1. The Lunchtime Layoff Text
November 30, 2023. A Thursday. Just before the lunch hour.
Approximately 240 call-center agents working for a subcontractor handling KB Kookmin Bank's customer service received termination notices on their smartphones simultaneously. Contract terminated.1 Kim Hyun-ju (a pseudonym) had been lifting the lid off her lunchbox. "I got it right before lunch, so I couldn't eat. I had no idea how I was going to survive."
There had been no prior notice. The agents had not even been formally informed that an AI customer-service system was being deployed. They learned of it through angry callers. "What is this AI agent?" Customers demanded answers the human agents could not give, so the agents themselves started calling in to discover what had replaced them.
Hyun Jin-a (a pseudonym) said after the dismissal: "In the end, weren't we the ones who trained that AI? Countless employees were mobilized to catch its errors. We corrected the text. That data will be used."
It was a structure of self-destructive labor. Every time the AI gave a wrong answer, the agents corrected it. The corrected data fed back into the AI's training set. The more accurate the AI became, the narrower the agents' foothold grew. They were training the machine that would replace them, and when the lesson was over, they were expelled from the classroom.
Kim Hyun-ju had worked for twenty years. "My professional pride and expertise were shattered overnight by a single word: 'efficiency.'"
But KB Kookmin Bank did not issue the termination. The agents were not KB employees. They belonged to a subcontractor — an outsourced labor firm. KB introduced the AI; the subcontractor cut the workforce. The principal employer stood outside the perimeter of responsibility. Indirect employment — a structure that outsources even the accountability for dismissal.
In Chapter 3, we saw the victims of subprime mortgages. Financial products engineered by banks took away homes. The victims did not understand the structure of those products and were told they did not need to. "It's safe," they were assured. The call-center agents of Chapter 9 inhabit the same structure. The data they corrected every day improved the AI's performance, and on the day that performance peaked, they vanished with a single text message. Under the banner of efficiency, financial innovation once took homes; now AI innovation takes jobs.
One agent testified: "There was a single mother raising two children alone... This isn't simply about someone losing a job. It's a situation that drives people toward death."
2. Three Tiers of Transition
It is not only call-center agents. The people displaced by AI are stacked in three layers.
Blue-Collar — Subordinates of the Algorithm
Driver Lee from Chapter 7 returns. His story grows darker.
Two a.m. In the darkness of a one-room apartment in Suwon, Driver Lee's smartphone screen lit up. One notification: "Your account has been temporarily restricted." He was halfway through removing his delivery uniform and stared at the screen. No reason given. He tapped the app's customer center. A chatbot appeared. "Restriction due to policy violation." Which policy? The same answer. He tried again. The same answer. The typing animation looped — a machine pretending to think. He tried phone support. Hold music played. He waited forty minutes, then hung up. He called again. The same music.
When an invisible boss fires you, there is no window to protest through. Account deactivation — the platform's version of a pink slip. There is no employment contract. There is no employer. Applying for unemployment benefits fails at the intake stage. You must prove "involuntary separation," but whether an algorithm's account suspension constitutes a dismissal or a contract expiration is legally ambiguous.
As Chapter 7 showed, the algorithm's design forces a choice between safety and livelihood — and livelihood wins every time.
A delivery rider who participated in an academic study testified: "AI determines who gets how much volume, and which route to take."2 As the daily average of one casualty in Chengdu — documented in Chapter 5 — demonstrates, the speed engineered by algorithms exceeds human limits.3 South Korea was no different.
In July 2024, a popular rider who had been documenting fifteen-to-seventeen-hour delivery days on YouTube was killed in a traffic accident. The Riders' Union (라이더유니온) held a rally for industrial-accident testimony, and the Seoul Metropolitan Government granted platform workers' unions their first official legal status.4
Baedal Minjok (배달의민족, South Korea's dominant food-delivery platform) held a market share of roughly 60 percent as of 2023–2024. The unemployment-benefit receipt rate for the riders powering that market was 0.68 percent.5 Driver Lee's monthly income was 2.8 million won (roughly $2,100) — the reward for ten-to-twelve-hour days, rain or snow. The overall monthly average for platform workers was just 1.45 million won (roughly $1,090).6 Industrial-accident insurance was mandatory for special-employment workers in principle, but an opt-out provision undercut its effectiveness. Employment insurance enrollment was technically possible, but actual enrollment rates were dismal because platform companies refused to be recognized as "employers."
