The Rupture of the Blue-Collar Myth: When Physical Labor Is No Longer Safe

Opening: Dallas, a Turning Point in April 2026

On an ordinary workday morning in April 2026, in Dallas, Texas, an Aurora autonomous truck was traveling smoothly on Interstate 45. The cabin was empty—no human driver, no safety operator, only a sophisticated suite of sensors and computing systems piloting this 80-ton moving machine.

That same day, another logistics company announced the layoff of its last 200 truck drivers.

This was no longer news; it had become a kind of monthly routine. But this time was different. This time symbolized the final shattering of a myth—a myth that everyone once deeply believed: physical labor is safe, at least for the next decade.

The 350 million blue-collar workers worldwide, especially the more than 100 million blue-collar workers in the United States, once enjoyed a brief sense of historical superiority. When ChatGPT began sweeping through the white-collar world and programmers started worrying about whether AI would replace their jobs, blue-collar workers could confidently say: “Our jobs require physical presence, on-site judgment, and manual skills. Machines will have a hard time doing that.”

This conclusion used to be correct.

Now it is dead.

Part I: The Origins of the “Blue-Collar Myth”

To understand why the entire society believed blue-collar work was safe, we need to go back to the early period of 2015-2023, after the “AI winter.”

During that period, AI's major breakthroughs were concentrated in one direction: text and data processing. NLP (natural language processing) made revolutionary progress, from GPT-2 to GPT-3, and then the explosion of ChatGPT. What did this mean? It meant the jobs AI could most easily kill were those involving “sitting in an office typing.”

When the 2023-2024 layoff wave swept through the tech industry, those cut were primarily: content editors, junior analysts, customer service representatives, data labelers, junior programmers. These people all shared one characteristic: their work could be digitized into text or code, and then taken over by a neural network.

Meanwhile, people observed that manufacturing unemployment remained relatively stable. Why? Because industrial robotics progressed slowly. Yes, robots had been working on assembly lines for decades, but they mainly did repetitive, monotonous work—tightening screws, welding, painting—tasks that had already been automated in the 2000s, not something new in the 2020s.

Based on this observation, a comfortable conclusion emerged: automation has a “natural dividing line.”

What is this dividing line? Academia and the media offered a tidy explanation: automation excels at handling “structured, codable, rule-clear” work, but is powerless against “unstructured work requiring on-site judgment and adaptation to complex environments.”

What does a truck driver do? Drive in different weather conditions, make judgments when encountering sudden accidents, interact with other drivers or law enforcement, handle mechanical failures. This looked like a job “beyond the reach of automation.” What does a restaurant worker need? Quick reactions, follow the head chef's instructions, understand customers' special requests, stay calm under high pressure. This looked entirely like a human job.

This was not an absurd conclusion. It was based on the actual state of technology at the time. But it overlooked one small detail: the direction of technological progress was changing.

Part II: The Beginning of the Myth's Rupture—The 2025-2026 Evidence Chain

Transportation: No Longer Hypothetical, But Real

Aurora Trucking's progress shows a striking acceleration. In April 2026, the company announced it had completed over 250,000 miles of driverless operation in Texas, with zero safety incidents. More importantly, they had achieved observer-free operation on I-45 and other commercial routes—meaning truly unmanned trucks, no longer tests with “a safety operator sitting in the cab just in case.”

But even more noteworthy is the commercial commitment. Hirschbach Carrier signed a memorandum of understanding with Aurora to purchase 500 autonomous trucks, with delivery starting in 2027. This is not a proof of concept, not a pilot project—this is a real, binding commercial order involving a major U.S. logistics company.

Update (2026-09): Aurora's 2026 Q2 financial report (released 2026-07-29) shows Q2 revenue of only $2 million and a net loss of $270 million; management reiterated full-year revenue guidance of $14-16 million (+400% YoY, but the absolute figure is still a rounding error), and stated that year-end capacity of 200 vehicles is “fully allocated” (pre-sold to customers). The CFO gave the conversion on the earnings call: “200 vehicles ≈ $80 million annualized revenue”—this volume is still insufficient to make a dent in the $175 billion U.S. truck driver payroll pool mentioned later in this article. In the same period, Aurora released its second-generation hardware, Aurora Driver 2, mounted on the newly introduced International LT series truck platform; CEO Chris Urmson called it a “superhuman standard for safety” (FirstLight lidar with 1 km detection range). This update supports rather than refutes the original judgment: commercialization speed is accelerating, but absolute scale remains extremely small.

But the numerator-denominator comparison paints a much more restrained picture than the news headlines. Placing “200 Aurora + 28 Kodiak + 30 other vendors ≈ 260 vehicles” into the total U.S. Class 8 heavy truck fleet of 4,000,000, driverless trucks still represent only 0.0073%—the denominator is five orders of magnitude larger than the numerator. In the same period, the ATA driver shortage is approximately 60-80K. This means: autonomous trucks are filling the shortage, not replacing the existing workforce. This state may persist until a 2027-2028 inflection point.

If Aurora maintains the “200% YoY growth” guidance from its 2026 Q1 earnings, cumulative deliveries could reach 5,000-7,500 vehicles by 2028. In the same period, the ATA shortage is projected to expand to 160K. The two curves first approach each other around 2028—this is the inflection point where substitution truly begins consuming “existing drivers,” not 2026. Some language in the original § 2 equates “autonomous truck commercialization” with “structural disappearance of truck driver jobs,” which needs to be dialed back a notch.

New coordinate point: Aurora expanded its driving network to 10 routes in February, including the Fort Worth → Phoenix 1,000-mile corridor, which is the first commercial driverless freight route exceeding the federal 11-hour single-driver limit. The significance of this route: Aurora is no longer “replacing drivers” but undertaking “transport that human law does not allow”—the long-haul segment must legally be unmanned, while the last mile is handed back to human drivers. This is an industry turning point first explicitly identified in a CNBC May 6 report: McLane is owned by Berkshire Hathaway, and its largest customer is Walmart—Aurora is being locked into the logistics chain of America's largest retailer.

The United States has 3.5 million truck drivers. Their average annual salary ranges from $55,000 to $70,000. The total annual payroll for this industry is approximately $175 billion.

When Hirschbach and other logistics companies begin replacing their truck drivers at scale, what happens? An analysis from the Michigan Journal of Economics indicates that in the transportation and warehousing sector, unemployment is expected to rise 34% within 5 years due to automation.

Food Service: From Pilot to Scale

Miso Robotics' Flippy robot was once an amusing novelty. In mid-2024, it was being tested in a handful of restaurants. In February of this year, Miso Robotics acquired Zignyl—a restaurant operations software company—and launched the integrated product Zippy, an AI-driven dashboard that restaurant owners can interact with via a chat interface to monitor robot status, ROI, and maintenance data in real time.

