Chinese Economists: Make AI Pay for the Jobs It Replaces
PKU's Zhang Dandan & SMU's Li Jia warn that AI is closing off entry-level jobs for China's young workers, calling for an AI Employment Compensation Fund.
A few weeks ago, I shared Zhuo Xian’s piece on how AI is eroding the three pillars of China’s social security system. In recent days, I have seen an increasing number of Chinese economists, such as Luo Zhiheng of Yuekai Securities and Cai Fang from CASS, discussing the impact of AI on the job market. But today, I want to follow up with an article focusing on public policy mechanisms.
The article is co-authored by Li Jia 李嘉, Dean of the School of Economics and Lee Kong Chian Professor of Economics at Singapore Management University, and Zhang Dandan 张丹丹, Vice Dean and Professor of Economics at the National School of Development, Peking University. Both are well-regarded labor economists, and Zhang in particular has done extensive empirical work on China’s labor market using large-scale recruitment data
.
Their research, based on roughly 1.63 million job postings scraped from Zhaopin (智联招聘), constructed an AI-LLM occupational exposure index showing that white-collar jobs are disproportionately concentrated in high-AI-exposure occupations. More importantly, the higher the occupational exposure risk, the higher the entry barriers employers are setting for junior-level applicants. In other words, AI is closing off the on-ramps for young workers at the very start of their careers. This also echoes Zhuo Xian’s concern about the “learning by doing” ladder being broken, but here it is grounded in microdata from China’s actual labor market.
Li and Zhang argue that AI-driven labor substitution generates an aggregate demand externality; each firm internalizes the cost savings from automation but does not internalize the consequence that displaced workers, with reduced incomes, consume less, which eventually feeds back to the economy as weaker demand. It’s treated not just as a job-loss problem, but also as a market failure.
Li and Zhang lay out four distinct national approaches:
Singapore: “State-led coordinated transformation” — integrating industrial upgrading, Company Training Committees through NTUC, and the new SkillsFuture Jobseeker Support Scheme into a single transition framework.
EU: Procedural constraints via the AI Act, requiring prior consultation with worker representatives before deploying high-risk AI systems — slowing substitution to buy time for adjustment.
US: Federal retrenchment under the Trump administration, combined with fragmented state-level regulation (California’s SB951, NYC’s algorithmic bias audits) — focused more on the fairness of AI substitution than on its pace.
China: An emerging “governance-within-development” path that emphasizes the creation effect of AI rather than just limiting the substitution effect.
At the policy proposal level, Li and Zhang sketch out a “one fund, two pillars, three supporting measures” framework. The “one fund” is a dedicated AI Employment Compensation Fund, financed through fiscal allocations, surpluses from the unemployment insurance system, employer contributions, and possible surcharges tied to data and computing power revenues. And that fund should be used for targeted interventions, such as retraining displaced workers and support groups like young and low-skilled workers.
The two pillars then divide the intervention timeline. In the short term, an AI-exposure-based employment risk monitoring and early warning system, with more flexible working hours and income support arrangements in high-risk industries. In the long run, they propose institutionalizing a lifelong learning system through portable “skills accounts” and “learning accounts” that move with workers across jobs, alongside formally incorporating AI employment governance into the Employment Promotion Law and future AI-related legislation.
Tech progress does not determine social outcomes on its own. Whether AI broadens shared prosperity or deepens social divisions depends on how the government mediates the relationship between technology and labor. This line of thinking also reflects the Chinese scholarly understanding of the relationship between the state and technological progress: on the one hand, rapid technological development is welcomed and actively promoted, viewed as a key engine for boosting productivity and national competitiveness; on the other hand, the government keep vigilance to the employment shocks and social risks that technology may bring, with a strong emphasis on publishing new policies as a safety net for technological change, preventing technological transformation from tipping over into social disintegration.
Thanks to Professor Zhang’s kind authorization, I can share her piece in English.
