AI Job Displacement and Labor Market Restructuring: How Can Humanity Find Its Place When AI Redefines Work?

In May 2025, IBM announced it would pause hiring for approximately 7,800 back-office positions, citing that these jobs would be replaced by AI within five years.^[1] That same year, the Writers Guild of America (WGA) reached an agreement after a 148-day strike in which AI usage restrictions became a core provision -- the writers were not fighting for higher pay, but for "the right not to be replaced."^[2] These two events may seem unrelated, but they point to the same fundamental question that is reshaping the global economic order: when AI systems become increasingly capable of performing human work tasks, where will hundreds of millions of workers turn? This is not science fiction. McKinsey Global Institute (2023) estimates that generative AI could automate 60% to 70% of current work activities before 2030, affecting the equivalent of 300 million full-time jobs worldwide.^[3] The World Economic Forum's Future of Jobs Report 2025 predicts that by 2030, 92 million jobs will be eliminated globally while 170 million new positions will be created -- a net gain of 78 million.^[4] But the optimistic narrative of "net growth" conceals a brutal distributional problem: those who lose their jobs and those who gain new ones are often not the same people. Drawing on my experience conducting cross-national policy research at Cambridge University and currently leading Meta Intelligence in assisting enterprises with AI transformation, I have observed profoundly that AI's impact on employment is not a purely technical issue, but a governance challenge involving distributive justice, educational systems, and the fundamental restructuring of the social contract.

I. AI Automation's Impact on the Labor Market: Data, Models, and Debates

To understand AI's true impact on the labor market, one must first distinguish three different but frequently conflated concepts: technical feasibility -- whether AI can technically perform a given task; economic feasibility -- whether replacing human labor with AI is economically worthwhile; and adoption rate -- the actual pace at which enterprises and society deploy AI. Significant gaps exist among these three dimensions, and most sensational predictions claiming "AI will replace XX% of jobs" typically consider only technical feasibility while ignoring the constraints imposed by the latter two.

The pioneering research of Frey and Osborne. In 2013 (formally published in 2017), Oxford University's Carl Benedikt Frey and Michael Osborne published a profoundly influential paper estimating that 47% of U.S. jobs faced a high risk of automation.^[5] This figure has been widely cited in public discourse but has also provoked serious methodological criticism. The Frey-Osborne model analyzes at the "occupation" level -- determining whether an entire occupation is "automatable." However, the OECD's Arntz, Gregory, and Zierahn (2016) pointed out that automation operates at the "task" level rather than the "occupation" level -- most occupations contain a mixture of automatable and non-automatable tasks, and far fewer occupations will be entirely eliminated than the 47% figure suggests. Re-estimating at the task level, they concluded that on average only 9% of jobs in OECD countries face a high risk of automation.^[6]

The equilibrium model of Acemoglu and Restrepo. MIT's Daron Acemoglu (2024 Nobel Laureate in Economics) and Pascual Restrepo constructed a more refined theoretical framework for analyzing automation's effects on the labor market.^[7] Their model distinguishes between two types of effects: the "displacement effect" -- automation replaces human labor in certain tasks, depressing wages and employment; and the "productivity effect" and "reinstatement effect" -- automation boosts productivity, reduces costs, increases output, and simultaneously creates new tasks that require human labor. The net effect of automation on employment depends on the relative magnitude of these effects. Acemoglu and Restrepo's empirical study, using the diffusion of industrial robots in the United States from 1990 to 2007 as their sample, found that each additional robot per thousand workers reduced local employment by 0.2 percentage points and wages by 0.42% -- the displacement effect was significantly positive but partially offset.^[8]

The "cognitive revolution" of generative AI. Most of the research cited above was based on industrial automation (robotics) or early AI, but the explosion of generative AI since 2022 has brought a qualitative shift. Previous automation primarily affected "routine" tasks -- whether physical (factory assembly) or cognitive (data entry). However, large language models such as GPT-4 and Claude have demonstrated astonishing capabilities in "non-routine cognitive" tasks -- copywriting, programming, legal analysis, medical reasoning -- precisely the white-collar professional work that was previously considered resistant to automation.^[9] Eloundou et al. (2023) analyzed the task exposure of every occupation in the U.S. labor market, finding that approximately 80% of the American workforce has at least 10% of their work tasks affected by GPT models, and 19% of workers have more than 50% of their tasks affected.^[10] A more critical finding: the higher the income, the greater the proportion of work affected by generative AI -- a complete reversal of the previous understanding that "automation primarily impacts low-skill work."