Flippy is now deployed in seven states and has cooked 5 million baskets of food. The latest version of the robot is twice as fast as the previous generation, half the size, and profitable from day one.

What does this mean? The U.S. has approximately 2.9 million fast food and food service workers, many of whom do frying, grilling, and cooking. Flippy isn't targeting all of these jobs—it “only” handles the fry station. But once this link is broken, once a McDonald's or Chick-fil-A realizes they can replace two employees with one Flippy, other chains will follow. And Miso has signed a nationwide installation and support agreement with Roboworx—this is the infrastructure for scaling.

Construction and Manufacturing: The Specter of Humanoid Robots

Perhaps nowhere is the pace of change faster than in the rise of humanoid robots. Foundation Robotics' bricklaying robot is already working on several construction sites in the United States. Tesla Optimus and Figure AI's robot prototypes are improving every month, with expanding functional scope.

These robots are not yet mature enough to fully replace a skilled construction worker. But they are accelerating toward that point. Once they reach it—which the industry broadly expects to happen between 2027 and 2029—the risk facing construction workers will be catastrophic.

The U.S. construction industry employs 11 million workers. This is one of the largest blue-collar occupational categories in America. When this sector begins to automate, the impact will be visible.

According to Demandsage's data aggregation, by the end of 2026, AI-driven robots have already replaced approximately 2 million manufacturing workers globally. This number increases every quarter.

Placing the cumulative operating hours of humanoid robots across four factories side by side reveals a continuous spectrum from “demo video” to “double-shift production”: Agility Digit at the GXO warehouse has achieved 16 hr/day continuous double-shift operation; Figure 02 at BMW Spartanburg averages 1.7 hr/day per unit, still in the “demo + process validation” phase; Apollo and Optimus don't even have disclosable numbers. This shows that “humanoid robots in factories” is real, but there is still an order-of-magnitude gap before “replacing human workers hour by hour.”

But a prerequisite change in May 2026 is worth including: After Tesla produced the final batch of Model S/X at Fremont, it officially dismantled the main line and converted it into an Optimus mass production base. Optimus mass production is planned to start in late July-August 2026, with an initial capacity target of 1 million units per year. Musk again delayed the Optimus V3 launch in April, giving everyone a 12-month breather, but the Fremont conversion itself is an irreversible decision—the auto business is yielding factory physical space for the robot business, a stronger commitment signal than any demo video.

Update (2026-09): Optimus officially entered production at Fremont in late August 2026, slightly delayed from the “late July-August” window given in April (the July shareholder letter had revised this to “sometime this year”). But Musk simultaneously warned: the production line involves 10,000 entirely new parts, and this year's output is “literally impossible to predict”; he declined to give any 2026 production target, only promising the robots would first learn “simple skills in the factory”—a public and conspicuous downshift from his early 2025 promise of “10,000 units per year.” An independent tracking site estimates that as of mid-2026, Tesla has deployed approximately 1,000-1,200 Optimus units total across Fremont and Giga Texas, with zero external sales and no published yield or runtime data; a second dedicated Optimus factory is planned to begin producing higher-capacity Gen 4 versions at Giga Texas in summer 2027.

Another overlooked counter-example: BMW's second-phase pilot in Leipzig, Germany does not reuse Figure 02—switching to the AEON platform. Two implications: (a) BMW is placing simultaneous bets on two independent humanoid robot supply chains, hedging against single-vendor risk; (b) Figure has been validated as producible, but BMW still insists on “diversification”—no single humanoid brand has reached the “winner-take-all” stage yet.

Agriculture: The Invisible Revolution

While not strictly in the “industrial blue-collar” category, the pace of agricultural automation is equally astonishing. Drone spraying, robotic harvesting, precision agriculture AI systems—these are no longer lab prototypes but technologies already deployed at scale on farms worldwide. There are 1 billion agricultural workers globally. When 30%-40% of this sector begins to automate, we will see truly massive unemployment.

Part III: Why Blue-Collar Workers Are More Vulnerable Than White-Collar Workers

This is the most critical point in the story. Both white-collar and blue-collar workers face the impact of automation, but the nature, depth, and recovery capacity of that impact are entirely different.

1. The Geographic Specificity and Particularity of Skill Migration

What can a laid-off software engineer do? He/she can:

• Pivot to different programming languages and tech stacks
• Pivot to product management, technical writing, data analysis
• Work remotely, free of geographic constraints
• Restart a career anywhere with an internet connection

What can a truck driver displaced by autonomous trucks do? Options:

• Warehouse work (also facing automation threats)
• Construction (also undergoing automation)
• Service industry (requires retraining, lower pay)
• Unemployment (often prolonged and deep)

The key difference is geographic dependency. A truck driver's entire career is built on the local highway network. When the transportation company headquarters is in Dallas and the supply chain runs along the coast, you can't simply move to Seattle just because Dallas jobs have disappeared. Your skills are hired on Dallas roads.

This creates a massive geographic trap. Once a region's transportation industry begins to automate, the entire regional economy could collapse.

2. The Asymmetry of Retraining Resources

When a white-collar employee is laid off, he/she typically receives severance pay, possibly including career transition support. Many large companies provide retraining programs for white-collar workers. Even without them, an engineer earning $100,000 a month has savings, a credit score, and can afford the cost of learning new skills.

The situation for blue-collar workers is different. The average unemployed blue-collar worker in the U.S. has a retraining budget of $2,500-5,000. By contrast, white-collar retraining investments are typically 5-10 times that amount.

More importantly, blue-collar workers usually have no savings. In the U.S., 40% of workers would be in trouble if faced with a $400 emergency expense. A truck driver who is displaced doesn't have a year of savings to support re-education. He must find work immediately—even if that work pays less and has worse prospects.

3. The Erosion of Union Protection

Union coverage in the U.S. workforce has declined from 35% in the 1950s to 10% today. This decline has been especially steep in manufacturing and transportation. What does this mean? When large-scale automation arrives, unorganized labor cannot collectively bargain.

A useful contrast: When General Motors or Ford automates an assembly line, unionized workers at that plant can at least obtain some form of compensation or redeployment through collective agreements. But an independent truck driver or a small restaurant employee has no power to negotiate with a large company. They are replaced one by one, with no bargaining power whatsoever.

4. The Lethal Combination of Industry Concentration and Geographic Dependency

Many blue-collar jobs are characterized by high geographic concentration. Take truck drivers—they are concentrated in the South, Midwest, and continental interior, typically around logistics hubs. When autonomous trucks arrive, some regions could lose 50%-60% of jobs within 2-3 years.

This is not a matter of individual career changes. This is a matter of regional economic collapse. When Dallas loses 200 truck drivers, the spending of those 200 people decreases, revenues at local restaurants and retail stores decline, potentially causing more job losses. This is an economic multiplier effect.