White-Collar — The Betrayal of the Certificate
"Was I born at the wrong time? I studied myself half to death, and the job market was scorched earth by the time I got there."
The words of a young job-seeker. CPA — Certified Public Accountant. People who had cleared a single-digit pass-rate gate could not find employment. Among the 2025 cohort, only 338 of roughly 1,200 successful candidates had secured trainee positions by October — approximately 72 percent remained unemployed.7
Park Hyun-jin (a pseudonym) spent five years preparing. After graduating from university, he shuttled between cram schools and a goshiwon — a cramped study dormitory common in South Korea — failing three times before finally earning his certificate. The night the passing notification arrived on his phone, he called his mother. She wept. He believed he could finally do something with his life.
That belief collapsed within six months. Major accounting firms rejected him in the document-screening round, one after another. Mid-sized firms and small tax offices were no different. "I made it to the interview stage at two places. Both said they had been significantly reducing junior hires since adopting AI." More than half of the colleagues who passed in the same year were in a similar position. In the group chat of successful candidates, notices of rejection outnumbered news of employment.
"Senior colleagues told me that using AI saves more time and money than training a first-year trainee." Hwang Byeong-chan, a fifteen-year veteran accountant, was blunter: "Hiring as many junior accountants as we used to simply doesn't fit the current era."
IT headhunter Noh Sang-beom summarized the structure in a single sentence: "From a company's perspective, paying an AI 300,000 won a month for a year is far more cost-effective than hiring a junior developer at an annual salary of 50 million won."
Fifty million won a year versus 300,000 won a month. A 167-fold price gap. International data already show signs of wage inversion. OECD occupation-level wage analysis reveals a converging wage premium between clerical and administrative roles — high in AI-displacement exposure — and skilled trades with low exposure. This is a harbinger of a world where work done by hand becomes more valuable than work done by mind. Goldman Sachs (2023) estimated that 40 to 60 percent of accounting and data-processing tasks are automatable. That was the market Park Hyun-jin's certificate had delivered him into.
Park Hyun-jin now works the night shift at a convenience store. The certificate he spent five years earning sits in his wallet. "I don't know what to tell the juniors preparing for the next exam."
Global — Same Wave, Different Heights
Amazon internally estimated that it could replace up to 600,000 new hires with robots by 2033.8 Across the U.S. tech sector, mass layoffs continued through 2024 and 2025.910 In the same retail industry, Walmart chose the opposite path — investing in large-scale reskilling for its 2.1 million employees and declaring upskilling rather than downsizing.11 Safety nets are not determined by technology. They are determined by people.
Meanwhile, on a television set, a panelist was saying: "I asked companies where they most want to deploy AI, and they said call centers. Handling complaints is so emotionally taxing that if AI takes over... the people who used to take those calls would be free from all that stress."
Under studio lights, the celebration of "liberating agents from stress" aired at the same moment that, in reality, 240 agents were receiving layoff notices at lunchtime and could not eat. The distance between the two worlds was the width of a camera lens, and the gap in perception spanned decades.
Ku Gyo-jun, a professor at Korea University, identified the gap precisely: "When a new technology appears, people ask whether new jobs will emerge. The probability is high. But the timing will be at least twenty to thirty years from now. The most urgent problem, therefore, is the young people who will lose their jobs in the next ten to twenty years, during the period when the technology is being adopted."
Ten to twenty years. That is the gap in the safety net.
Yet this narrative has a counter-lens. AI does not only "replace" jobs. What history shows is that technology destroys, transforms, and creates occupations simultaneously. When the Luddites smashed weaving frames at the dawn of the Industrial Revolution, employment in the British textile industry actually surged — machines lowered production costs, demand exploded, and new occupations emerged to operate and maintain the machines. The 1990s fear that ATMs would eliminate bank tellers proved wrong. ATMs reduced the cost of operating a branch, so banks opened more branches, and the number of U.S. bank employees rose through the 2000s. In the AI field itself, job titles that did not exist five years ago — "prompt engineer," "AI trainer," "AI ethics consultant" — are already appearing. The World Economic Forum's 2025 report projected a net gain of 78 million jobs, reflecting this creation effect.12
The problem is not the aggregate. The problem is the time lag. The speed at which jobs vanish differs from the speed at which new ones appear. Employment did rise in the long run after the Industrial Revolution, but that "long run" spanned more than a generation. The suffering endured by that generation does not show up in aggregate statistics. As Professor Ku put it, the new jobs arriving in twenty to thirty years offer no comfort to those fired today. The optimism that AI will create jobs and the reality that people drown in the gap between are not contradictions — they are simultaneously true.