II. Who Is Most Vulnerable? An Economic Analysis of Technological Displacement

In his seminal paper "Why Are There Still So Many Jobs?", David Autor proposed a profoundly influential "task-based framework."^[11] Autor classifies work tasks along two dimensions: routine vs. non-routine, and cognitive vs. manual. The past four decades of technological progress have primarily replaced "routine cognitive" (e.g., bank tellers, administrative assistants) and "routine manual" (e.g., factory operators) tasks, while increasing demand for "non-routine cognitive" (e.g., management, analysis, creative) and "non-routine manual" (e.g., caregiving, maintenance) tasks -- explaining why the labor market has exhibited a "polarization" phenomenon: middle-skill jobs declining while both high-skill and low-skill jobs increase.^[12]

But generative AI is challenging the boundaries of Autor's model. Tasks traditionally classified as "non-routine cognitive" -- such as writing reports, analyzing data, generating creative proposals -- are being executed by large language models with remarkable efficiency. Brynjolfsson and McAfee foresaw this trend as early as 2014 in The Second Machine Age: the exponential progress of digital technologies would eventually reach virtually all types of cognitive work.^[13]

Based on the latest research, we can identify several particularly vulnerable occupational groups:

First, mid-level cognitive workers. This includes junior legal assistants, basic financial analysts, general administrative staff, and standardized report writers. These jobs share key characteristics: tasks are highly structured, rely on the combination and application of existing knowledge, and quality standards are clearly evaluable -- precisely the domain where large language models excel. Goldman Sachs (2023) estimates that tasks within approximately 300 million full-time jobs globally could be automated by generative AI, with legal services and administrative support among the most affected industries.^[14]

Second, the "middle ground" of creative industries. Graphic designers, advertising copywriters, illustrators, translators, and journalists -- professions once thought to require "human creativity" -- are facing direct disruption from generative AI. Importantly, AI is not replacing all creative work, but dramatically compressing demand at the "execution level" while potentially increasing the value of the "strategic level" and "aesthetic judgment level." A senior creative director wielding AI tools may replace the output of five junior designers -- this is not the disappearance of work, but the reorganization of value chains and the deepening of inequality.

Third, customer service and sales frontlines. McKinsey estimates that customer service and sales are among the areas most directly affected by generative AI, with approximately 40 to 50 million customer service positions globally likely to be substantially reshaped within the next decade.^[15] AI chatbots can already handle most standardized customer inquiries, and advances in voice AI are placing telephone customer service under automation pressure as well.

In contrast, the following types of work retain relatively high "AI resistance" for the foreseeable future. Work requiring deep interpersonal interaction and emotional connection -- such as psychotherapists, social workers, and nursing staff. The core value of these jobs lies in trust relationships and emotional resonance between people, which current AI cannot replicate. Work requiring complex physical manipulation and environmental adaptation -- such as plumbers, landscapers, and emergency room nurses. Hans Moravec's "Moravec's paradox" still holds here: perception and motor tasks that are effortless for humans remain extraordinarily difficult for machines.^[16] Work requiring cross-domain judgment and institutional knowledge -- such as senior lawyers' litigation strategy, physicians' complex diagnoses, and corporate executives' strategic decisions. These jobs require not merely knowledge, but the capacity to make decisions based on experience, intuition, and value judgments under conditions of uncertainty.

III. Emerging Occupations and Skills Transformation: What Jobs Is AI Creating?

History has repeatedly demonstrated that technological revolutions destroy old jobs while creating new ones. The proliferation of ATMs did not eliminate bank tellers -- instead, it enabled banks to open more branches at lower cost, shifting the teller's role from "cash processing" to "customer relationship management."^[17] The automobile replaced coachmen but created millions of truck drivers, auto mechanics, and gas station attendants. The question is not whether new jobs will emerge -- they almost certainly will -- but whether the time lag, skills gap, and geographic mismatch of the transition will cause massive human suffering.