Part IV: The Overlooked Policy Blind Spot

This is the most ironic part of the whole story: policy discussions about AI and automation employment almost entirely ignore blue-collar workers.

When you read articles about “the future of AI work,” what do you see? You see discussions about programmers, data scientists, marketers, lawyers. You see debates about how universities should revise curricula for the AI era. You see discussions about “needing to learn prompt engineering” or “AI will create new creative jobs.”

What don't you see? You don't see discussions about how a 55-year-old truck driver can restart in 2027. You don't see policy recommendations about how to create retraining infrastructure in a small Texas town. You don't see national strategies about how to prevent entire regions from falling into decline due to transportation automation.

Why?

Because blue-collar workers lack voice. They are not the people journalists in newsrooms interview, not the topic discussed at venture capital conferences, not the subjects of thought experiments on university campuses. Blue-collar unemployment is a “already priced in” phenomenon in the minds of economists and technologists—something that has already happened countless times over the past 30 years of globalization and deindustrialization.

This has created a policy vacuum.

Contrast: The “Epicenter” of White-Collar Impact vs. Blue-Collar Impact

There is a temporal dimension worth noting. The “epicenter” of the AI impact facing white-collar workers falls in 2023-2026. At this point, it is easy to observe, monitor, and discuss. Every month brings news about AI job losses. Policymakers, educational institutions, and companies all have time to react.

The “epicenter” of blue-collar impact won't arrive until 2027-2032—when autonomous trucks reach scale, robots mature enough to work on real construction sites, and food service automation becomes cost-effective. By then, it may be too late. No retraining program can be created and deployed within a year. No new industry can create millions of jobs overnight.

The policy window has closed.

One-Year Retrospective: The Policy Window May Not Be “Closed”

The original text asserts “the policy window has closed.” But two developments in 2026 require revising this judgment.

The EU AI Act's August implementation may be delayed by 16 months. The EU AI Act's high-risk clauses were originally scheduled for full application on 2026-08-02, and AI systems in the workplace would default to the high-risk category (Annex III). But on 2026-05-07, the European Council and Parliament reached a provisional agreement that could delay the application of high-risk AI clauses by up to 16 months. If this goes through, the original expectation of “EU AI Act triggering forward liability for workplace injuries” would be pushed to 2027-2028—giving capital an additional 16-month window to accelerate automation deployment, a policy constraint that perversely incentivizes the very trend the original text worried about.

Update (2026-09): This “possible delay” has materialized, and more definitively than the original text anticipated. The Digital Omnibus AI Simplification Act was voted on by the European Parliament on 2026-06-16, approved by the Council on 2026-06-29, and officially entered into force on 2026-07-27 (published in the EU Official Journal on 2026-07-24). According to legal advisory firm Ogletree's reading of the final text, it delays the compliance deadline for Annex III stand-alone high-risk systems—including hiring, promotion, termination, task allocation, and performance monitoring employment-category AI—from 2026-08-02 to 2027-12-02; embedded high-risk systems (Annex I, AI built into machinery/medical devices) are delayed to 2028-08-02. This delay is written as two fixed dates, no longer carrying conditional triggers like “6-12 months after standards confirmation.” It should be clarified: Article 50 transparency obligations and regulatory authority over general-purpose AI (GPAI) providers still take effect on 2026-08-02 as originally planned and were not delayed—only the heaviest compliance burden of the “high-risk” tier was pushed back.

Trump tariffs did not bring manufacturing reshoring. After the tariff war began in April 2025, U.S. manufacturing actually lost 89K jobs (BLS May employment data + Reason April 29 retrospective + Yahoo Finance “Liberation Day”). Total blue-collar employment has declined by 190K since April 2025; truck transportation is at an 8-year low. This data poses a causal attribution challenge to the original text: on one hand, blue-collar employment is shrinking overall (supporting the myth-rupture thesis); on the other hand, the shrinkage is primarily driven by tariff uncertainty + weak demand, not robot substitution. Attributing all job losses to stories like Aurora / Figure / Flippy is wrong—shutdowns from automation failures (e.g., Kroger closing 3 automated fulfillment centers) and demand recessions caused by tariffs contributed a larger share of unemployment.

Part V: Is There a “Future of Blue-Collar Work”?

The honest answer to this question is: some blue-collar jobs will survive, but the range is narrow.

What Jobs Might Survive?

1. Highly customized work: For example, a custom furniture maker, a specialized restoration mechanic—these jobs require artistic judgment and creativity and are difficult to standardize. But these jobs are characterized by small numbers; pay may be high (but the market is small).

2. Work requiring complex on-site problem-solving: For example, HVAC repair, electrical installation (in new construction). These jobs cannot be fully automated because every site is unique. But even these jobs are being automated through better diagnostic AI and more flexible robots.

3. Care work: Elder care, disability care—these jobs involve elements of human interaction and empathy that are difficult for machines to fully replace. But care work is typically low-paid, physically demanding, and lacks union protection.

What Is the More Realistic Scenario?

The more realistic scenario is that most blue-collar workers will be forced into the service industry—retail, food service, cleaning, delivery. And these jobs are themselves being automated. A truck driver cannot become a restaurant manager because (a) this requires entirely different skills, and (b) the restaurant manager's job itself is also disappearing.

This creates a skill depreciation trap: what was once a dignified, middle-class-wage job (truck driver, earning $4,500-5,800/month) becomes a low-wage service job (warehouse assistant, earning $2,500-3,000/month).

For a 45-year-old truck driver with a family, this is not just a job change. It is a permanent decline in living standards.

Part VI: Policy Recommendations (And Why They Probably Won't Be Implemented)

If we assume policymakers want to prevent this disaster, what would they do?

Short-Term (Now-2027)

1. Immediately launch a national retraining program: Targeting workers in high-automation-risk industries (transportation, manufacturing, food service). This requires $50-100 billion in federal investment for: - Establishing training centers in areas concentrated with truck drivers and manufacturing workers - Subsidizing living costs during unemployment - Helping workers relocate to areas with job opportunities

2. Industry-specific social assistance: In regions where automation is imminent (e.g., Texas, Oklahoma, Kansas), establish “transition funds” to support local business diversification.

3. Adjust tax incentives for automation: Tax companies that use automation to replace large numbers of workers, and use that revenue to retrain affected workers.

Medium-Term (2027-2030)

4. Large-scale education reform: Reform high school curricula to teach “blue-collar work transition” skills earlier—diagnostic thinking, technical literacy, adaptability.

5. Rebuild community infrastructure: Invest in new industries (renewable energy, grid modernization, etc.) in regions where unemployment may surge, to create alternative employment.

Long-Term (2030+)

6. Universal basic income or similar programs: If blue-collar jobs shrink substantially, society needs a new social contract to support those who cannot integrate into the new economy.

