Compensation Forces Don’t Appear on Their Own; You Have to Find Them

When cotton cloth becomes cheaper, people buy more clothes—this is intuitive. When opening bank branches becomes cheaper, banks open more branches—also intuitive. These two stories are repeatedly used to argue that “AI will not eliminate jobs”: efficiency improves → costs decrease → demand expands → employment increases.

But there’s a question skipped over here.

People need clothes; this is a clear, perceivable need. People need convenient banking services; this is also true. You can see the demand exists; it was just too expensive to fulfill before. AI makes “writing code” cheaper—and then what? Who is “buying” more code? AI makes “writing articles” cheaper—who is “buying” more articles? How much room is there for these demands to expand? Is there a ceiling?

Economics has a framework for this question called the Jevons Paradox. But most people who cite it omit a crucial detail: this paradox has conditions for it to hold. Not all efficiency improvements trigger a demand surge. Figuring out what those conditions are is key to judging whether AI will actually create new employment.

The more fundamental question is: even if new demand will emerge, how do you find it?

---

The Half Economists Can’t Predict

The Jevons Paradox has three necessary conditions: 1. Technological progress improves efficiency. 2. The efficiency improvement translates into lower consumer prices. 3. The price decrease must lead to a significant increase in the quantity demanded—i.e., demand elasticity must be sufficiently high.

The third condition is the crux of the entire debate.

High demand elasticity means that a small price drop leads to a large increase in purchases. Cotton cloth was a high-elasticity good—poor people in the 18th century didn’t not want to wear cotton; they couldn’t afford it. When prices fell, demand exploded. Banking services were the same—people in remote areas didn’t not want to save money; they lacked branches. Once branches opened, customers flooded in.

But not everything has high demand elasticity.

You only need to translate an article once. You won’t translate the same article ten times just because translation gets cheaper. A product needs one set of copy, not a thousand. A function only needs to be coded once. In these scenarios, efficiency improvement doesn’t lead to a “demand surge” but rather “the same amount of demand is handled with fewer people”—which is standard labor substitution. The Jevons Paradox doesn’t trigger.

So the question becomes: in the areas where AI lowers costs, is there more “massive, previously suppressed demand,” or more “demand that’s basically saturated, where efficiency gains directly translate into workforce reduction”?

MIT’s Erik Brynjolfsson believes fields like radiology, translation, and programming meet the three conditions—because there truly is a lot of “demand that was previously desired but too expensive to pursue” (previously, it wasn’t worth double-checking every X-ray; now AI can do it all. Previously, documents in small languages weren’t worth translating; now they all can be). Acemoglu, however, thinks only about 5% of tasks truly meet the conditions; in most scenarios, demand elasticity isn’t that high.

Who’s right? The honest answer is: nobody knows.

This isn’t an evasion. It’s a structural blind spot in economics.

Look back at history. When Edison built the first power plant in 1882, no economist predicted that “refrigerator” would become a product or that “home appliance salesperson” would become a profession. When the internet commercialized in 1995, Newsweek was still mocking the internet for having no economic value—no one predicted that “search engine optimizer” or “app developer” would become livelihoods for millions. In 2020, the term “Prompt Engineer” didn’t exist in any career guide.

Schumpeter called this unpredictability “creative destruction.” His core insight was: new demand comes from “new combinations” (neue Kombinationen), not linear extensions of existing demand. You can’t derive “people will need automobiles” from “people need horse-drawn carriages,” let alone derive “people will need auto insurance, driving tests, gas stations, roadside assistance.” Every new technology spawns a set of completely new demands that were unimaginable before the technology existed.

Economics can elegantly explain after the fact why demand expanded. But beforehand, it can hardly tell you “where exactly it will expand.”

This is both bad news and good news. The bad news is: no one can give you a “list of future jobs AI will create”—anyone who claims they can is lying to you. The good news is: precisely because it’s unpredictable, everyone willing to explore has the opportunity to be the first to discover it.

---

Will Demand Still Follow When Supply is Unlimited?

Even if you accept the general direction that “new demand will emerge,” there’s a sharper critique to address.

Past efficiency improvements had a built-in brake: supply expansion requires time and capital. No matter how fast the loom, you still need to build factories, buy cotton, hire workers, and wait for shipping. This physical constraint meant supply was released gradually—demand had time to grow in parallel.

In the AI era, the supply side has almost no such brake. An article? AI writes ten in ten minutes. A logo? Midjourney generates a hundred in three minutes. An app? One person with Cursor can do it in days. Marginal production cost approaches zero; supply can explode to near-infinity instantly.