The World Economic Forum's Future of Jobs Report 2025 identifies several fast-growing emerging occupations.^[4] AI and machine learning specialists top the list of fastest-growing occupations -- but the number of such highly technical positions is limited and cannot absorb the large volume of displaced workers. More noteworthy are the types of occupations indirectly created by AI diffusion:

"Human-machine interface" occupations. As AI systems are deployed across industries, a large number of new "intermediary roles" are emerging -- prompt engineers, AI trainers, AI ethics auditors, and human-AI collaboration workflow designers. The common characteristic of these occupations is that they do not require deep machine learning expertise, but demand a profound understanding of AI's capabilities and limitations, as well as the ability to integrate AI tools with specific business contexts. Harvard Business Review calls this capability "fusion skills" -- neither purely technical nor purely business skills, but the intersection of both.^[18]

"Trust and verification" occupations. As AI-generated content grows explosively, demand for "human verification" is also increasing -- AI output quality reviewers, fact-checking specialists, and algorithmic bias detectors. In high-stakes fields such as healthcare, law, and finance, the "human-in-the-loop" requirement makes these verification roles indispensable new occupations. This directly relates to the "human oversight" principle emphasized in AI governance frameworks.

The revaluation of "uniquely human" occupations. When AI can execute an ever-expanding range of cognitive tasks, jobs that depend on uniquely human capabilities may actually increase in value. Nursing and caregiving work (where Japan already faces severe shortages), experiential education (as opposed to knowledge transmission), mental health services, and community building -- these jobs are grounded in human emotion, empathy, and social connection, not in codifiable knowledge or skills. Ironically, these are often the lowest-paid jobs in the current economic system -- the AI revolution may force us to reassess the true social value of such work.

The core challenge of skills transformation. However, the transition from "displaced jobs" to "newly created jobs" does not happen automatically. Research by Mark Muro and colleagues at the Brookings Institution indicates that the skills gap between occupations with the highest AI exposure (such as accountants and administrative assistants) and the fastest-growing emerging occupations (such as AI specialists and data scientists) is enormous -- an administrative assistant displaced by AI cannot possibly retrain as a machine learning engineer in the short term.^[19] This "skills chasm" is the thorniest problem in the AI employment transition -- it means that even if, in aggregate, new job creation exceeds old job destruction, large numbers of individual workers may still fall into prolonged structural unemployment.

IV. Policy Responses Across Nations: The UBI Debate, Retraining, and Social Safety Nets

Facing the employment impact of AI automation, policy responses across countries can be broadly categorized along three dimensions: preventive measures (slowing the pace of impact), compensatory measures (providing safety nets for those affected), and transformative measures (helping workers acquire new skills).

The Universal Basic Income (UBI) debate. The most controversial policy proposal of the AI era is arguably Universal Basic Income -- unconditionally distributing a fixed cash amount to every citizen regardless of employment status. Silicon Valley's tech elite -- from Sam Altman to Elon Musk -- have voiced support for UBI, arguing that AI automation will render the traditional "work-for-income" model unsustainable.^[20]

From an economics perspective, the UBI debate involves several core questions. First, fiscal feasibility. Taking the United States as an example, distributing $1,000 per person per month would cost approximately $3.9 trillion annually -- equivalent to over 60% of the federal government's annual expenditure. Even with productivity growth driven by AI expanding the tax base, the fiscal pressure remains enormous.^[21] Second, labor supply effects. Traditional economics fears that UBI would reduce work motivation -- if one can live without working, why work? But a large-scale randomized controlled trial published in Nature Human Behaviour (2024) showed that unconditional monthly cash transfers of $1,000 reduced beneficiaries' working hours by only 1.3 to 1.4 hours per week, far less than expected.^[22] Finland's UBI experiment (2017-2018) reached similar conclusions: beneficiaries' employment rates did not decline significantly, while well-being and life satisfaction improved markedly.^[23]

Acemoglu takes a cautious stance on UBI. He argues that the key question is not "how to distribute the wealth AI creates" but rather "how to enable more people to participate in the AI economy" -- UBI is a passive compensation mechanism, not an active empowerment strategy. He advocates that policy priorities should focus on steering the direction of AI technology development so that it becomes a "complement" to human labor rather than a "substitute."^[24]

Retraining policies in practice. Large-scale workforce retraining is the most commonly adopted response strategy among governments, but its effectiveness varies widely.