How to Build an “Interactive Compensation” Public Policy Mechanism in Response to AI’s Impact on Employment
By Li Jia and Zhang Dandan
The new wave of artificial intelligence (AI) technology, represented by large language models, is evolving from an auxiliary tool into a key factor of production at an unprecedented pace, rapidly reshaping the structure of the labor market. The AI-LLM Occupational Exposure Index, constructed by our research team based on approximately 1.63 million online job postings from Zhaopin, reveals that white-collar positions are more heavily concentrated in occupations with high AI exposure risk. At the same time, the higher the occupational exposure risk, the more pronounced the rise in entry barriers that employers impose on applicants for junior positions. AI is not only reshaping occupational structures, but also reinforcing labor market stratification at the very starting point of career development.
Even more noteworthy, this round of technological disruption is overlapping with trends such as population aging, rising youth employment pressure, and the expansion of flexible employment, presenting new governance challenges to the “employment-first” strategy. Against this backdrop, an unavoidable core question emerges: when the pace of AI substitution systematically outstrips the creation of new jobs and the reallocation of labor, are traditional employment policy tools still sufficient? And how are countries around the world responding to this wave of AI disruption?
Aggregate Demand Constraints in AI Substitution and the Logic of Government Intervention
Before discussing specific public policy options, it is necessary to first address why government intervention in the governance of AI-driven employment restructuring is necessary.
When deciding whether to use AI to replace workers, each firm only calculates its own returns—such as how much wage cost it saves and how much operational efficiency it gains—while ignoring the “social cost”: once replaced workers experience income declines, their consumption capacity weakens accordingly, and the resulting contraction in demand ultimately feeds back into the entire economic system in the form of fewer orders and declining sales.
In other words, the cost-saving benefits from corporate automation are typically captured exclusively by the firms themselves, while the demand decline triggered by employment contraction is collectively borne by society as a whole. This means that under a fragmented, competitive market structure, the automation process driven by rational decision-making at the individual firm level is unlikely to produce a socially optimal outcome. The more competitive and dispersed the market, the harder it is for any individual firm to “internalize” the impact of its own behavior on aggregate demand, and the stronger the “aggregate demand externality” caused by employment displacement becomes.
What makes this round of AI technological revolution unique is that its “breadth” and “speed” simultaneously and significantly exceed those of most historical technological transitions, making this aggregate demand externality logic especially important.
Historically, technological revolutions usually had relatively clear industry boundaries. Even when some jobs were displaced, workers generally still had relatively sufficient time and space to migrate to other industries or occupations. For example, industrial automation primarily affected manufacturing jobs, while the spread of computers mostly transformed office administrative workflows.
This round of AI technology represented by large language models, however, exhibits highly horizontal diffusion. A large number of cognitive jobs—including copywriting, customer service, junior programming, drafting legal documents, translation, and research report writing—are simultaneously entering the scope of AI substitutability. Technological disruption is no longer confined to a single industry, but is unfolding synchronously across industries and occupational tiers.
At the same time, the speed of technological diffusion far exceeds historical experience. It took electricity about 30 years to permeate the industrial system, personal computers about 20 years to become widespread, and the internet roughly 10 years to spread. By contrast, once a large language model goes online, it can be rapidly deployed and replicated globally via cloud computing platforms and digital infrastructure.
This means that when AI’s pace of labor substitution systematically outstrips the economic system’s pace of creating new jobs and workers’ pace of retraining and occupational transition, the problems brought by technological progress are no longer merely short-term frictional unemployment, but may evolve into persistent aggregate demand shortfalls and social welfare losses.
Precisely for this reason, the “externality” perspective provides theoretical space for public policy intervention. The use of AI is not in itself a “deviation” from corporate norms; on the contrary, it is often a rational choice under the pressure of market competition. The problem is that when all firms move in unison toward “reducing labor costs,” individual rationality can accumulate into systemic social welfare losses through a mechanism akin to the “prisoner’s dilemma.”
This kind of aggregate demand externality, which the market mechanism itself cannot easily repair, is precisely one of the areas where institutional intervention is most justified. In this sense, establishing new cooperative mechanisms among government, firms, and workers around AI transformation has implications not only for social stability, but also for correcting externalities and improving overall economic efficiency.
