This Time, the Loom Might Actually Win

Sarah Chen has worked as a translator for eleven years. English-Chinese and Chinese-English, specializing in technical documents and legal contracts. Her clients include two Fortune 500 companies and an international law firm. The work was stable, the pay decent, with a small raise each year.

In the spring of 2024, her first client notified her: from now on, the initial drafts of technical documents would be completed by AI, and she would only be needed for “human review.” Her workload was cut in half, and the rate slashed by 60%. She accepted reluctantly.

By autumn, even the review work disappeared. The client implemented an AI + automated quality check workflow, eliminating the human step. Her second client soon followed suit. The volume of contract translations from the law firm also plummeted—legal AI tools could now handle the multilingual versions of standard contract clauses.

In early 2025, Sarah opened Upwork to find new clients. Translation job postings were down nearly 20% compared to a year prior. The remaining postings had rates pushed below her cost line—because countless others were using AI to generate initial drafts and then undercutting on price. Competing with AI on price, she was bound to lose.

She wasn’t trying hard enough. She took a course in prompt engineering. She tried to reposition herself as an “AI Translation Quality Consultant.” But too few companies needed this role, and even fewer were willing to pay for it.

Eleven years of professional accumulation saw most of its value eaten away in twelve months by a $20/month tool.

Meanwhile, her LinkedIn feed was filled with inspirational articles on “How to Embrace Change in the AI Era,” citing the story of ATMs and bank tellers, telling her that history proves everything will be fine.

She wanted to ask: When exactly does that “fine” arrive? And before that, who pays the rent?

The ATM Story: You Only Heard the First Half

The ATM story is brought up in every discussion about AI and jobs. “ATMs appeared, yet the number of bank tellers increased. So don’t be afraid.”

After carefully dissecting the data, Paul Kedrosky pointed out a fact overlooked by most who retell this story: ATMs did indeed reduce the number of tellers needed per bank branch. Since 1985, the number of tellers per branch has been steadily declining. This is the true effect of ATMs—entirely consistent with the intuitive prediction of “machines replacing humans.”

So why didn’t total teller employment fall? Because during the same period, the U.S. passed the Riegle-Neal Act (1994), which lifted restrictions on banks establishing branches across state lines. Suddenly, banks could open branches nationwide, and the total number of branches surged. With more branches, even if each branch had fewer tellers, the total number was sustained.

What saved teller employment was not the “new demand” created by ATMs, but a regulatory reform completely unrelated to ATMs. Kedrosky’s conclusion is stark: “Unless you can identify a specific compensating force, employment survival may be coincidental, not mechanistic.”

The Loom story similarly does not withstand scrutiny.

Machines made cotton cloth cheaper → demand surged → the total number of textile workers increased. This causal chain is correct, but it omits a critical external condition: the surge in demand for Britain’s 19th-century textile industry came largely from the global market of its colonial empire—hundreds of millions of people in India, Africa, and the Americas became consumers of cheap British cotton. This was not a pure economic mechanism of “automation creating demand,” but market expansion backed by geopolitical and military power.

The employment expansion of the electrification era relied on the overall rise of a new consumer class (the middle class)—a one-time structural dividend jointly created by industrialization, urbanization, immigration, consumer credit, and the global landscape between the two World Wars. The Internet created the App economy, but it also created the gig economy—a form of employment characterized by instability, lack of benefits, and volatile income. Among the 64 million American “freelancers,” how many are truly voluntary, and how many are forced into “flexible employment” because they cannot find full-time jobs?

Every historical analogy used to comfort us has special, non-replicable external conditions behind it. Strip away these conditions, and how much confidence remains in the judgment that “technological revolution always creates more jobs”?

The most critical difference lies in speed. From Watt’s improved steam engine to full industrialization took about 80 years. From Edison’s power station to electrification covering 75% of manufacturing took nearly 50 years. From commercialization to full penetration, the Internet took about 20 years. From its launch to covering 100 million users, ChatGPT took 2 months.

A core reason past technological revolutions did not cause long-term mass unemployment was that the pace of substitution was slow enough to allow the labor market time to adjust—the speed at which old jobs disappeared roughly matched the speed at which new jobs emerged, giving workers a generation to transition.

If AI substitution is exponential, but the creation of new jobs remains linear—requiring adjustments to the education system, the formation of new industries, and the establishment of new consumption habits—the “friction period” in between is not a few years of pain, but the predicament of an entire generation.

No Retreat from General Cognitive Substitution

Every previous automation has replaced a specific, bounded task. ATMs replaced “counting money” and “deposits/withdrawals.” Spreadsheets replaced “manual calculation.” In each case, humans had a large area of “things machines can’t do” to fall back on.

AI is not replacing a process. It is replacing a capability—the ability to understand language, generate text, analyze data, write code, and perform reasoning. These are not links in a specific industry; they are the underlying infrastructure of almost all knowledge work.

When machines could only “count money,” people could retreat to “chatting with customers.” When machines can “count money,” “chat,” “analyze behavior,” “write copy,” and “generate reports,” where do people retreat to?