When supply is infinite and demand is finite, what happens? Prices drop to near zero. This is already happening in information products—blog articles, AI-generated images, short videos have a “price” of zero (monetized through attention and ads, not direct charges).

Herbert Simon pointed out this problem in 1971: “A wealth of information creates a poverty of attention.” Humans only have 24 hours a day. AI can produce infinite content, but human consumption capacity has a ceiling. This is a hard constraint—it means the Jevons Paradox may fail for “attention-intensive products.”

But here’s a key distinction.

The attention constraint mainly limits information products aimed at end consumers—articles, videos, images, music. You won’t watch 48 hours of video a day just because there’s more content.

But for another major category of demand—productivity tools, B2B services, infrastructure—the constraint isn’t attention, but budgets and organizational inertia. The elasticity for this type of demand might be far higher than people intuitively think.

For example. Previously, a small-to-medium enterprise (SME) doing customer behavior analysis needed to hire a data analyst (annual salary $60K+). Most SMEs therefore didn’t do it at all. AI has reduced this cost to $200/month. There are tens of millions of SMEs globally. The demand elasticity here is extremely high—not “existing customers buy more,” but “tens of millions of buyers previously completely outside the market suddenly enter.”

Similar logic applies to: - Personalized education (previously only rich kids had tutors; AI gives every student one) - Legal services (previously freelancers couldn’t afford lawyers; AI legal tools let them in) - Psychological counseling (previously avoided due to cost and stigma; AI solves both privacy and cost issues simultaneously) - Multilingual services (1.2 billion people have never enjoyed quality digital services in their native language)

These are not attention-consuming demands like “more people reading more articles.” These are demands for substantially improving life and productivity; their ceiling is far higher than content consumption.

World Bank data shows: in emerging markets, 86% of EdTech companies have integrated AI; AgTech (agricultural technology) 70%; Fintech 54%. These aren’t Silicon Valley toys. These are basic services covering billions of people.

Whether the Jevons Paradox holds in the AI era depends on which type of demand you’re looking at. If you’re staring at “AI writing articles,” you’ll see oversupply and price collapse. If you look at “AI turning previously rich-only services into something everyone can use,” you’ll see an almost infinite demand pool waiting to be tapped.

Both things are happening simultaneously. Pessimists and optimists are looking at different sides of the same coin.

---

Three Methods for Discovering Compensation Forces

Returning to Kedrosky’s quote: “Unless you can identify specific compensation forces, job survival may be coincidence, not mechanism.”

Historically, four compensation forces have appeared: geographical market expansion (colonial empires, globalization), the rise of new consumer classes (middle class), institutional change (deregulation of banking), and the emergence of entirely new demand categories (home appliances, app economy).

In the AI era, all four forces could reappear. But the key is not to wait for them to emerge on their own—but to actively find them. Here are three specific methods.

Method 1: Globalization × AI = Releasing Suppressed Demand

Your users’ judgment is correct: globalization will accelerate the emergence of compensation forces.

The logic is simple. There are 1.2 billion people globally who have never enjoyed personalized education services—not because they don’t need them, but because human costs were previously too high for anyone to afford. AI pushes the marginal cost of services to near zero. The remaining problem is localization: language adaptation, cultural adjustment, regulatory compliance, channel building. This “bridging layer” work still requires humans, and specifically requires many people who understand the local market.

The Africa AI Education Summit (November 2025, Nairobi) brought together 100+ developers and policymakers; the core discussion was: how to make AI education tools truly adapt to the languages and teaching scenarios of Sub-Saharan Africa. Behind the adaptation for each language is a batch of new jobs in content moderation, instructional design, and localization testing.

Specific operational path: - Choose a service category that was “previously too expensive to offer” (education, law, health, finance). - Choose a market with “demand but lacking supply” (Tier 2/3 cities in Southeast Asia, South Asia, Africa, Latin America). - Use AI to bring service costs down to a level acceptable to the target market. - Your “human work” is: understanding the real needs of that market, doing localization, building trust.

This isn’t the Silicon Valley story of “using AI to start a business.” This is the globalization story of “using AI to enter markets previously inaccessible.” Among the 30.4 million U.S. solopreneurs, more and more are succeeding not by building another SaaS for English users, but by packaging AI tools into solutions for a specific vertical market, serving a group of users who previously had no choice.