International Perspectives: Different Approaches in Singapore, the EU, and the United States
Facing the impact of AI on the labor market, different economies are forming sharply different response paths. The differences reflect not only the policy choices themselves, but also each country’s institutional structure, governmental capacity, and traditions of labor-management relations.
1. Singapore’s “Collaborative Governance” Approach
Entering 2026, Singapore is accelerating the formation of a governance framework that links AI industrial upgrading with employment stability. Its core logic is not simply to restrict technological substitution, but to actively promote AI diffusion and industrial upgrading while cushioning technological shocks, enhancing labor adaptability, and maintaining social stability through national coordination, skills training, and employment protection.
Overall, Singapore’s AI governance can be summarized as a “state-led collaborative transformation” model: industrial policy, labor policy, and social security policy are not isolated from one another, but are integrated into a single national transformation strategy.
First, at the national governance level, Singapore has elevated AI to the core of its national competitiveness and industrial upgrading strategy. In the February 2026 Budget, Singapore announced the establishment of a “National Artificial Intelligence Council,” chaired by Prime Minister Lawrence Wong. This is the first time Singapore has elevated AI governance to the level of cabinet coordination, attempting to integrate the dispersed regulatory and developmental functions of the Infocomm Media Development Authority, the Monetary Authority of Singapore, and the Personal Data Protection Commission, while advancing national-level AI action plans in priority areas such as advanced manufacturing, logistics, finance, and healthcare.
This arrangement reflects Singapore’s fundamental positioning of AI: AI is not merely a technological tool, but key infrastructure for raising national productivity, driving industrial upgrading, and reshaping international competitiveness. Therefore, the governance focus is primarily on “promoting adoption” and “facilitating transformation,” rather than restricting technological diffusion.
At the same time, Singapore does not treat AI solely as an industrial policy issue, but simultaneously incorporates it into the labor market governance framework.
In April 2026, the Ministry of Manpower, the National Trades Union Congress (NTUC), and the Singapore National Employers Federation (SNEF) jointly established the “Tripartite Employment Council” as a unified coordination point for the employment effects of AI transformation. This mechanism is built upon Singapore’s long-standing tradition of “tripartite consultation” among government, business, and unions, with the goal of minimizing the employment impact of technological substitution while promoting corporate digitalization and AI transformation.
An important foundation of this system is the “Company Training Committee” institution, which the NTUC has been advancing since 2019. To date, Singapore has established more than 3,800 Company Training Committees, covering more than 300,000 blue-collar and white-collar workers; among these, AI-related training programs doubled in 2025 compared with the previous year.
A key feature is that training is not “external re-education” detached from corporate needs, but is directly embedded in the firm’s internal transformation process. Firms, unions, and government jointly identify which positions will be affected by AI and which skills need upgrading, and they organize training and job-transition arrangements around actual production processes. In other words, Singapore seeks to push “corporate upgrading” and “worker transformation” forward in parallel, rather than waiting until jobs disappear before providing ex-post relief.
Therefore, the logic of Singapore’s AI governance is not “preserving jobs,” but “preserving employability.” This concept was reflected in Prime Minister Lawrence Wong’s emphasis at the May 1, 2026 Labor Day Rally that “We may not be able to save every job, but we must protect every worker.”
This approach is also reflected in institutional integration. The 2026 Budget announced that the former “Singapore Workforce Development Agency” and “SkillsFuture Singapore” would be merged into a unified “Workforce and Skills Agency,” creating—for the first time at the institutional level—a single entry point for both “job seeking” and “skills learning.” Against the backdrop of AI-accelerated occupational restructuring, employment services, vocational training, and skills upgrading can no longer be treated as separate policy modules, but must form a continuously connected “employment-training-reemployment” cycle.
At the same time, Singapore is also beginning to establish employment buffer and relief mechanisms under AI transformation.
The “SkillsFuture Jobseeker Support Scheme” was fully implemented in 2025–2026. This is the first time in Singapore’s history that a cash bridging support mechanism has been established for involuntarily unemployed workers. Eligible unemployed workers can receive up to S$6,000 in phased support over six months.