Singapore's SkillsFuture initiative is the world's most systematic case. Since its launch in 2015, the government has provided every citizen aged 25 and above with a S$500 training subsidy (increased to S$4,000 in 2025), redeemable for over 24,000 certified courses.^[25] More importantly, Singapore has linked retraining to industrial policy -- providing enhanced subsidies for priority areas such as AI, the green economy, and care services, guiding the workforce to transition toward areas of demand.

Germany's "Work 4.0" (Arbeit 4.0) policy framework emphasizes social partnership -- tripartite negotiations among government, enterprises, and trade unions on the pace and manner of automation rollout. Germany's short-time work scheme (Kurzarbeit), which demonstrated its ability to buffer labor market shocks during the pandemic, could be adapted for the AI transition period -- rather than laying off workers when deploying AI, companies would reduce working hours and use the freed time for on-the-job training, with government subsidizing part of the wage differential.^[26]

Denmark's "flexicurity" model offers yet another approach: a highly flexible labor market (where hiring and firing are easy), an extremely robust social safety net (generous unemployment insurance covering up to 90% of previous salary), and active labor market policies (intensive vocational training and job matching).^[27] The core logic of this model is to protect not "specific job positions" but "the workers themselves" -- enabling people to transition smoothly between different jobs without a single bout of unemployment triggering a life crisis.

However, retraining policies face a fundamental critique: MIT Technology Review notes that the tens of billions of dollars the United States has invested in vocational retraining programs over the past several decades have yielded results far below expectations.^[28] Many mid-career workers who complete "retraining" still end up in lower-paying jobs with worse benefits. The problem lies not in the training itself, but in the fact that the pace of structural demand change in the labor market exceeds individuals' capacity to adapt -- particularly for older workers and those with lower educational attainment.

V. Taiwan's Labor Market: Unique Challenges and Opportunities

Taiwan's labor market possesses several structural features that distinguish it from global trends when confronting the impact of AI automation -- these features constitute both challenges and opportunities.

First, the dual pressure of declining birth rates and labor shortages. Taiwan's total fertility rate (approximately 0.87 in 2024) ranks among the lowest in the world.^[29] The National Development Council projects that Taiwan's working-age population (ages 15-64) will decline from approximately 16 million in 2024 to roughly 13 million by 2040. In this context, AI automation carries a dual implication for Taiwan: on one hand, it may displace existing jobs and exacerbate unemployment among certain groups; on the other, it is a necessary tool for alleviating labor shortages -- particularly in manufacturing, services, and caregiving. In other words, Taiwan's challenge is not "whether AI will displace too many jobs" but rather "whether AI can displace the right tasks in the right fields while enabling the human workforce to shift toward higher-value work." This is directly connected to the proposition that talent is national power.

Second, the transformation bottleneck of SMEs. Over 97% of Taiwan's enterprises are small and medium-sized, employing approximately 78% of the workforce.^[30] Most of these enterprises lack the technical capacity, data infrastructure, and financial resources to adopt AI. While large technology companies are already using AI to reshape their workflows, most Taiwanese SMEs have not yet completed even basic digitalization. Without policy guidance, AI may widen the "digital divide" between large corporations and SMEs, and between the tech sector and traditional industries, further exacerbating structural imbalances in the labor market.

Third, the automation frontier in manufacturing. Taiwan's manufacturing sector accounts for approximately 33% of GDP, far exceeding most developed countries. Within this sector, semiconductors, electronic components, and precision machinery are already highly automated -- TSMC's wafer fabs operate at over 95% automation. Yet Taiwan also has a large base of traditional manufacturing (textiles, food processing, metalworking) that still relies on labor-intensive production models. AI-driven smart manufacturing will deliver a dual impact on these traditional sectors: automating part of the workforce while enhancing industrial competitiveness to maintain Taiwan's position in global value chains.

