technology be taxed? A more feasible approach would be to implement differentiated tax rates based on the type of AI technology: granting tax credits for “labour-augmenting” technologies such as exoskeletons and augmented reality glasses that assist workers, while withholding tax incentives or imposing moderate taxes on technologies that purely substitute for labour.
On a “tax-contribution coordinated” approach to social security financing. Unlike the model in continental European countries such as Germany and France, which relies primarily on employer and employee contributions, countries like Denmark have chosen to fund social security mainly through general taxation, with a smaller share of contributions. Japan, one of the world’s most aged societies, raised its consumption tax rate from 8% to 10% in 2019, with the increased revenue explicitly earmarked for social security expenditures, including pensions, healthcare, and long-term care. Although the social security financing structures of Denmark and Japan’s reforms were not originally designed to address AI disruption, a “tax-contribution coordinated” approach to social security financing can channel the wealth dividends created by AI back into the social safety net, mitigating the shocks to the three cornerstones of social security. As for specific tax instruments, based on policy practices in several countries, value-added tax (or consumption tax), environmental taxes, and capital gains taxes are options, and some research institutions have also proposed levying an AI “excess profits tax.”
On sovereign AI infrastructure. If AI computing power, as some researchers suggest, will become the currency of the future, then controlling AI infrastructure means controlling future seigniorage. Building “sovereign AI infrastructure” is not only a national security issue but could also become a new channel for social security financing. Countries such as the United Kingdom, France, Canada, and Singapore are investing in building state-owned “national research clouds” or sovereign AI computing clusters. Through national investment in core computing infrastructure, governments can directly capture the economic rents generated by AI in the future. After the large-scale commercialisation of AI, this “AI dividend” could play a role similar to Norway’s current petroleum fund, directly injecting into the social security system, achieving a shift from “taxing labour” to “sharing in AI dividends,” and allowing the social security system to share in the capital appreciation brought by AI.
On human capital accumulation in the AI era. A study by the European think tank Bruegel found that in AI-related job postings, mentions of university degrees declined by 23%, while mentions of specific skills increased significantly. At the basic and higher education stages, as the half-life of specific professional backgrounds and skills shortens, education must pivot toward cultivating “metacognitive” abilities, critical thinking, and cross-disciplinary systems integration capabilities. On the youth employment front, as AI takes over entry-level work and the “learning by doing” pathway for human capital narrows, new incentive mechanisms for graduate internships and apprenticeships must be designed. One option is for fiscal funds to subsidise the wages or social security contributions of young people entering the workforce, encouraging enterprises to hire young workers and develop human-AI collaboration and co-growth on the job.