3. The 0.68 Percent Safety Net
A single number carries the weight of this chapter.
0.68 percent.
That is the benefit receipt rate for platform and freelance workers enrolled in employment insurance. Not the enrollment rate. The enrollment rate itself stands at roughly 46.4 percent (as of 2022) — many are enrolled. But the proportion who actually receive unemployment benefits is just 0.68 percent. Compared with the 6.24 percent receipt rate for standard wage workers, it is roughly one-ninth. The source is a report by the Working Citizens' Research Institute (일하는시민연구소) and Union Center (유니온센터), published in October 2024: "Improving the Social Safety Net Blind Spots for Precarious Workers Outside the System."
Why are they enrolled but unable to collect? Four barriers stand in the way. First, the irregular income inherent to platform work makes it difficult to meet the threshold of 800,000 won per month. Second, proving "involuntary separation" is a hurdle — when an algorithm deactivates an account, whether that constitutes a dismissal or a contract expiration is legally unclear. Third, platform companies refuse to issue confirmation letters, claiming "the platform bears no fault." Fourth, the sheer complexity of the administrative process leads applicants to give up.
According to a 2025 Human Rights Watch (HRW) report, six of seven major U.S. gig platforms used opaque wage-calculation algorithms.13 Workers learned their compensation only after completing a task. More than a third of gig workers had experienced account deactivation, and nearly half of those cases were later confirmed as "errors." An algorithm's mistake resulting in a firing — with a probability approaching fifty percent. No reason to assume South Korea is different.
Consider the scale of South Korea's platform labor force. Two figures exist, and the reason requires explanation. The Ministry of Employment and Labor (고용노동부) and the Korea Employment Information Service (한국고용정보원) 2021 survey on platform workers uses two definitions. Broad — anyone who obtains work through a platform (including side gigs and occasional participation): approximately 2.2 million people (8.5 percent of employed persons aged 15–69). Narrow — full-time workers whose tasks are directly assigned by an algorithm and who rely on platform income as their primary source: roughly 660,000 (2.6 percent) in the same year.14 By 2023, the narrow-definition figure had risen to 883,000, an 11.1 percent increase from the previous year's 795,000.15 The total non-standard workforce — delivery riders, special-employment workers, freelancers combined — exceeds 8.62 million.
The safety net stretched over those 8.62 million people is thinner than paper.
Article 24 of the current Labor Standards Act (근로기준법) specifies the requirements for managerial dismissal: urgent managerial necessity, efforts to avoid dismissal, rational criteria for selecting those to be let go, and notification plus consultation at least fifty days in advance. But these requirements apply to "workers" (근로자). Platform workers are legally classified as "independent contractors" (개인사업자). They stand outside the Labor Standards Act. No unfair-dismissal remedy, no automatic industrial-accident insurance, no severance pay.
On July 25, 2024, the Supreme Court (대법원) acted. It recognized a TADA app-based driver as a "worker" under the Labor Standards Act.16 It was the first Supreme Court ruling to look at substance rather than form. That same year, the EU passed the Platform Work Directive, establishing a presumption that platform workers are employees and shifting the burden of proof to companies. In South Korea, the reverse still holds — workers must prove their own employee status. In 2025, the Supreme Court issued additional rulings further relaxing the criteria for determining worker status, and from March of that year, legal and educational support for freelancers, non-regular workers, and platform laborers was expanded. These were significant steps forward. Yet for delivery riders, call-center agents, and freelance translators, each case requires individual litigation and individual verdicts. One case at a time, one person at a time.
In 2025, the government advanced the "Basic Act on the Rights of Working People" (일하는 사람의 권리에 관한 기본법) as its flagship labor bill. The act would grant eight categories of rights — safety, fair contracts, adequate compensation, access to social security, and more — to "all working people." Labor groups responded coldly: "Most provisions amount to declarative language — 'shall endeavor,' 'may recommend' — imposing no substantive obligations on employers."17
Rights were declared, but protection was absent.
4. Historical Precedent — Safety Nets Have Always Come After Blood Was Spilled
From the outset, this book has tracked a single pattern: institutions are slower than crises.