Notably, the scheme sets a threshold of “an average monthly income of no more than S$5,000 over the past year,” which roughly corresponds to the median income of full-time resident workers in Singapore. This means that the scheme mainly covers low- and middle-income workers, with its core objective being to reduce the short-term income risks faced by vulnerable groups during the process of technological transformation.
Thus, from industrial upgrading, corporate training, and worker skills transitions to employment relief, Singapore is in effect building a relatively complete AI transformation governance system: the government is responsible for overall coordination, firms are responsible for implementing transformation and training, unions are responsible for organizing workers and articulating their interests, and the fiscal system bears part of the transformation costs.
2. The EU: Slowing the Impact of Substitution Through Procedural Constraints
Unlike Singapore, which emphasizes the simultaneous promotion of technology adoption and labor transformation, the EU has chosen to incorporate AI into a high-risk governance framework through horizontal legislation and ex-ante regulation, raising the institutional cost for firms to engage in large-scale labor substitution through procedural constraints.
The core carrier of this approach is the “EU AI Act,” formally adopted in 2024. The Act will become fully mandatory in August 2026, and for the first time, it provides systematic regulation through unified legislation of the use of AI in the workplace.
The Act explicitly classifies AI applications involving labor relations—such as recruitment screening, job assignment, performance evaluation, promotion and dismissal decisions, and employee behavior monitoring—as “high-risk AI systems.” Before deploying such systems, firms must fulfill obligations such as ex-ante risk assessment, algorithmic bias testing, technical documentation retention, design of human oversight mechanisms, and ongoing compliance review.
Particularly important is the requirement that before officially using a high-risk AI system, firms must notify worker representatives and affected employees in advance.
This means that the EU has effectively established a “prior consultation mechanism” for AI deployment at the institutional level. AI is no longer merely an internal corporate technology upgrade decision, but has been incorporated into the labor relations governance framework. Even if firms possess the technical capacity and economic incentives, they must procedurally address concerns regarding worker rights, transparency, and fairness.
In a sense, the EU is not directly preventing AI substitution, but rather attempting to slow the pace of substitution by raising institutional friction and procedural costs, thereby buying time for labor market adjustment.
A similar approach is reflected in the “EU Platform Workers Directive,” which took effect at the end of 2024. The directive, for the first time at the level of EU law, stipulates that platform workers cannot be dismissed or subjected to adverse treatment solely on the basis of automated algorithmic decisions; decisions involving major labor rights must retain “human intervention” and appeal mechanisms. EU member states must complete domestic transposition by the end of 2026.
It can be seen that the EU’s institutional focus on AI and employment is not primarily on “promoting technological diffusion,” but rather on “procedural legitimacy in the use of technology.” Behind this lies the core idea of Europe’s long-standing “social market economy” tradition: technological efficiency cannot automatically take precedence over worker protection.
The limitations of the EU approach are equally evident. The EU is relatively active in “limiting the pace of substitution,” but its institutional capacity in “bearing the consequences of substitution” remains limited. The EU’s strengths are primarily reflected in ex-ante regulation, information disclosure, and procedural constraints, but its existing policy tools give insufficient consideration to the income losses, reemployment transitions, and aggregate demand shocks that AI substitution may cause.
In addition, the EU approach also faces the typical problem of “high consensus thresholds.” AI governance involves complex tradeoffs among member states’ industrial interests, labor regulations, and competitiveness in innovation, making it difficult for the EU to form stable and consistent policy consensus internally. Many seemingly ambitious institutional designs have repeatedly encountered member state disagreements and industry resistance during implementation.
A typical case is the “AI Liability Directive” proposed by the European Commission in 2022. The proposal sought to lower the burden of proof for AI victims and establish a unified liability framework for harm caused by AI systems. However, because member states and the business community could never reach agreement on the boundaries of liability, the proposal was formally withdrawn in 2025.
