In the decade following China's accession to the World Trade Organization in 2001, the United States lost approximately 5.8 million manufacturing jobs. Economists estimate that between one and 2.4 million of those losses were attributable to what researchers call the "China Shock" — the concentrated displacement caused by import competition.
That wave of unemployment eroded workers' livelihoods and self-worth, destabilized family structures, and hollowed out local economies. It ultimately contributed to the social fragmentation and polarization that, analysts argue, created the conditions for right-wing populist movements such as Donald Trump's rise to power.
Agentic AI has been gaining traction since roughly last year, initially targeting document processing, then software development, and more recently expanding into law, finance, medicine, accounting, consulting, and auditing — fields defined by information gathering, synthesis, analysis, judgment, and decision-making. The effect has been a rapid contraction in entry-level hiring and, in some cases, large-scale layoffs.
The social disruption caused by this white-collar automation, many technologists and economists warn, is likely to prove more acute and far-reaching than the China Shock of the early 2000s.
Not Your Father's Chatbot
Unlike the large language models (LLMs) of three years ago, the LLMs powering today's agentic AI systems are capable of deep reasoning on specific problems, structured planning, and the use of external tools. This enables them to handle complex, time-intensive tasks that would previously have required significant human cognitive effort — for example: "Identify all market competitors for a given product over the past three years and compare them by features, performance, price, and user ratings"; "Retrieve all past contracts similar in nature to a submitted agreement, compare their differences, flag clauses requiring human review or negotiation, and suggest improvements"; or "Scan an entire codebase for potential security vulnerabilities and propose remediation for each."
Anthropic's Claude Code is currently regarded as a leading agentic AI system, with comparable products from OpenAI and Google following closely. Claude Code also inspired the open-source project Open Claw, which surpassed its progenitor by incorporating event-triggered task execution — a feature that significantly improves practical usability and quickly gained a devoted following in developer circles.
The Money Doesn't Lie
Although agentic AI systems are currently used primarily by professionals, their extended usage sessions have driven a sharp increase in demand for backend LLM services — specifically inference workloads, rather than training workloads. Since mid-last year, data centers operated by LLM service providers have experienced rapidly rising inference demand with no sign of deceleration.
This dynamic explains why the four major cloud providers — Amazon, Google, Meta, and Microsoft — are projected to increase their combined capital investment in AI data centers from approximately $400 billion in 2025 to at least $600 billion in 2026, a growth of at least 50 percent.
Anthropic's own revenue trajectory reinforces this picture: roughly $1 billion in 2024, approximately $9 billion in 2025, and $10 billion in just the first two months of 2026 — figures that effectively challenge the narrative that AI investment is heading toward a speculative bubble.
The Man Building the Storm Sounds the Alarm
Dario Amodei, Anthropic's chief executive officer, has expressed deep concern that his company's agentic AI products could fundamentally transform knowledge-intensive industries within a remarkably short timeframe. Over the past several months, Amodei has written multiple public essays and participated frequently in forums and interviews, urging broader awareness of the coming wave of white-collar automation and its potential harms.
LLM technology, Amodei has said, is approaching "the end of exponential growth," and within a few years humanity will effectively possess "a country of geniuses in a data center" — each genius being a task-specialized agentic AI system that, in most cases, outperforms the best human experts in its domain. In specific scientific fields, he has suggested, such AI systems could surpass Nobel laureates in research capability.
Amodei has said he is 90 percent confident this scenario will materialize within the next decade, and that at least 50 percent of entry-level white-collar positions will disappear entirely within five years. What concerns him most, he has said, is that "society has absolutely no idea how close we already are to this reality."
Worse Than China Shock — and Faster
If Amodei's projections are accurate, the unemployment wave triggered by agentic AI will extend well beyond entry-level white-collar roles and will affect most countries in a compressed timeframe — making its economic damage and political consequences substantially greater than those of the China Shock.
Few analysts have yet examined how different segments of society will respond to such a fundamental disruption in labor markets, and it remains genuinely difficult to predict what the world will look like once the transition stabilizes. What historical precedent does suggest, however, is that large-scale unemployment reliably produces social fragmentation, political polarization, and populist backlash — and that anticipating and mitigating these outcomes deserves urgent attention.
So What's Left for Humans?
Given that LLMs can now retain effectively all accumulated human knowledge and apply it to novel problems they have never previously encountered, a meaningful question arises: what roles remain for humans?
Some observers argue that physically demanding trades — auto mechanics, plumbers, electricians, construction workers — should be relatively insulated from near-term AI displacement. Even so, humanoid robotics poses a longer-term risk to those occupations as well, and the total number of such jobs is inherently limited.
Elon Musk's proposal that most people will no longer need to work and can subsist on universal basic income (UBI) may be economically feasible in principle, but may prove insufficient to sustain healthy social functioning.
Follow-on research into the China Shock has shown that displaced workers do not simply seek a replacement income — they seek the sense of purpose and social recognition that comes from contributing meaningfully to society. UBI, on this analysis, may address material deprivation without resolving the deeper social injury of involuntary idleness.
The Jobs That Get Better With Age
What meaningful roles, then, might humans retain in an LLM-dominated knowledge economy? Several categories appear relatively durable.
Human taste and aesthetic judgment can inform product specifications tailored to specific user communities. Humans can supervise, review, and refine AI outputs to ensure accuracy and responsiveness to subjective client needs. Humans can also certify AI-generated work product, thereby assuming ultimate legal accountability.
On the supply side, LLMs require continuous streams of new material to sustain their development — which means journalists who report on events, scientists who conduct and document experiments, and artists who generate genuinely novel forms of poetry, music, and visual art will continue to generate value that AI systems cannot independently produce.
Finally, roles requiring deep interpersonal engagement — sales, negotiation, litigation, lobbying, political campaigning, and live entertainment — are likely to remain human domains for the foreseeable future. Notably, nearly all of these categories share a common feature: practitioners become more valuable, not less, as they accumulate age and experience.
*The author, Dr. Tzi-cker Chiueh (闕志克), is a jointly appointed professor in the Department of Computer Science at National Tsing Hua University.
英文站用!! 授權自奔騰思潮