In Chapter 2, we saw the British Factory Act (1833). Sixty-four years had passed since Arkwright obtained his spinning-frame patent in 1769. During those sixty-four years, tens of thousands of children worked sixteen-hour days. In Chapter 3, we saw American financial regulation. From the onset of the Great Depression in 1929 to the Social Security Act (1935), six years elapsed — yet the regulatory vacuum that failed to prevent the 2008 financial crisis had formed over decades.
Return to the Rome of Chapter 1, and the gap stretches longer still. From the point when latifundia began displacing smallholders around 264 BC — the onset of the Punic Wars, which accelerated land consolidation — to Tiberius Gracchus's attempted reform (133 BC): 130 years. Add Gaius's grain law: 140 years.
Safety nets have always come after blood was spilled. Laid out in sequence:
Roman Republic: latifundia and the ruin of smallholders → Gaius's grain law (123 BC) → a delay of more than a century. British Industrial Revolution: child labor → the Factory Act (1833) → 64 years. The American Great Depression: mass unemployment → the Social Security Act (1935) → 6 years. The AI transition: structural job displacement → ? → ongoing.
The intervals are shrinking, and that could be cause for hope. But if they are shrinking because crises arrive faster, the speed of safety nets must accelerate in kind. It has not.
All five structural forces analyzed in Chapter 1 operate in this pattern — information asymmetry, incumbent capture, and ideological barriers are each present — but two are particularly intense.
The first is the gradualism of crisis. AI-driven displacement unfolds slowly. Translators thin out, customer-service agents thin out, data-entry clerks thin out — but each instance is not a dramatic event. The layoff of 240 call-center agents made the news, but the individual frustration of each unemployed CPA certificate holder does not. Just as smallholders were pushed out over a generation, AI displacement advances one occupation at a time, one office at a time. It has not yet registered clearly in macroeconomic employment indicators. The WEF projects that by 2030, 92 million jobs will be displaced and 170 million new ones created — a net gain of 78 million. Behind that optimistic aggregate, the turmoil of the transition period hides. Cumulative catastrophe that never crosses the detection threshold at any single point — a structure repeated in Rome, in Manchester, and on Wall Street.
The second is the cost of consensus. South Korea's universal employment insurance was declared in 2020. In 2021, coverage began for twelve categories of special-employment workers; in 2022, it was extended to platform workers. But actual enrollment and receipt rates remain dismal. Between expanding the scope of coverage and delivering substantive protection lies a vast gap called the cost of consensus. Platform companies refuse recognition as "employers." Negotiations over premium sharing stall. Just as a single tribune's veto blocked reform in Rome, the resistance of one stakeholder group can neutralize the effectiveness of an entire safety net.
"Are we once again in the stage 'before blood is spilled'?"
Recall Professor Ku's words. Twenty to thirty years for new jobs to emerge, ten to twenty years for existing ones to vanish. The decade in between is the gap in the safety net. What history tells us is that the gap is filled only after someone has bled enough.
5. Experiments in Other Countries
Can a safety net be woven before blood is spilled? Several experiments are underway.
Finland — The Question of Basic Income
In 2017, Finland launched the world's first national-level basic income experiment, paying 2,000 long-term unemployed individuals 560 euros per month for two years, regardless of whether they found work. The results overturned conventional wisdom. Participants worked an average of six more days per year than the control group, and their mental health improved significantly.18 The premise that "people won't work if you give them money" was wrong. The government did not extend the experiment — the political cost was too high — but the counterargument of "lost work ethic" lost its force in the face of this data.
Denmark — The Paradox of a Country Where Firing Is Easy
Denmark's "golden triangle" (flexicurity): dismissal is easy (flexibility), the unemployed receive up to 90 percent of their former wages for up to two years (income security), and the government is obligated to provide retraining and job placement (active labor market policy spending at approximately 2 percent of GDP, the highest in the OECD).1920 When all three interlock, dismissal becomes not a catastrophe but a transition. When the Odense Lindø shipyard closed in 2009 and approximately 2,600 workers lost their jobs, the EU Globalisation Adjustment Fund and Denmark's vocational training system supported their transition into energy technology and robotics. A robotics cluster now stands on the former shipyard site.
But will it hold for the AI transition? The knowledge work that AI displaces is work AI already does better, and the boundary of "what AI cannot yet do" — the destination of retraining — shifts every month. Moreover, the gap between Denmark's union membership rate (approximately 70 percent) and South Korea's (approximately 13 percent) means the same blueprint produces a net that bears a different weight.