Fourth, transformation pressure in the service sector. Taiwan's service sector accounts for approximately 60% of GDP and employs over 60% of the workforce. In retail, food and beverage, finance, and insurance, AI and automation are rapidly penetrating -- self-checkout systems, AI customer service, and robo-advisors. However, a structural problem in Taiwan's service sector is "low-wage stagnation" -- wages in a large segment of service jobs have been stagnant for years. AI may further compress the number of these low-wage positions, but if the transition is managed well, it could also create higher-value service jobs by improving service quality and efficiency.

Policy recommendations. Tailored to Taiwan's specific context, I propose the following policy framework. First, establish an "AI Employment Impact Assessment Mechanism" -- requiring key industries to regularly assess AI technology's potential impact on employment, providing a basis for industrial and labor policy formulation. Second, provide SMEs with "AI Transformation Packages" -- integrated support including technical consulting, financial subsidies, and employee retraining, lowering barriers for SMEs to adopt AI. Third, reform the unemployment insurance system -- Taiwan's current unemployment benefits last a maximum of only six months (for involuntary separation), far too short for structurally unemployed workers requiring extended retraining. Drawing on the Danish model, the benefit period should be extended and tied to active retraining programs. Fourth, develop AI applications in the "silver economy" -- Taiwan's aging trend creates enormous demand for care services, and AI-assisted care (such as remote health monitoring and smart home systems) can both alleviate caregiver shortages and create new employment opportunities.

VI. Transforming the Education System: From Knowledge Transmission to Capability Development

If AI is redefining "valuable human labor," then the education system -- as the source of the labor supply -- must undergo fundamental transformation. This issue has been touched upon in discussions of the crisis in higher education, but AI's impact on education is far more profound than we imagine.

The devaluation of knowledge transmission. When AI can instantly answer virtually any factual question, generate structurally complete analytical reports, and even pass bar exams and medical licensing examinations,^[31] the traditional educational value proposition centered on "knowledge transmission" faces a fundamental challenge. If AI can accomplish in seconds -- and better -- the knowledge and skills that students spend four years memorizing and practicing, then the return on investment for that form of education will decline precipitously. This does not mean knowledge is no longer important -- it means that "possessing knowledge" alone is no longer a valuable differentiator.

Redefining capability development. In the AI era, education's core objective should shift from "knowledge transmission" to cultivating several categories of capabilities that AI finds difficult to replicate. A Foreign Affairs in-depth analysis identifies the most valuable human capabilities of the future:^[32]

Critical judgment -- not the ability to "find answers" (at which AI excels), but the ability to "judge whether answers are correct, appropriate, and ethical." In a world where AI can produce massive volumes of content, the ability to discern, evaluate, and filter is more valuable than the ability to produce. Complex problem framing -- AI excels at finding optimal solutions within well-defined problem spaces, but "how to define the problem" itself is often the most critical cognitive challenge. As Einstein reportedly said: "If I had an hour to solve a problem, I'd spend 55 minutes thinking about what the problem is." This capacity to "ask the right questions" is a current weakness of AI. Cross-domain integration -- AI models typically perform exceptionally within specific domains, but creatively combining knowledge and methodologies from different fields -- for example, integrating business thinking from executive education with cutting-edge technology -- remains a human advantage. Social-emotional competencies -- leadership, empathy, collaboration, and conflict resolution -- the importance of these competencies in human society will not diminish as AI advances; indeed, they may become even more critical as technical work is automated.