Why These Probably Won't Happen

The reality is that none of this is likely to happen—at least not at sufficient scale or speed to effectively prevent the disaster.

First, political will does not exist. Blue-collar workers are numerous in American politics but lack voice. They are scattered across the country with no strong union to represent their interests. Meanwhile, the beneficiaries of automation (tech companies, logistics companies, restaurant chains) have powerful lobbying forces in Washington.

Second, temporal asymmetry. Automation is driven by market forces and can be implemented within months. Policy responses take years. By the time they arrive, it is usually too late.

Third, psychological denial. Many policymakers and elites still believe the story that “new jobs will be created.” This story has sometimes been true over the past two centuries, but there is no reason to believe it will remain true in the AI era.

Conclusion: The Overlooked Disaster

The automation of blue-collar work will not be a sudden, dramatic collapse. It will be a slow, silent decline. One month, a logistics company loses 50 truck drivers. The next month, a restaurant chain in another region closes 100 fry stations. No one will call it a “crisis.” The media will continue discussing white-collar jobs. Policymakers will be surprised by a problem they never truly paid attention to.

By then, it will be too late.

This is not a story about technology. This is a story about power, attention, and social injustice. Blue-collar workers once believed their jobs were safe because they believed a story—a story told by economists, reinforced by the press, repeated by technologists. But this story was never validated. It was simply a comfortable assumption, a story that allowed people to feel reassured while automation ravaged white-collar work.

Now this story is rupturing. And when it fully ruptures, we will find ourselves in a world that is prepared for neither the white-collar crisis nor the blue-collar crisis.

That will be the real crisis.

Final Note

According to the U.S. Bureau of Labor Statistics, the transportation and warehousing sector has approximately 3.5 million workers. According to BLS Occupational Employment Statistics, the median annual salary for these jobs is $62,000. If even half the workers in this one industry are replaced by automation, we are talking about 1.7 million people unemployed and $105 billion in lost annual income. Multiply by food service, construction, manufacturing, and other industries, and the total becomes an unimaginable, life-changing number. This number is not a guess. It is based on real commercial decisions we are seeing today. The question is: will we act before the disaster arrives, or will we only start reacting after it has already arrived? Based on history, the answer is usually the latter.

Data Update: 2025–2026 Field Progress

Aurora: From Concept to Commercial Reality

In April 2025, Aurora Innovation officially launched commercial driverless truck services in Texas—becoming the first company globally to conduct truly unmanned commercial operations with heavy trucks on public roads. The initial routes were Dallas to Houston, with customers including Uber Freight and Hirschbach Motor Lines.

Data has since accumulated rapidly: - By early 2026, completed over 250,000 miles of zero-incident driverless operation - Routes expanded to Fort Worth → El Paso, with announced expansion to Phoenix, Arizona - 2026 plan to deploy 200+ autonomous trucks - New-generation hardware costs reduced 50% from the previous generation, with simultaneous improvements in performance and durability

Hirschbach has signed an agreement with Aurora to purchase hundreds of autonomous trucks. This is not an MOU—it is a commercially binding deployment plan.

Aurora's numbers show the industry is crossing the threshold from “concept phase” into scaled commercial deployment.

Humanoid Robots: From Experiments to the Factory Floor

In early 2026, Tesla had deployed over 1,000 Optimus Gen 3 humanoid robots in its manufacturing plants. Musk admitted on the Q4 2025 earnings call that these robots are currently primarily used for learning and data collection and have not yet entered the “useful work” stage—a rare candid admission indicating we are still in a transitional period.

However, other companies are pushing forward:

• Agility Robotics' Digit: Has entered Amazon fulfillment centers, performing real warehouse tasks
• Apptronik's Apollo: Is conducting logistics delivery pilots at Mercedes-Benz factories in Europe
• Boston Dynamics Atlas: Commercialization projects for industrial customers continue to advance

Important reality check: humanoid robots' capabilities in 2026 remain limited—they can only complete a small number of specific tasks, and their speed and reliability have not yet reached the levels of traditional industrial robots from a decade ago. But the direction of the trend is clear, and the speed is faster than anyone expected.

Further Reading: - Aurora Innovation (2025). Aurora Begins Commercial Driverless Trucking in Texas. Press release - Aurora Innovation (2026). Aurora Triples Driverless Network to 10 Routes and Prepares to Expand Across U.S. Sun Belt. Original press release (the originally cited nationaltoday.com article is 404; replaced with Aurora's official primary source) - Michigan Journal of Economics (2026). AI on the Job: How Blue-Collar and White-Collar Workers are Impacted. Analysis

Academic Frontier (arXiv preprints): - del Rio-Chanona, R.M. et al. (2025, Sep). AI and jobs: A review of theory, estimates, and evidence. International Labour Organization / University of Oxford. The most comprehensive AI employment impact review to date: RCT experiments show AI boosts productivity 20-60%, field experiments 15-30%; junior workers benefit more on simple tasks but simultaneously face shrinking demand for entry-level positions—skill development pathways are being truncated. arXiv:2509.15265 - Gupta, R. & Kumar, S. (2026, Mar). Agentic AI and Occupational Displacement: A Multi-Regional Task Exposure Analysis of Emerging Labor Market Disruption. Submitted to the IMF-OECD-PIIE-World Bank “Labor Markets and Structural Transformation” conference. First introduction of the “Agentic Task Exposure (ATE)” scoring framework—targeting not single sub-task substitution but AI agents taking over entire workflows: by 2030, 93.2% of six information-intensive occupational categories (finance, law, healthcare, sales, etc.) will cross the medium-risk threshold (ATE≥0.35), with the blue-white collar boundary fully breached. arXiv:2604.00186 - Frank, M.R. et al. (2026, Jan). AI-exposed jobs deteriorated before ChatGPT. econ.GN / cs.AI. Using U.S. monthly unemployment insurance records + millions of LinkedIn profiles + millions of university course catalogs as evidence: unemployment risk for AI-exposed jobs was already rising in early 2022—months before ChatGPT's release—and the proportion of graduates entering AI-exposed jobs began declining from the class of 2021 onward—damage preceded public awareness and the generative AI wave. arXiv:2601.02554

Temporal Dimension: How Academia and Industry Incrementally Revised Their Judgments on AI's Employment Impact

This was not a sudden cognitive revolution. It was a series of judgment corrections, collapsed assumptions, and ex-post rationalizations of “we knew it all along”—distributed across a timeline of more than a decade.

Phase I (2003—2012): The Theoretical Foundation for Blue-Collar Anxiety

The intellectual starting point traces back to the Routinization Hypothesis proposed by Harvard economist David Autor and colleagues in 2003: computers excel at executing explicit, codable rules, so repetitive manual and cognitive tasks are most vulnerable, while complex manual and non-routine cognitive tasks are relatively safe.