Method 2: Start from Jobs to Be Done, Not from Technology

Most “failed AI startup” cases share a common trait: founders got the AI hammer first, then went around looking for nails. Their starting point is “what can AI do,” not “what do people need.”

Clayton Christensen’s Jobs to Be Done framework provides a reverse approach:

Don’t ask “what can AI do”; ask “in what situations do people want to complete what task, and what obstacles do they face?”

Specific method: 1. Observe, don’t survey. Watch what people are doing, where they get stuck, what clunky workarounds they use. Reddit, Zhihu, industry forums are full of such “painful workarounds.” 2. Define the Job. Format: “When I am in [a certain situation], I want to [complete a certain task] so that [I achieve a certain result].” Example: “When I am a non-native English-speaking small business owner, I want to write a proper business email to an overseas client so that I don’t lose an order due to language issues.” 3. Evaluate whether AI can significantly reduce the cost of this Job. If yes, you’ve found a potential high-elasticity demand. 4. Validate scale: How many people face the same situation? How much money/time are they currently spending on workarounds?

The advantage of this method is: Jobs are relatively stable; technology changes. The Job “wanting to write proper emails to overseas clients” won’t disappear because of technology changes. It has existed for decades; only the method of completing it changes. Services built around stable Jobs are more risk-resistant than products built around popular AI technologies.

The PwC report says “design skills have surpassed pure technical skills as the most sought-after skill by employers.” The “design skills” here are essentially JTBD skills: understanding what users want to do in what situations, then planning a path from the current state to the goal state.

Method 3: Use AI Itself to Discover Demand

This is the concrete expansion of your intuition to “use AI to dig.”

AI is not good at judging “what is worth doing,” but it is extremely good at “large-scale scanning of signals.” Combine these two, and you have an efficient demand discovery engine:

Step 1: AI does signal scanning. - Use AI to bulk-crawl forums, communities, product reviews, and social media for the target industry. - Have AI perform sentiment analysis and pain-point clustering: “What are the three things users in this industry complain about most frequently? What unsatisfactory workarounds are they using?” - One person + AI can complete a preliminary scan in 48 hours that a traditional market research firm would take two weeks for.

Step 2: Humans make judgments. - AI gives you a hundred “signals.” Your job is to filter out three to five “problems truly worth solving.” - Judgment criteria: Does the problem appear frequently? What cost (time/money/emotional) are users currently paying for workarounds? Can AI reduce the solution cost by an order of magnitude? - This step cannot be done by AI. It requires industry experience, business intuition, and understanding of people.

Step 3: AI does rapid validation. - Once a direction is chosen, use AI to create a Minimum Viable Product (MVP) in 48 hours. - Put it in front of target users. It doesn’t need to be perfect—just “good enough for users to be willing to try it and give feedback.” - User feedback determines the next step: double down or change direction. The cost is extremely low; the number of trials can be high.

This cycle—AI scanning → Human judgment → AI validation → User feedback → Iteration—is the core operating system for “discovering compensation forces” in the AI era. It doesn’t rely on luck or talent; it relies on discipline and repetition.

---

How Large a Demand Pool Can Globalization Actually Release?

Globalization is the biggest candidate source for compensation forces. But it’s not unlimited.

Positive signals are strong. The Indian EdTech market grew from $2B in 2020 to $6B in 2023, projected at $30B+ by 2030. African mobile financial users grew from a few million in 2012 to hundreds of millions in 2024. Latin American SaaS grew by 35%+ year-over-year. Southeast Asian AI applications are covering everything from agriculture to healthcare.

The common characteristic of these markets is: massive population × long-suppressed demand × AI is bringing service costs down to affordable levels. In India and Africa alone, there are over 2 billion people who have never enjoyed basic knowledge services that developed countries take for granted (personalized education, basic legal consultation, mental health support, financial planning).

But the negative constraints are also very real.

Geopolitical fragmentation is making globalization harder. Data sovereignty regulations (GDPR, China’s data export management) limit the efficiency of AI services operating across borders. Language and cultural barriers are harder to overcome than technical barriers. Regulatory fragmentation makes the cost of entering different markets with the same product rise significantly.

The net judgment is: the demand pool that globalization can release is indeed huge, but it won’t be a “one product for the world” scenario like the internet era. The more likely model is: each regional market will require deeply localized solutions. This is actually good news—because “localization adaptation” is a type of work that highly relies on human judgment, which AI itself doesn’t do well.