6. The Mirror — The Second Flood
In Chapter 3, we spoke of water. When the first flood — subprime mortgages — surged in, there were no levees. Millions of households lost their homes. In the United States alone, more than six million foreclosure proceedings followed.
The structural parallels are striking. The product designers of Chapter 3 said that mortgage-backed securities (MBS) bundled tens of thousands of individual loans to diversify risk. The platform companies of Chapter 9 say that AI processes millions of individual inquiries to boost efficiency. In both cases, the system's beneficiaries and its cost-bearers were different people. The profits of financial innovation concentrated on Wall Street; the losses were distributed among mortgage borrowers. The productivity gains of AI innovation concentrate in platform companies; the transition costs are distributed among dismissed workers.
Brooksley Born warned a decade in advance, but Greenspan, Rubin, and Summers blocked her, and by the time her warning was proved right, six million households were already underwater. The Dodd-Frank Act was built as a levee after the fact, but it was useless for those already submerged. Greenspan said at the congressional hearing: "Yes, I found a flaw." The admission came only after the damage was done.
The pattern of Chapter 9 is the same. Indirect employment outsourced accountability. AI adoption created the incentive for workforce reduction. The termination notice arrived at lunchtime. Someday, a corporate executive may acknowledge that "there were shortcomings in the AI transition process." But when that admission comes, Kim Hyun-ju's lunch that day will not return. Park Hyun-jin's five years will not be restored. The account-restriction notification Driver Lee received at two a.m. has already cut into his income.
The most important commonality between Chapters 3 and 9 is that the victims participated in building the very structure that harmed them. Subprime borrowers made monthly interest payments that sustained the financial system's revenue base. Call-center agents corrected errors daily, generating the AI's training data. The reward for participation was exclusion. The second flood moves in the same way.
Now, the second flood is rising.
The first flood was called financial innovation. The second flood is called AI innovation. The first flood took homes. The second flood takes jobs. In both cases, the operative word was "efficiency." Efficient financial products, efficient automation. The fruits of efficiency concentrate among the few; the costs of efficiency are distributed among the many.
A KBS investigative program asked: "Is the arrival of AI a liberation of human labor, or an expulsion?"
The answer has not yet been decided. But if there is no safety net, the answer is already determined.
The Korea Institute for Industrial Economics and Trade (산업연구원, KIET) estimated that 3.27 million domestic jobs are at risk of AI displacement.21 According to analysis by the Stanford Digital Economy Lab, early-career employment for ages 22–25 in high-AI-exposure occupations in the United States declined by approximately 16 percent in relative terms.22 It is also worth noting the spread of "AI washing" — the practice of branding cost-cutting as "AI transformation" regardless of actual AI adoption. A survey of 1,006 global executives found that only 14 percent of companies were ready to deploy AI solutions, and only 11 percent had them in operation.23 The actual impact of AI may be overstated. But whether or not it is, the people who have been fired are already fired.
If there is no safety net, who will weave the net?
The Supreme Court's TADA ruling was a single thread. Finland's experiment, a fragment of a blueprint. Denmark's flexicurity, a fabric tested in a different climate. Walmart's upskilling, a patch woven voluntarily by the private sector. But there is still no net.
If there is a fragile hope, it is this: throughout history, safety nets have always arrived late, but they have never failed to arrive. Even in Rome, Gaius's grain law eventually passed. Even in Manchester, the Factory Act was eventually enforced. Even on Wall Street, the Dodd-Frank Act was eventually enacted. The question is not "will it come?" but "how many people will drown before it does?"
Kim Hyun-ju has not yet found a new job. If it rains tomorrow, Driver Lee will once again run the route assigned by the algorithm, rain streaking over his helmet. His account restriction was lifted, but next time it might not be. Park Hyun-jin's certificate sits in a drawer. For them, the new jobs arriving in twenty to thirty years are no comfort.
Meanwhile, in Silicon Valley, six-figure software engineers who have been laid off are opening food-delivery apps. The algorithm assigns them routes — built by the very kind of software they helped create. The structural irony has no visible end.
In the next chapter, we examine an attempt to regulate technology with technology. RegTech — regulatory technology. An experiment in which slow institutions use technology itself as a tool to keep pace with fast technology. From Singapore's AI Governance Framework to Australia's disastrous failure. If weaving a safety net is also a question of what material to weave it from, Chapter 10 examines that material.