This framework dominated labor economics throughout the 2000s. Its policy implication was clear: manufacturing blue-collar workers were at risk; lawyers, doctors, engineers were not. Automation was a “class” problem—it punished those who already had little bargaining power.

This was not wrong. But it was incomplete—in a way that later proved fatal.

Phase II (2013): Frey & Osborne's 47% Shock

In September 2013, Carl Frey and Michael Osborne of Oxford University's Martin School published the working paper The Future of Employment. They analyzed 702 occupations from the U.S. Bureau of Labor Statistics, training a classifier to determine whether each occupation was computerizable.

The conclusion was disturbing: 47% of U.S. employment faced high automation risk.

This number immediately became media headlines and sparked broad policy discussion. But Frey & Osborne's methodology had a neglected detail: they assessed entire occupations, not individual tasks within occupations. In other words, their question was “can all the work of this occupation be done by machines,” not “how many tasks within this occupation can be taken over by machines.”

This methodological choice set the stage for the great debate two years later.

Phase III (2016): OECD's “9%” and the Task Decomposition Argument

In 2016, Arntz, Gregory, and Zierahn of the OECD published a study directly rebutting Frey & Osborne. They used the same national data but changed the unit of analysis: not occupations, but specific tasks within occupations.

Conclusion: In the U.S., only about 9% of workers face truly high automation risk—because even in “high-risk occupations,” workers actually perform tasks that machines struggle to replace (improvisational judgment, interpersonal communication, physical dexterity, etc.) on a daily basis.

This 9% vs. 47% gulf gave policymakers enormous breathing room: “The problem isn't that severe.” Technology optimists counter-attacked, emphasizing that every previous technological revolution had created more jobs, and “this time won't be different.”

Academia descended into methodological trench warfare; policy discussion effectively stalled.

Phase IV (2018—2021): The First Causal Evidence of the Robotics Era

Daron Acemoglu and Pascual Restrepo's 2018 paper broke the impasse—using more rigorous econometric methods and exploiting differential penetration of industrial robots across regions as an instrumental variable, they confirmed at the causal inference level for the first time:

Each standard deviation increase in industrial robot density reduces the commuting zone's employment rate by 0.18-0.34 percentage points and wages by approximately 0.25-0.5%.

Blue-collar workers, especially mid-skill blue-collar workers, bore the most direct impact. This was no longer a prediction—it was ex-post econometric confirmation.

But note: research in this period focused on industrial robots—welding, painting, assembly lines. Their capability boundaries were clear, the market mature, and data available. Cognitive work, services, and tasks requiring natural language were still considered safe territory.

Phase V (November 2022—2023): ChatGPT's Paradigm Reversal

On November 30, 2022, OpenAI released ChatGPT.

This was the watershed for the entire discussion—not because it immediately replaced anyone, but because it shattered the “protection myth” for cognitive work. Legal briefs written by lawyers, code written by programmers, drafts written by journalists, emails replied by customer service…suddenly, all these tasks appeared on the “replaceable” list.

The academic reaction split into two poles:

One side (Goldman Sachs, 2023): approximately 300 million full-time jobs globally could be automated by generative AI; up to two-thirds of occupations in the U.S. and Europe would be affected by AI to varying degrees.

The other side (Acemoglu, 2023): In the paper “Simple Macroeconomics of AI,” the economist renowned for causal evidence maintained rare composure. He estimated that over the next decade, only about 5% of tasks could be performed by AI in a cost-effective manner—meaning the real impact on GDP and employment is far smaller than the media narrative.

Meanwhile, industry exhibited a noticeable “qualitative shift” during this period: IBM announced a pause on hiring for AI-replaceable positions (approximately 7,800). Tech media began tracking “AI-related layoffs.” But most companies still packaged layoffs as “workforce restructuring” rather than “AI replacement.”

Blue-collar work virtually disappeared from public discussion during this phase—ChatGPT turned everyone's gaze toward white-collar workers.

Phase VI (2024—2025): The Field Evidence Period, Two Cracks Widening Simultaneously

The hallmark of this period: macro debates receded, field data poured in.

Blue-collar side: Aurora Innovation officially launched commercial driverless truck operations in Texas in April 2025. Digit robots entered Amazon warehouses. Humanoid robot cost curves began to bend. These were no longer demo videos—they were contractually bound commercial deployments.

White-collar side: Duolingo, SAP, Shopify, Workday, and other companies explicitly cited AI efficiency as a reason for layoffs—this was the first time “AI replacement” discourse was openly adopted by companies rather than euphemistically packaged.

Academia welcomed its most comprehensive review to date. ILO and Oxford University's del Rio-Chanona et al. (September 2025, arXiv:2509.15265) integrated dozens of RCT experiments and field studies, arriving at a seemingly paradoxical conclusion:

AI boosted worker productivity by 20-60% (RCT experiments) or 15-30% (field experiments), but demand for entry-level positions is continuously shrinking—the junior workers who benefit the most are simultaneously the group whose career pathways are most severely truncated.

Efficiency gains and job elimination coexist. This paradox is the most important finding of this period.

Phase VII (2026): Damage Precedes Awareness; Market Failure Theoretically Confirmed

Three papers from 2026 constitute the current frontier of understanding:

Frank et al. (arXiv:2601.02554) used monthly unemployment insurance records, LinkedIn profiles, and university course catalogs to demonstrate empirically: deterioration of job quality in AI-exposed occupations had already begun in early 2022—nearly a year before ChatGPT's release. The damage was silent, structural, and preceded public awareness and media discussion.

Gupta & Kumar (arXiv:2604.00186) introduced the “Agentic Task Exposure (ATE)” framework: unlike earlier research focusing on single tasks, they assessed the ability of AI agents to take over entire workflows. Conclusion: by 2030, 93.2% of six information-intensive occupational categories will cross the medium-risk threshold.

Falk & Tsoukalas (arXiv:2603.20617) offered the most powerful theoretical explanation to date: why do firms lay off workers beyond the collectively optimal level? The answer is demand externalities—when one firm in an industry automates, market demand doesn't expand accordingly but is captured by competitors; this forces other firms to automate as well, creating an arms-race-style lock-in. Even knowing the collective outcome is bad, no individual firm can stop.

What does this mean? Capital taxes, employee stock ownership, UBI, vocational training—none of the conventional policy tools can resolve the distortion in competitive incentives itself. Only a Pigouvian automation tax (imposing a tax rate on automation equal to its marginal social damage) can correct the problem at its root.

This is the complete cognitive journey from “AI will affect employment” to “structural market failure requiring specific policy instruments.”

Viewpoint Evolution Comparison Table

A Noteworthy Structural Bias

Looking back at this history, a systematic cognitive bias has persisted throughout: each round of discussion lagged actual damage by at least two years.