For individuals, the strategy is clear: - If you have language and cultural understanding of a non-English market, you possess an advantage that AI cannot currently replace. - Use AI for the product and technology layer; use human judgment for market selection and localization. - The most competitive market is always the English market. The more niche, the more vertical, the more culturally sensitive the market, the greater the leverage of the AI tool + local knowledge combination.

---

Compensation Forces Don’t Appear on Their Own; You Have to Find Them

Returning to the opening question. AI makes writing code cheaper—who is “buying” more code?

The answer is: those who previously didn’t even know they needed code. A small shop owner in an Indian Tier 2 city doesn’t know he can have an inventory management system. A community clinic in Africa doesn’t know it can have an AI-assisted preliminary screening tool. A freelancer in Latin America doesn’t know he can have an automated tax assistant.

These demands exist today, but they are latent. They won’t jump out and say “here’s a new job.” They need to be discovered, defined, and translated into specific products and services.

This is precisely the substance of the capability to “systematically understand user needs, define problem boundaries, and plan solutions.” It’s not some abstract “soft skill”—it’s a set of methodologies that can be practiced:

1. Start from the situation. Don’t ask “what can AI do”; ask “who is stuck on what problem in what situation?” Every complaint post in a forum, every rant in an industry group, every comment saying “I’d pay someone to solve this” is a demand signal.
2. Use JTBD to define problem boundaries. Transform vague pain into precise Job descriptions. “Customer management is too cumbersome” is not a solvable problem. “When I have 50+ overseas clients across three time zones, I want to automatically send a personalized greeting on each of their birthdays to maintain relationships without spending half an hour of my day each time”—this is.
3. Use AI for low-cost validation. Make an MVP in 48 hours. If it doesn’t work, switch. If it works, double down. Traditional startups need three months and a hundred thousand dollars to validate an idea. The AI era needs two days and three hundred dollars. The number of validations can be 50 times higher.
4. Seek blue oceans in the long tail of globalization. The English market is a red ocean. Small languages × vertical industries × AI tools = an almost uncontested blue ocean. Your local knowledge is the moat.

Kedrosky is right: unless you can identify specific compensation forces, job survival may be coincidence.

But what he didn’t say is: compensation forces don’t fall from the sky. The middle class in the electrification era didn’t just naturally emerge—it was “created” by entrepreneurs’ products (Ford’s Model T, GE’s refrigerators) and policymakers’ systems (labor laws, consumer credit). The app economy didn’t automatically grow from the internet—it was “built” by Jobs’ App Store, Google’s Android ecosystem, and millions of developers.

The compensation forces of the AI era won’t appear on their own either. Someone needs to find those latent demands, define them, use AI to bring service costs down, and deliver products into the hands of people who never had a choice.

This task is both the source of the next generation’s employment and a capability—a capability that can be practiced, mastered, and leveraged.

Economics can’t calculate this account. But you don’t need to wait for economics to finish calculating before you act. What you need is a discover-validate-deliver cycle, and low enough trial costs to let you run more cycles than others.

AI is what brings the cost of trial and error to near zero. The only scarce resource left is your judgment.

---

Appendix: Sources Cited in This Article

• Necessary conditions for the Jevons Paradox (Wikipedia / economics literature): https://en.wikipedia.org/wiki/Jevons_paradox
• Herbert Simon’s attention economy theory (1971) and subsequent research: https://www.sciencedirect.com/science/article/pii/S0016328723001477
• Schumpeter’s creative destruction (Econlib): https://www.econlib.org/library/Enc/CreativeDestruction.html
• Jobs to Be Done framework (Tony Ulwick / Christensen): https://jobs-to-be-done.com/jobs-to-be-done-a-framework-for-customer-needs-c883cbf61c90
• AI diffusion in emerging markets (World Bank): https://blogs.worldbank.org/en/psd/how-ai-travels–diffusion-among-firms-in-emerging-markets
• Africa AI Education Summit (World Bank): https://blogs.worldbank.org/en/education/the-future-is-africa–shaping-ai-enabled-edtech-for-skilling-the
• Refuting the ATM fable (Paul Kedrosky): https://paulkedrosky.com/ai-and-the-fable-of-the-atms/
• Acemoglu’s AI productivity estimates (MIT / NBER): https://economics.mit.edu/news/daron-acemoglu-what-do-we-know-about-economics-ai
• PwC 2025 Global AI Jobs Barometer: https://www.pwc.com/gx/en/news-room/press-releases/2025/ai-linked-to-a-fourfold-increase-in-productivity-growth.html
• Erik Brynjolfsson on the Jevons Paradox and AI (cited via Wikipedia)