Daron Acemoglu, the 2024 Nobel laureate in Economics, provides a sobering estimate in his paper: AI’s contribution to U.S. GDP over the next 10 years will be only about 0.93%-1.16%. Far below Goldman Sachs’s prediction of 7%. His core argument: the cost-effectiveness of most AI application scenarios does not support large-scale deployment; only about 5% of work tasks can be effectively and economically automated by AI.

But here lies a counter-intuitive trap. If Acemoglu is right and AI’s economic benefits have been overestimated, then the optimistic narrative that “AI will create a flood of new jobs” loses its foundation—because the creation of new jobs depends on market expansion brought by AI; if the expansion is limited, so are the new jobs. But if the optimists are right and AI’s impact is indeed revolutionary, then the scale of displacement will also be far greater than they are willing to admit.

In either case, a large number of workers in the middle face an awkward situation: AI replacing them does not require “revolutionary” AI, just “good enough” AI. And “good enough” has already arrived.

Four Falling Knives

The First: The Collapse of the White-Collar Middle Tier.

In May 2025, Anthropic CEO Dario Amodei warned: AI could eliminate 50% of entry-level white-collar jobs within the next few years. A year later, data shows his direction was correct—top tech companies saw entry-level hiring fall 30-50% year-over-year. The “fresh graduate software engineer” position, which defined Silicon Valley for 20 years, is being quietly phased out: a senior engineer plus an AI coding assistant can now do the work that previously required a senior plus two juniors.

Major law firms have slowed hiring of first-year lawyers. The Big Four accounting firms are restructuring entry-level audit positions. Wall Street banks plan to cut about 200,000 jobs over the next 3-5 years, with cuts concentrated in junior analysts and back-office operations.

Optimists say these people will “upskill” to higher positions. But the pyramid is narrow at the top and wide at the bottom for a reason. A company needs 100 junior analysts to feed data but only 10 senior analysts to make decisions. When AI replaces the work of those 100 juniors, 100 senior positions will not magically appear.

The deeper issue is the talent pipeline. The junior employees laid off in 2026 are the missing mid-level backbone of 2030, and the non-existent executive candidates of 2035. Two Fortune 100 Chief Human Resources Officers have already realized this problem. But most companies are still viewing this through the lens of quarterly earnings reports; the pipeline breakage won’t become apparent for another three to five years.

The Second: The AI-fication of Freelancing.

Bloomberry analyzed 5 million Upwork job postings: After ChatGPT’s release, writing jobs decreased by 33%, translation by 19%, and customer service by 16. Translation hourly rates fell by over 20%. The mid-tier freelance market is experiencing a systemic collapse—small orders like logo design, social media management packages, and brand refreshes, which previously supported many freelancers, are now being handled by clients using AI themselves. The remaining orders concentrate at two ends: the low end is better handled directly by AI, and the high end requires genuine creative judgment, but such jobs were always scarce. Those in the middle are being squeezed out.

The Third: The One-Person Company is a Privilege, Not a Universal Good.

The most touching story in the optimistic narrative is the “one-person company”—using AI tools to start a business and earn a six-figure monthly income. But Indie Hackers data shows the median monthly income for independent founders is $3,000, annualizing to $36,000. The names earning $1M+ are visible due to survivorship bias.

A more fundamental issue is the barrier to entry. Those who can start a business with AI need English proficiency, technical intuition, product sense, market judgment, risk tolerance, and a solid grasp of the AI toolchain. Taken together, these conditions exclude the vast majority of workers facing AI substitution risk. The “one-person company” is an opportunity for young techies in Silicon Valley, but for a laid-off 45-year-old bank teller, it’s a story that has nothing to do with them.

The Fourth: Overcapacity.

Jevons Paradox assumes that the demand expansion from efficiency gains can absorb the growth in supply. But the problem in the AI era is: supply-side constraints have virtually disappeared. Previously, writing an article required an author half a day; now AI can write ten in ten minutes. Previously, developing an App required a team several months; now one person can finish it in a few days.

When everyone can create content, products, and services, has demand grown proportionally? No. A day still has 24 hours. Corporate budgets have not increased tenfold. The result is oversupply, falling prices, and cutthroat competition. “Everyone can do it” is not an expansion of jobs, but a dilution of income. This is already clearly visible in the freelance market—not that jobs have disappeared, but that the value of each job has been flattened.

The Other Half of the Story in the Data

The optimistic data doesn’t hold up under scrutiny.

“AI skill job wage premium of 56%”—this applies to those who already possess AI skills. For ordinary white-collar workers without AI skills, total hiring volume fell 11.3% year-over-year during the same period. On one side, the premium for AI-skilled workers is soaring; on the other, the market for non-AI skills is shrinking. The “average growth” in the data masks a polarization in distribution.