• The blue-collar risk predicted by Frey & Osborne was considered exaggerated by policymakers before empirical data on actual robot impact emerged.
• Before ChatGPT, AI's impact on white-collar work had already begun—Frank et al.'s data shows this started in early 2022—but public awareness didn't erupt until 2023.
• The logic of market failure (arms races, excessive layoffs) only received a theoretical model in 2026, but the behavior predicted by the model had been underway for some time.

This is not an unintentional oversight. When damage is still accumulating, there are no dramatic events to report; when damage has become fact, there is finally sufficient data to support analysis. Structural unemployment is inherently harder to detect in real time than asset bubble bursts.

This also means: current academic judgments about the post-2026 period are most likely still lagging behind unfolding reality.

Latest Developments (May-June 2026)

1. Aurora Lands McLane: Driverless Trucks Move from “Mileage” to “Daily Operations”

On May 6, 2026, Aurora announced it had launched fully driverless long-haul transport with McLane, one of North America's largest food distributors, on the Dallas-Houston corridor, expanding from twice-weekly round trips to seven-day continuous operations. The long-haul segment is completed by the Aurora Driver, with the terminal last mile handed back to McLane's human drivers. This is the company's second “revenue-generating” commercial contract after Detmar, and its significance is that shippers have shifted from the old “complete a pilot and write a report” posture to slotting driverless trucks into daily dispatch schedules. Aurora itself set the target in its Q1 2026 shareholder letter of “running 200+ driverless trucks in the Sun Belt by year-end” and activating the second-generation hardware package, no longer requiring customers to retain observer seats.

Advancing in parallel with Aurora is Kodiak AI: the company disclosed in its Q1 2026 earnings that it has delivered 28 Customer-Owned driverless trucks, with cumulative paid driverless driving hours exceeding 23,500. The two companies occupy the Southwest-Central South transport corridors without overlap, meaning the “disappearance curve” for the truck driver occupation is shifting from “occasionally mentioned in quarterly news” to “observable on a monthly basis.”

2. Figure 02 at BMW Completes 11-Month, 30,000-Vehicle Production Closed Loop

In early May, Figure AI officially published the milestone settlement with BMW's Spartanburg plant: two Figure 02 humanoid robots in the body shop accumulated 1,250 operating hours, loaded over 90,000 sheet metal parts, and participated in the production of over 30,000 BMW X3s, with single-shift placement accuracy exceeding 99% and 84-second per-part cycle time compliance. This is the first time a humanoid robot on a mainstream automotive assembly line has completed a billable production contract—no longer a “demo video,” but a real process counted toward BMW's North American plant's annual capacity.

Shortly after, BMW announced in May that it was replicating the same project in Europe: the Leipzig plant launched Germany's first humanoid robot production pilot, and established a “Physical AI Production Capability Center” to coordinate global deployment. The political significance of this news line exceeds its technical significance—Germany is one of the developed economies with the strongest union power and strictest employment protections; BMW choosing Leipzig rather than a more permissive region signals that manufacturing capital has judged that “humanoid + IG Metall negotiations” is workable.

Update (2026-09): The Spartanburg workstation has been upgraded—Figure AI released Figure 03 on June 25, 2026, taking over Figure 02's position, with tasks shifting from “body shop welding assistance” to “material sorting” (organizing unclassified parts into carts in sequence for “just-in-time delivery” to the production line), adding soft safety components, wireless charging, voice interaction, and hands with tactile sensors. The task switch between two hardware generations itself indicates: humanoid robots are still in rapid iteration, not “install once and run for a decade”定型production lines. The Leipzig AEON project has also advanced: entering second-phase testing in April 2026, with full piloting starting in summer of that year, and two robots expected to be formally deployed on the high-voltage battery assembly and parts production lines by year-end—but as of September, “two” remains the only disclosed deployment scale for that plant.

3. Agility Digit Crosses the 100K-Item Threshold and Lands at Toyota's North American Plant

The “most-shipped humanoid robot” in the industry is currently Agility Robotics' Digit. According to Sacra's May comparative review, Digit at GXO's Flowery Branch, Georgia warehouse has cumulatively moved over 100,000 totes and completed a full year of continuous full-time operation; simultaneously, Agility signed a Robots-as-a-Service contract with Toyota Canada, deploying 7+ Digit units to assist with material handling at the RAV4 plant. Both numbers crossing thresholds simultaneously means: humanoid robots are shifting from “specially authorized prototypes” to “monthly-billed assets”—the billing unit shifting from “events” to “hours,” a standard precursor before any labor substitution curve enters its exponential phase.

4. Actual Blue-Collar Job Losses: The “Non-Automation” Layoff Wave in Logistics and Manufacturing

Notably, the blue-collar unemployment at the start of 2026 was not entirely caused directly by robots. FreightWaves' May summary shows that over 2,000 logistics and manufacturing workers had already been laid off early in the year, involving RailCrew Xpress (lost CSX contract, 400+ people), AVI Food Systems (297), UPS, FedEx, and others. Even more worth tracking is Kroger closing three automated fulfillment centers in January and cutting 1,000+ employees—this is unemployment caused by “automation failure,” not “automation success,” and the two have entirely different political narrative implications for labor. Appearing simultaneously is a reverse migration of white-collar workers transitioning to plumbing/electrical work: blue-collar wages are being pushed up by construction and skilled trades shortages, forming reverse data within the same time window as transportation/warehousing job shrinkage. This suggests the original text's “blue-collar myth rupture” needs a footnote: what is rupturing is scalable, routinizable blue-collar work (long-haul trucking, restaurant fry stations, assembly line production), while highly contextual, on-site-service blue-collar work (electricians, HVAC, roofers) remains in a wage-upswing cycle.

Update (2026-09): The wave of non-automation layoffs hasn't stopped. FreightWaves' August summary shows that in August alone, logistics, warehousing, and manufacturing added 7,000+ more layoffs, including Postal Center International (457 across Florida/Texas/Massachusetts), GXO Logistics (layoffs at two Kentucky and Pennsylvania sites due to lost distribution contracts), and DHL Supply Chain (layoffs triggered by service contract changes). This continues to confirm a judgment: in 2026, the direct trigger for blue-collar job losses is most often still “lost contracts / plant closure restructuring / weak demand,” not “robots replacing people”—but for the laid-off workers, the result of both causes is the same: the job disappears. The counter-examples are also expanding: CNBC reports that young electricians on the front lines of data center construction in Texas can earn pre-tax annual salaries of $240,000-280,000 with zero student loan burden; the same Randstad analysis of 50+ million job postings shows that from 2022-2026, robotics technician postings grew 107%, HVAC/refrigeration engineer postings grew 67%, industrial automation technician postings grew 51%, and traditional electrician postings grew 18% over three years—demand-side hiring volume itself is still accelerating, the most direct evidence that “these jobs are not yet covered by automation.”