This reflects a deeper tension in EU AI governance: its institutional design tends to carry strong normative idealism, yet under the pressures of rapid technological evolution and global industrial competition, maintaining a balance between “regulation” and “innovation” remains a persistently difficult policy challenge.
3. The United States: Federal Retrenchment and Fragmented State-Level Governance
The United States exhibits a pattern of “light federal intervention and decentralized state-level advancement” on AI and employment issues.
After 2025, the policy priorities of the US federal government have shifted back toward “technological leadership,” “reducing regulatory burdens,” and “enhancing national competitiveness,” while issues of worker protection have clearly receded into a secondary position.
The federal government’s attitude on the issue of “AI and worker protection” has clearly contracted. In January 2025, the Trump administration formally revoked the Biden-era executive order that had designated “AI’s impact on the labor market” as a federal priority. That executive order had required relevant agencies to conduct assessments of AI-induced job losses, establish labor consultation mechanisms, and study the long-term effects of automation on employment and wage structures. Under the Trump administration’s new executive order framework, much of this content has been substantially scaled back.
The “America’s AI Action Plan” released in July of the same year proposed more than 90 federal-level AI policy actions, covering areas such as infrastructure, energy, computing power, defense, export controls, and research support. Content related to the labor market, by contrast, was only briefly mentioned as an ancillary issue, with no new fiscal budget or systematic institutional arrangements.
To some extent, this reflects the basic orientation of US AI policy: prioritizing technological innovation and industrial competitive advantage, rather than first establishing large-scale employment buffer mechanisms.
From the second half of 2025 to early 2026, multiple bipartisan bill proposals have emerged in the US Senate, including requiring large firms to disclose AI-driven layoffs, establishing a national-level AI workforce research center, and strengthening firms’ obligations to notify employees about automation deployment. However, as of May 2026, most of these proposals have not yet completed formal legislative procedures.
In the United States, what is genuinely driving concrete restrictive measures are certain state governments and local regulatory bodies.
California’s SB951, the “Worker Technological Displacement Act,” currently under deliberation, is somewhat representative. The bill would require firms conducting large-scale layoffs due to AI or automation to give workers at least 90 days’ notice in advance and submit a technological substitution statement.
New York City has moved even earlier. Since 2023, New York City has formally required firms using automated recruitment tools to conduct annual “algorithmic bias audits” and to disclose to job applicants the use of AI recruitment systems. Subsequently, Illinois, Colorado, and other jurisdictions have also advanced local legislation related to workplace AI.
Overall, however, most current US state-level regulation still focuses on “AI should not produce discriminatory outcomes”—such as algorithmic bias, fairness in recruitment, and information transparency—rather than truly addressing the deeper issue of “the scale of AI substitution itself.”
In other words, the US institutional framework is currently more concerned with “whether AI is fairly replacing labor” than with “whether AI is replacing labor too quickly.”
Behind this difference lies a long-standing feature of the US political-economic system: compared with Europe, the United States is more inclined to view technological progress as a matter of market competition and innovation efficiency, rather than primarily incorporating it into frameworks of social protection and redistribution. The federal government has always remained cautious about directly intervening in corporate automation decisions, while labor market adjustment relies more heavily on local governments, market mechanisms, and individual workers’ adaptive capacities.
Therefore, the US model may achieve higher efficiency in technological diffusion and innovation incentives, but the cost is that the employment risks and income volatility brought by AI shocks must, to a greater extent, be borne by individual workers themselves.
4. China’s Policy Direction: From “Shock Response” to “Interactive Compensation”
China’s understanding of the relationship between AI and employment is accelerating into an institutionalized stage.
In recent years, policy formulations on “AI and employment” have clearly shifted from “paying attention to impacts” to “active governance.” In May 2024, General Secretary Xi Jinping, during the 14th group study session of the Politburo, clearly stated the need to “strengthen the positive employment-creating effects of new technologies and prevent the substitution effect from being released in a concentrated way over the short term.” In the same year, the “State Council’s Opinions on Deeply Implementing the ‘AI Plus’ Initiative” proposed establishing a mechanism for assessing the employment impact of AI applications, and preventing and mitigating technological shocks. Subsequently, the Third Plenary Session of the 20th CPC Central Committee and the related arrangements in the “15th Five-Year Plan” further emphasized “promoting high-quality and full employment” and “building an employment-friendly development model,” and proposed a systematic response to the impact of new technological changes on employment structure.