“Employment growth of 38% in AI-exposed industries”—the growth is mainly in AI engineers, machine learning specialists, and data scientists. Can a translator or junior customer service rep whose job was replaced by AI directly transition to become an AI engineer? BLS predicts data scientist job growth of 33.5% from 2024-2034, but in the same report, information entry clerk jobs are expected to decline by 35%. What’s growing is the peak of the pyramid; what’s shrinking is the base and the middle.

Where do those in the middle go? Brookings research offers an unsettling answer: many workers displaced by technology ultimately do not move into “higher-skilled” jobs, but slide into lower-paid service sector work. Factory workers replaced by robots did not become engineers; they became delivery drivers and security guards.

Now look at the company level. Klarna laid off people, found it was a mistake, and the CEO admitted “cost became the sole consideration, which was wrong,” and started rehiring. Block laid off 40% at once, cheered on by Wall Street, but analysts questioned the layoffs were too drastic. HBR’s survey reveals a startling fact: companies are laying off workers based on AI’s potential capabilities, not its actual performance—AI hasn’t truly proven it can replace these positions, but companies are laying off anyway. Robert Half survey: 29% of hiring managers have already reopened positions previously cut due to AI.

This is an absurd cycle of layoff-and-regret. In each cycle, those laid off bear all the costs, and these costs are almost invisible in the macro employment data.

Optimism is a Luxury, Preparation is a Necessity

Back to Sarah.

Not everyone replaced by AI can learn Python within six months, transform into an AI product manager, and then use AI to start a business earning a six-figure monthly income. This narrative exists, but it belongs to a very small group. For more people, the reality is: a decade of professional accumulation has been flattened by a $20/month tool, and the path forward through retraining is uncertain.

Brookings points out three structural difficulties with the retraining narrative. First, the number of high-skill positions available to upgrade into may simply be insufficient. Second, mid-career workers with family obligations cannot afford the time and income loss of retraining. Third, retraining programs often result in the absurd outcome of “training from one job about to be automated to another job about to be automated.”

“History proves everything will be fine” is a dangerous anesthetic. What history proves is: in the long run, technological revolutions create more jobs. But how long is “the long run”? If it’s 20 years, a 35-year-old laid off today will be 55 before new jobs appear. Mortgages, tuition fees, and medical bills won’t wait 20 years.

At the individual level, three things are worth doing.

Identify the parts of your work that AI truly cannot replace. Not “what AI hasn’t learned to do yet”—that’s just a matter of time. But what inherently requires human participation: taking responsibility for high-risk decisions, managing complex interpersonal relationships, and operations in the physical world that require physical presence. Focus your energy in these directions.

Don’t wait for your company to arrange your transition. Most companies’ perspective is quarterly earnings reports, not your career. Klarna laid off people, found it was a mistake, and rehired—in this cycle, it’s not the company that bears the cost, but the person who was laid off. Proactively build cross-industry skills and networks, and start exploring alternative paths before passive layoffs arrive.

Take the social safety net seriously. If the pace of transition truly exceeds the pace of retraining—current evidence increasingly points this way—then relying purely on individual effort is not enough. Skill subsidies, transitional income guarantees, structural adjustments to the education system—these are not “welfare” issues, but infrastructure for economic transition.

AI may ultimately create more jobs. The broad direction of history may indeed be optimistic. But “ultimately” and “may” are not answers for those losing their income today.

Sarah doesn’t need someone to tell her that “there are more tellers after ATMs appeared.” She needs someone to tell her how to pay this month’s rent.

Optimism is a luxury. Preparation is a necessity.

Appendix: Sources Cited in This Article

Rebuttal of the ATM Fable (Paul Kedrosky): https://paulkedrosky.com/ai-and-the-fable-of-the-atms/
Acemoglu “The Simple Macroeconomics of AI” (NBER / MIT): https://economics.mit.edu/news/daron-acemoglu-what-do-we-know-about-economics-ai
Dario Amodei Prediction + Entry-Level White-Collar Job Data: https://shawnkanungo.com/blog/dario-amodei-was-right-entry-level-white-collar-jobs-are-disappearing-fast
Analysis of 5 Million Upwork Jobs (Bloomberry): https://bloomberry.com/blog/i-analyzed-5m-freelancing-jobs-to-see-what-jobs-are-being-replaced-by-ai/
Limits of AI Retraining (Brookings): https://www.brookings.edu/articles/ai-labor-displacement-and-the-limits-of-worker-retraining/
Companies Laying Off Workers Based on AI Potential (HBR): https://hbr.org/2026/01/companies-are-laying-off-workers-because-of-ais-potential-not-its-performance
Analysis of Block’s 40% Layoffs (Forbes): https://www.forbes.com/sites/ronshevlin/2026/02/27/block-lays-off-40-of-staff-and-blames-it-on-ai-dont-buy-the-excuse/
Klarna AI Layoff Reversal (Forbes / Fast Company): https://www.digitalapplied.com/blog/klarna-reverses-ai-layoffs-replacing-700-workers-backfired
AI Job Creation / Layoff Statistics (Novoresume): https://novoresume.com/career-blog/ai-job-creation-statistics