5. Policy Clock: EU AI Act High-Risk Clauses Take Effect in August (Overturned—See 2026-09 Update)

The EU-side policy clock is also approaching: the EU AI Act's high-risk clauses were originally scheduled for full application on August 2, 2026, and AI systems in the workplace would default to the high-risk category, triggering obligations for employee notification, human oversight, discrimination monitoring, and log retention. The absolute prohibition on workplace emotion recognition (Article 5) is a separate matter that took effect in February 2025 and is not affected by the timeline discussed in this section. The original text judged here: this means BMW Leipzig's humanoid robot project would, from August onward, be among the first blue-collar automation deployments in Europe operating under a legal context governed by the AI Act—any malfunction, misoperation, or workplace injury would become the first test case under the Act.

Update (2026-09): This judgment did not hold. The EU's Digital Omnibus AI Simplification Act entered into force on 2026-07-27, delaying the compliance deadline for employment-category high-risk AI (hiring/promotion/termination/performance monitoring/task allocation) from 2026-08-02 to 2027-12-02 (see the full timeline in the “One-Year Retrospective” section above). In other words, BMW Leipzig's AEON pilot will not be “among the first blue-collar automation deployments governed by the AI Act's high-risk clauses”—at least not until late 2027 will it truly enter this tier of regulatory obligations; until then, any malfunction, misoperation, or workplace injury remains in a compliance vacuum with respect to employment-category high-risk rules. The original text's prediction about the direction of the policy clock was correct (regulation will eventually cover such deployments), but the timing judgment was falsified.

Link Verification (L3)

Reflections and Boundaries: Under What Conditions Do This Article's Claims Hold

The above uses extensive real data to argue for “the rupture of the blue-collar myth,” but every core assertion has edges that can be challenged. Below, these vulnerabilities are placed on the table—not to overturn the conclusions, but to let readers know: this article is a list of hypotheses, not a final verdict.

1. High-End Customized Blue-Collar Wages Are Rising, Not Falling

The article portrays blue-collar workers as “a unified class about to be torn apart by automation,” but 2025-2026 field data shows divergence within blue-collar, not uniform decline. Fortune reported in March 2026 that data center construction is driving skilled blue-collar wages up, with commercial electrician median salaries rising to $85K-95K, with top earners reaching $120K-180K; JLL's April “silent army” report priced the U.S. skilled trades shortage gap at $1 trillion; WEF data for the same period shows 37% of Gen Z graduates are actively entering blue-collar tracks, forming a labor flow in the opposite direction from the “blue-collar collapse” narrative.

What this means for the article: Part III's sweeping assertion that “blue-collar workers are more vulnerable than white-collar workers” needs a qualifier—what is rupturing is scalable, routinizable, geographically concentrated blue-collar work (long-haul trucking, fry stations, assembly line production), while contextual, on-site-service, judgment-requiring blue-collar work (electricians/plumbers/HVAC/roofers/data center construction workers) is in the middle of a wage upswing cycle. The conclusion should be “divergence within the blue-collar myth,” not “overall rupture of the blue-collar myth.”

China's 2013-2025 blue-white collar wage gap narrowed from ¥3,344 to ¥2,250 (-32.7%)—a direction entirely opposite to the U.S. pattern of “white-collar gains + entry-level blue-collar losses.” Specific data: in 2025, China's maternity matrons earned ¥10,128, delivery drivers ¥8,325, truck drivers ¥8,279; in 2026, certified electricians/welders earned ¥6,800-12,000+, with core positions in tier-1 cities paying 15%-20% more than local administrative white-collar workers (aggregated from Statista + ERI + China-Briefing 2026).

But note, this does not refute the article; it supplements a boundary condition the article already acknowledges in §1. The title “Rupture of the Blue-Collar Myth” should be read as “rupture of the scalable, routinizable, geographically concentrated blue-collar myth”—long-haul trucking, fry stations, and assembly line production will be hit by automation; electricians, HVAC technicians, roofers, and data center construction workers are seeing wage increases. “Blue-collar” is a label flattened by narrative; in reality, it encompasses two industries with opposite trajectories.

6. The Optimus Production Line Launch Could Reset the Validity Period of Acemoglu's 5% Assumption

Acemoglu's 2024 NBER w32487 judgment that “only 5% of tasks can be performed by AI in a cost-effective manner in the next decade” is predicated on general-purpose humanoid robots not being mass-produced. Tesla converting Fremont into an Optimus production line means the validity period of the “5% assumption” could be prematurely broken by Optimus's marginal cost curve—Musk's public target is to reduce Optimus's per-unit cost to $20,000-30,000. If achieved, the equation for “whether it's cost-effective to replace blue-collar workers with robots” is completely rewritten.

But this is a prerequisite change, not an accomplished fact. Musk admitted on the Q4 2025 earnings call that “Optimus is not yet in material use.” Interpreting this as “a potential change within 5 years” rather than “something that has already happened” is the more honest posture. The original § Reflections should add a falsifiable condition: if by the end of 2028, Optimus's per-unit cost still exceeds $50,000 and cumulative deployments remain below 10,000 units, this article's claim that “the window for blue-collar myth rupture has opened” should be significantly delayed.

2. Aurora and Optimus's Scale Shares Are Far Smaller Than the Narrative Implies

The article juxtaposes four threads—Aurora-McLane, Tesla Optimus, Figure 02 at BMW, Digit at Amazon—into “the automation inflection point has arrived,” but the denominator is omitted in all four cases:

• The total U.S. Class 8 heavy truck fleet is approximately 4 million vehicles; Aurora's 2026 target of “200+ vehicles” + Kodiak's 28 delivered—commercial driverless trucks currently represent less than 0.006% of all Class 8 trucks. The same ATA data shows the current U.S. truck driver shortage at approximately 60-82K, expected to expand to 160K by 2028—meaning autonomous trucks will fill the shortage for the foreseeable future, not replace the existing workforce.
• BMW's Spartanburg plant actually produced 412,799 X-series vehicles in 2025; the 30,000 vehicles Figure 02 participated in represent approximately 7.3% of that plant's annual output, but Figure deployed only two humanoid robots with 1,250 cumulative hours—translating to approximately 1.7 hours of effective operation per unit per day.
• Tesla Optimus: Musk himself admitted on the January 2026 Q4 earnings call that “It's not in usage in our factories in a material way. It's more so that the robot can learn”—while not contradicting the article body's statement that “Tesla has deployed over 1,000 Optimus Gen 3,” this severely undermines its force as evidence of an “automation inflection point.” It is recommended that the “thousand Optimus” paragraph in the body add Musk's own qualification.
• Update (2026-09) Agility Digit: The “100K-item threshold” cited in the body is a genuine commercial milestone, but an independent market analysis (April 2026) estimates Digit's global actual installed base at only approximately 40-150 units (median estimate ~75), and all 100,000 items were handled at the same single warehouse in Flowery Branch, Georgia—Agility Robotics' SPAC disclosure filing shows that as of mid-2026, Digit has expanded to 9 customer sites including GXO, Amazon, Toyota, and Schaeffler with over 65,000 cumulative operating hours, but “a milestone achieved through deep single-site penetration” and “scale across the entire network” are two different things; media headlines easily conflate the former with the latter.

What this means for the article: Part II's evidence chain needs to replace “absolute numbers” with a two-column “share + slope” presentation—looking only at absolute unit counts overestimates the current intensity of impact, while looking only at slope underestimates the speed once the substitution curve inflects upward. Both must be presented simultaneously.

3. Several Data Points' Source Chains Need Re-verification

• “By end of 2026, AI has globally replaced approximately 2 million manufacturing workers” is cited from Demandsage, a secondary data aggregation site, not a primary research institution. The original source most likely traces to IFR (International Federation of Robotics) or Oxford Economics prediction models, but both are predictions, not established facts. The body should be revised to “according to [institution] 2024 projections, by end of 2026…” or the claim deleted outright.
• “Data labelers 2022→2025: 3M→6.5M” — this number is used in the article to illustrate “AI both kills and creates jobs,” but does not define the proportion of AI-related labeling vs. general image/text/speech labeling—a significant portion comes from non-generative AI tasks carried over from the traditional machine learning era. If the original numbers come from Cognilytica or Grand View market reports, the scope should be explicitly noted.
• “Blue-collar workers lack voice”: This judgment contradicts 2025 field observations. Teamsters currently has approximately 1.25 million members; UAW forced the Big Three automakers to sign contracts with 25% wage increase clauses in late 2023, and in 2025 continued to demand restoration of COLA and elimination of tiered wages. Blue-collar collective action capacity is recovering, not disappearing. Part III's “erosion of union protection” is true, but “lack of voice” is false—these are two different facts.

4. Boundary Conditions: Geographic, Industry, and Temporal Sensitivity

• Geographic boundary: This article's claims hold under the employment structures of North America + EU. The Global South (India, Southeast Asia, Sub-Saharan Africa) is currently still actively importing blue-collar labor to fill the manufacturing vacuum from nearshoring/friendshoring; the automation curve is offset by at least 10 years from North America. Claiming “blue-collar myth rupture” as a global phenomenon would be inaccurate.
• Industry boundary: Electricians/plumbers/HVAC/roofers/old-home renovation/healthcare caregiving/precision agriculture (horticulture, vineyards)—these single-point customized, low-standardization tracks have no robots that can replace them wholesale under 2026 technology. Within a visible time window (5-8 years), wages in this portion of blue-collar work will remain structurally upward.
• Temporal sensitivity: All quantitative predictions in this article assume that the current three conditions—regulatory vacuum + cheap capital + abundant energy—persist. If any one changes—for example, EU AI Act high-risk clauses taking effect in August forcing forward liability for workplace injuries, or a U.S. state率先legislating a Pigouvian tax per ton-mile on autonomous trucks, or an 18-month-level disruption in the H100/B200 supply chain—the slope of the substitution curve would change immediately.

5. Competing Frameworks: WEF and Acemoglu's “Net Growth” Narrative

This article uses Falk-Tsoukalas's “Pigouvian tax as the sole solution” as its policy landing point, but at least two competing frameworks exist simultaneously that are equally citable with seriousness:

• WEF Future of Jobs 2025: By 2030, a projected net gain of 78 million jobs globally (170M created - 92M displaced), with the five fastest-growing job categories including delivery drivers, care workers, farm workers, construction workers—four blue-collar categories. WEF's methodology is not without flaws (employer surveys skew optimistic), but it represents a “AI also creates blue-collar jobs” framework accepted by both academia and industry.
• Acemoglu 2024-2025 revision: The same Acemoglu cited in this article for the “5% tasks” figure, in his 2024 NBER paper and 2025 IMF cross-country extension, continues to maintain a ten-year TFP increment ceiling of approximately 0.53%-0.66%—meaning AI's overall impact on macro employment is far smaller than Goldman Sachs' high estimate of 300 million jobs. In other words: the economist this article treats as a “prophet of blue-collar impact” himself maintains a conservative judgment on total impact magnitude. Categorizing him in the “blue-collar camp” is an over-deployment of his position.

Furthermore, looking back at history: Frey & Osborne's 47% figure from 2013—between 2013 and 2021, the U.S. actually added 16 million jobs and unemployment fell to 3.7%—the correlation coefficient between prediction and reality was only 0.26. This doesn't mean this article will necessarily repeat Frey-Osborne's error—LLM/embodied intelligence capability curves are indeed different from 2013 assumptions—but it reminds us: long-run pessimistic predictions about labor markets have had a near-zero hit rate over the past decade. This article should set a falsifiable condition for itself: if by the end of 2028, U.S. truck driver employment remains above 3.4 million, or the median actual wage for electricians/plumbers continues to grow at 5%+, this article's core claims would need to be rewritten.

Update (2026-09): Frey-Osborne's failure is not an isolated case. A peer-reviewed study by Coelli & Borland at the University of Melbourne directly tested whether Frey-Osborne's occupational risk classification could explain actual occupation-level employment changes in the U.S. from 2013-2018; the conclusion was “no”—they found that the Autor 2003 “task routinization” framework (scoring based on tasks rather than entire occupations) cited at the opening of this article had significantly stronger predictive power. This provides a belated validation for the foreshadowing planted in Part I: what history has shown to be problematic is not the directional judgment of “whether blue-collar will be replaced,” but the methodology of “packaging an entire occupation and judging whether it can be automated”—which has systematic flaws. This is why this article argues using “specific tasks + specific companies + specific numbers” rather than “occupation X% will disappear.”

Revised One-Sentence Summary (2026-06)

This article's claims are robust within the intersection of “long-haul trucking + assembly line production + standardized frying/packaging + North America/EU + 2027-2032 time window”; they do not hold within the reverse intersection of electricians / HVAC / roofers / data center construction workers / China blue-collar overall. Readers should treat this article as an alert targeting specific sub-sectors, not a verdict applicable to all blue-collar workers.

Falsifiable Conditions (added 2026-06): - If by end of 2028, driverless trucks as a share of the U.S. Class 8 heavy truck fleet remain < 0.5% (500 vehicles / 4M), this article's judgment that “the substitution curve enters an acceleration phase in 2027-2032” should be delayed. - If by end of 2028, Tesla Optimus cumulative deployments < 10,000 units or per-unit cost > $50,000, the “Acemoglu 5% assumption” should continue to be considered valid, and this article's core impact magnitude judgment should be dialed down. - If the U.S. data center electrician median salary continues to grow at 5%+ through 2028, this article's “total blue-collar collapse” title should be rewritten as “scalable blue-collar collapse.”