The institutional infrastructure for lifelong learning. In an era of accelerating technological change, the "one-shot education" model (completing a degree in one's twenties, then using the same skill set for 40 years) is no longer sustainable. Research from the Brookings Institution indicates that future workers may need to undergo three to five major skills transformations over their careers.^[33] This requires building institutional infrastructure that supports lifelong learning -- including flexible credit bank systems, work-integrated learning opportunities, and social security mechanisms that allow working professionals to "pause work, learn new skills, and return to the workforce." Singapore's SkillsFuture and France's Personal Training Account (Compte Personnel de Formation, CPF) are pioneering efforts in this direction.^[34]

The urgency of education reform in Taiwan. Taiwan's education system has achieved remarkable results in cultivating test-taking ability -- PISA scores have long ranked among the top globally. However, this exam-oriented education model cultivates precisely the capabilities that AI can most easily replace: memorizing large volumes of information and finding standard answers to structured problems. The 2019 Curriculum Guidelines' emphasis on "competency-based learning" and computational thinking sets the right reform direction, but at the implementation level -- teaching methods, assessment design, and the signaling effects of the college entrance system -- the depth and pace of reform are far from sufficient to meet the challenge posed by AI. What Taiwan needs is not incremental curriculum fine-tuning, but a fundamental reconsideration of "what is the purpose of education."

VII. Conclusion: The Future of Human-AI Collaboration -- Not Replacement, but Redefinition

Returning to the question posed at the beginning of this article: how can humanity find its place when AI redefines work? My answer is that we need to move beyond the binary opposition of "AI replacing humans" vs. "AI assisting humans" toward a more nuanced analytical framework.

The first cognitive shift: from "occupations" to "tasks." AI rarely eliminates an entire occupation -- its more common effect is to restructure the task composition within an occupation. Physicians will not disappear, but the task of "diagnosing diseases from medical images" will be performed faster and more accurately by AI; the physician's role will shift toward "making comprehensive treatment decisions based on AI analysis and communicating with patients." Lawyers will not disappear, but the tasks of "searching case law and drafting standardized documents" will be highly automated; the lawyer's value will concentrate on "strategic legal argumentation and complex courtroom advocacy." This "task restructuring" perspective -- derived from the theoretical frameworks of Autor (2015) and Acemoglu & Restrepo (2019) -- is closer to reality than sensational predictions that "XX% of jobs will be replaced."^[35]

The second cognitive shift: from "human capital" to "human capability." Traditional human capital theory (Becker, 1964) views education as an investment in labor productivity -- you learn specific knowledge and skills, then sell the services of those knowledge and skills on the labor market.^[36] But when AI can provide many "knowledge and skills services" at near-zero marginal cost, the traditional definition of human capital requires expansion. Amartya Sen's "capability approach" offers a more illuminating framework -- what matters is not what skills you "possess," but what you have the capability to "become" and to "do."^[37] In the AI era, judgment, creativity, empathy, and adaptability -- these are not "skills" in the traditional sense (standardized, teachable, and measurable), but deeper human "capabilities" -- and they are precisely what AI finds most difficult to replicate.

The third cognitive shift: from "work" to "contribution." If AI automation ultimately leads to a dramatic reduction in human weekly working hours -- Keynes's 1930 prediction of a "15-hour work week" may finally be realized a century later -- then we need to rethink the meaning of "work" in human life. The current social contract treats "paid employment" as the primary pathway through which people obtain income, social identity, and life purpose. But in a world where AI handles the bulk of productive labor, human "contribution" may need to be defined and valued in far broader terms -- including caring for family, community service, artistic creation, environmental stewardship, and volunteer work.^[38] This is not a distant utopian fantasy -- it is a social transformation that our generation must begin to envision and prepare for.

AI's impact on the labor market is a structural transformation already underway. It will not, as some optimists expect, "automatically" generate enough new jobs to replace those eliminated -- nor will it, as some pessimists fear, lead to massive permanent unemployment. The actual outcome will depend on how we -- as a society -- respond to this challenge: Can our education system cultivate the capabilities needed for the AI era? Can our social safety nets protect the most vulnerable during the transition? Can our business leaders deploy AI responsibly? Can our policymakers design institutional frameworks that both promote innovation and ensure equity? These questions have no standard answers. But one thing is certain: passive waiting is the worst strategy. Those nations and organizations that begin investing in human capability now -- rather than merely investing in technology -- will possess the greatest competitive advantage in the AI era.