Unlike some countries that mainly focus on “restricting substitution,” an important feature of China’s current policy logic is its greater emphasis on “enhancing the creation effect.”
In other words, the policy objective is not merely to slow AI’s impact on existing jobs, but more importantly to accelerate the formation of new employment demand and income sources through industrial expansion, the cultivation of new business forms, and skills upgrading—aiming to make technological progress a “blood-generating mechanism” rather than merely a “substitution mechanism.” This approach is highly consistent with China’s long-standing governance logic of advancing development and stability in tandem.
Against this backdrop, China’s policy framework around AI and employment is also gradually moving from fragmented response toward systematic construction. Its core direction is not simply to restrict technological diffusion, but to establish, while promoting AI development, a comprehensive governance system covering risk assessment, employment buffering, skills transitions, and social security adaptation, so that technological progress and labor market adjustment can be mutually coordinated.
From an international comparative perspective, China’s exploration is forming a path that places greater emphasis on “governance within development”: different from the US, which relies relatively heavily on spontaneous market adjustment; different from the EU, which focuses on procedural regulation; and different from Singapore, which relies heavily on tripartite consultation mechanisms—China seeks to maintain the pace of technological diffusion while simultaneously incorporating employment stability, skills upgrading, and social security into its national development framework.
Against this backdrop, the next step in policy design may need to shift further from “passively responding to technological shocks” to “actively building interactive compensation mechanisms.”
The core of so-called “interactive compensation” is not simply income redistribution, but the establishment, at the institutional level, of more stable adjustment mechanisms between the beneficiaries of technology and those harmed by technology, allowing the efficiency dividends created by AI to partially flow back to the worker groups bearing the costs of transformation, thereby easing the structural imbalances during technological diffusion.
Around this goal, a policy framework of “one fund, two pillars, three supporting measures” can gradually take shape.
First, “one fund” refers to exploring the establishment of a dedicated AI employment compensation fund. Its funding sources can adopt a diversified financing model, including fiscal support, surplus from unemployment insurance, corporate contributions, and supplementary revenue arrangements related to data and computing power. The focus of the fund is not on universal cash distribution, but on targeted support for job stabilization and expansion, retraining for job transitions, reemployment assistance, and support for key groups, so that part of the technological dividend can flow back to labor market adjustment.
Second, the “two pillars” correspond to short-term buffering and long-term institutional construction.
In the short term, an employment risk monitoring and early-warning system based on AI exposure should be established; more flexible working-hour and income support arrangements should be explored in high-risk industries; and more precisely targeted transition support should be provided for key groups such as young workers and low-skilled workers.
In the medium to long term, it is necessary to institutionalize lifelong learning systems and skills accumulation mechanisms. For example, arrangements such as “skills accounts” and “learning accounts” could be explored, allowing training resources to follow workers as they move; meanwhile, AI employment governance should gradually be incorporated into the framework of the “Employment Promotion Law” and future AI-related legislation, to enhance policy continuity and institutional stability.
Finally, the social security system must be made to adapt to the labor structure of the AI era. Priorities include improving protections for flexible employment and platform workers, expanding the role of unemployment insurance in supporting skills transitions and occupational shifts, and further exploring how to more stably channel the productivity dividends brought by AI back into the pension and social security systems.
Ultimately, AI’s impact on employment is not merely a technological upgrade, but rather an institutional test of “how growth is distributed.” Technological progress does not automatically determine social outcomes. What ultimately decides whether AI expands social prosperity or aggravates social divisions still depends on how institutions coordinate the relationships among technology, capital, and labor.
Authors: Li Jia is Dean of the School of Economics at Singapore Management University and Lee Kong Chian Chair Professor of Economics; Zhang Dandan is Vice Dean and Professor of Economics at the National School of Development, Peking University.
More to read:
