The author of this article is Chen Yuyu, a professor at Peking University's Guanghua School of Management and director of the Institute of Economic Policy Research at Peking University.

Today, AI can translate a million-word English text into Chinese in just a few dozen minutes. Yet, surprisingly, this hasn't triggered an explosion in translation and publishing. Bookstores haven't seen a thousandfold increase in translated works, nor have readers suddenly gained a thousand times more reading time. This reality reminds us: economic growth is driven not by the fastest machines but by the slowest links—humans, institutions, and markets.

The translator's paradox offers an excellent lens through which to view the AI era. It reveals that a surge in efficiency for a single task does not necessarily lead to a surge in the quantity of final products; exponential progress in technical capabilities does not mean macroeconomic growth rates will skyrocket. AI can rapidly reduce the cost of certain cognitive tasks but cannot simultaneously eliminate challenges like copyright, editing, accountability, trust, distribution channels, consumer attention, and market risks.

In a previous article, I attempted to clarify one point: the true significance of AI lies beyond the old industrial framework and cannot be reduced to mere "efficiency gains" or "workforce reductions." If we focus solely on whether a department can hire fewer people, whether a process can be sped up by minutes, or whether a job will be replaced by machines, we remain trapped in the coordinates of the old world, imagining a technology meant to redefine those coordinates.

The deeper meaning of AI lies in its ability to expand the feasible set of human production and consumption. Activities that were once impossible, unimaginable, unaffordable, or unorganizable are now within reach. While replaced jobs have names, most newly created activities do not yet. The old ledger lists "translators," "copywriters," "customer service," and "programmers," but the new world will give rise to forms of learning, companionship, healthcare, aesthetics, and organization that we currently lack names for.

This is the most fundamental wisdom of price theory: growth is never about scaling up quantities from a predetermined menu but about expanding the feasible set, rearranging relative prices, reorganizing activity boundaries, and engaging in continuous trial-and-error within localized knowledge.

However, acknowledging that AI expands the boundaries of the feasible set does not mean embracing technological utopianism. In recent years, a popular imagination has emerged: since model capabilities can climb exponentially—with parameters, computing power, reasoning, and automation all rising steadily—macroeconomic growth rates will follow suit in a straight upward trajectory. Some even envision AI propelling advanced economies from around 2% annual growth to 10%, 20%, or even more astonishing rates.

This reflects both a worship of technological capability and a forgetting of economics. The real world is not governed by a single production function. An economy is composed of property rights, organizations, trust, accountability, physical bodies, families, education, health, institutions, and time. If one link speeds up, the entire system does not necessarily follow; if one task becomes free, the final product does not necessarily become free; if one capability explodes, the macroeconomic total does not necessarily explode.

Thus, one of the most critical questions in the AI era is no longer "Can machines do it?" but "Can the entire social system absorb, reorganize, price, and diffuse it?" Why, despite such intense micro-level reorganization, do macroeconomic data remain calm? Why does technology seem to follow an exponential curve while economic aggregates are still pulled back to earth by some "gravity"? And why might the sectors truly expanding over the next thirty years be neither traditional commodity production nor narrowly defined digital production but the reproduction of humans themselves?

Before proceeding, we must clearly define our subject of discussion. What follows—whether "moderate growth" or "the gravity of 2%"—refers to leading economies at the technological frontier, with the United States being the most typical example today. Growth in frontier economies must come from forging new paths in uncharted territory, not from retracing familiar routes already taken by others. For economies still catching up, high growth rates are possible for extended periods because their growth stems from convergence and diffusion, not frontier advancement. That is a topic for another article and will not be discussed here.

Returning to the frontier: the computer age once saw the famous "Solow Paradox," where computers were ubiquitous but invisible in productivity statistics. Today, AI stands at a similar crossroads. Wafer-scale clusters, foundational models, startup valuations, capital expenditures, chip demand, and data center construction are all surging; yet macroeconomic total factor productivity has not risen to the same heights.

Much of the so-called AI boom at present manifests first as capital deepening. Firms buy more chips, build more data centers, hire more engineers, and incur higher cloud costs. This naturally sparks investment frenzies and reshapes capital market expectations of future cash flows. However, capital deepening does not equate to a fundamental leap in total factor productivity. While piling on more capital can raise local output, a true productivity revolution must manifest as stable increases in output from the same labor, capital, and organization.

From steam engines, electricity, and internal combustion engines to computers and the internet, the diffusion of general-purpose technologies has never been instantaneous. Technological invention is only the first step; what follows is a much longer journey: processes must be rewritten, institutions adapted, human capital retrained, legal responsibilities redefined, industry standards gradually formed, and consumer habits slowly changed. Initially, a new technology may even suppress observed productivity as society expends vast resources on trial-and-error, migration, and reorganization.

This is the "gravity of 2%" in macroeconomics.

Here, "2%" is not a mysterious constant but a historical observation: frontier economies rarely sustainably break free from moderate growth over extended periods. Technological revolutions can create local surges, reshape industrial landscapes, build vast enterprises and fortunes, and even accelerate productivity in some years; but they rarely enable entire frontier economies to permanently escape the real-world constraints imposed by education, health, organizations, law, families, cities, and physical bodies.

AI may slightly raise future growth rates compared to the past. If the U.S. can rise from 2% to 3% or even touch 4% in some phases over the next thirty years, that would already be a monumental historical shift. But to imagine such changes as twentyfold accelerations is to confuse model capability curves with macroeconomic curves.

Why is this the case? A critical reason is that Baumol's cost disease will reemerge in new forms in the AI era.

When AI rapidly drives down the marginal costs of cognitive tasks like text processing, logic, coding, retrieval, translation, calculation, and image generation, sectors that cannot be fully replaced by algorithms will become more expensive. Deep healthcare, psychological companionship, personality development in education, childcare, eldercare, organizational leadership, public responsibility, complex negotiations, aesthetic judgment, and trust endorsements are not merely information processing. They involve physical presence, emotional resonance, social commitments, and ultimately, human accountability.

In a highly automated economy, what becomes truly expensive is no longer computing power but humans themselves. For easily automatable tasks, relative prices tend to fall; for tasks that resist automation yet see no decline in demand, relative prices tend to rise. As a result, high-friction sectors account for a larger share of economic spending, acting as a drag on macroeconomic growth.

This is not a failure of technology but a manifestation of general equilibrium.

Viewing translation and publishing through the lens of production functions makes the issue clearer.

On the technological side, AI translation is near-miraculous. Today, translating a million-word English work into Chinese may take machines just a few dozen minutes; in the past, a skilled translator might need two to three years. For the single task of "text conversion," efficiency gains are not 20% or 200% but thousandfold or ten-thousandfold.

Yet in reality, we do not see an explosive surge in translated books. The number of high-quality books translated from English, French, German, and Japanese into Chinese has not multiplied by a thousand due to AI.

The reason is that a book's publication is never just about "translation" but a Leontief-style production process with fixed proportions. The final product requires many complementary elements to align simultaneously: topic selection, copyright acquisition, contracting, translation, proofreading, editing, review, design, distribution, marketing, channel management, reader targeting, academic endorsement, legal liability, and assumption of market risks. AI eliminates or drastically reduces costs in one link, but not the others.

If a production system has fixed proportions, its efficiency depends not on the fastest link but on the slowest bottleneck. The faster AI translates, the more glaring the bottlenecks become.

First is the friction of property rights and compliance. Cross-border copyright negotiations do not automatically succeed just because machines translate faster. Publishers must still locate rights holders, negotiate prices, sign contracts, and handle scope of authorization, electronic rights, derivative rights, regional restrictions, and legal liabilities. A text can be translated in half an hour, but a copyright contract may take six months to negotiate.

Second is the gatekeeping of trust and quality. Readers do not buy mere Chinese sentences but a trustworthy text. Who guarantees translation accuracy? Who handles concepts, contexts, terminology, cultural adaptation, and the author's style? Who bears reputational damage if errors occur? AI can generate drafts, but final quality still relies on experts, editors, and translators for endorsement—the stronger the machine, the more valuable human responsibility becomes.

Third is the discovery of market risks. Whether a book is worth translating is never determined by machines. Whether readers want it, whether the market can accommodate it, whether channels will promote it, whether critics will discuss it, and whether schools, media, and knowledge communities will recognize it are all highly uncertain discovery processes. AI can reduce production costs but not eliminate demand uncertainty.

Fourth are the constraints of attention and time. Even if all books were translated instantaneously, readers would not gain a thousandfold more reading time. The bottleneck for knowledge products often lies not in production but in consumption; human attention, patience, comprehension, and cognitive structures remain the ultimate scarcities.

This is the so-called "translator's paradox": translation as a technical task has been crushed, but translation and publishing as a social production process have not exploded in tandem.

The point of this example is that AI's economic impact cannot focus solely on single-task efficiency. Economics cares about final products, systemic complementarities, and general equilibrium. A task's marginal cost approaching zero does not mean an industry's marginal cost approaches zero. Machines may race ahead, but social systems must still navigate nodes like institutions, organizations, and humans.

This holds true for material production, digital production, healthcare, education, culture, law, finance, research, and governance alike.

In the short term, AI's primary impact will not be an explosion in macroeconomic aggregates but an intense reorganization of production factors.

Workflows within firms will be rewritten. Jobs will vanish and transform. Many roles reliant on rule-following, text processing, standardized analysis, and exam-oriented cognition will rapidly devalue; meanwhile, new activities, processes, and occupations lack stable names. This phase will be rife with frictional costs: employees must retrain, organizations must redivide labor, management must redefine responsibilities, laws must redraw boundaries, and consumers must relearn trust.

Thus, in the short term, we may witness a paradoxical combination: micro-level upheaval alongside macro-level calm. Everyone in firms feels AI has transformed their work, yet productivity growth in statistics remains unimpressive. This is not a contradiction but a typical state of transition—technological dividends are quietly offset by reorganization costs.

In the medium term, the most profound changes will occur in the structure of human capabilities.

For decades, modern education and labor markets have rewarded a specific set of abilities: exam performance, rule-following, text comprehension, standardized calculation, organizational advancement, and stable execution. AI first strikes at this very set. Any capability that can be clearly described, rule-based, covered by training data, or simulated by language models will experience a relative price decline.

So what will become more valuable?

Health, physical strength, appearance, expressiveness, charisma, leadership, risk-taking ability, aesthetic judgment, empathy, psychological resilience, sense of responsibility, trustworthiness, physical presence, organizational mobilization skills, and cross-disciplinary creativity—economics has often categorized these attributes as “non-cognitive abilities” or “soft skills.” However, in the AI era, these qualities are becoming increasingly “hard” and essential. This is because as machines take on more cognitive tasks, human differences will increasingly revolve around physicality, emotion, character, style, trust, and organizational capacity.

This observation leads to a broader conclusion: over the next three decades, the largest and most central sector of production in society may well be “human self-enhancement.”

“Human self-enhancement” does not refer to narrowly defined education or traditional human capital investment. Rather than training humans as mere supplements to machines, it involves shaping individuals into more complete, vibrant, trustworthy, and creative beings.

If we examine the reproduction of human beings themselves, at least three layers emerge. The first is biological reproduction, encompassing health, physical strength, nutrition, lifespan, chronic disease management, and child development. The second is psychological and personality reproduction, including resilience, self-control, responsibility, trustworthiness, and emotional stability. The third is social capability reproduction, involving expressiveness, aesthetics, leadership, charisma, cooperation, and organizational mobilization. As AI reduces the costs of standardized cognitive tasks, the “shadow prices”—or implicit value—of these three categories will rise accordingly.

Consequently, early childhood development, family companionship, sports training, nutrition, healthcare, mental health, aesthetic education, expressiveness training, social skills development, personality cultivation, risk education, leadership development, elderly care, chronic disease management, intimate relationships, community life, and spiritual order will increasingly resemble not peripheral consumption but the core productive activities of the future economy.

In the industrial age, the largest productive sector was material production; in the information age, digital and cognitive processing expanded rapidly; in the AI age, when ordinary cognitive tasks are automated on a large scale, the area with the highest marginal returns will shift back to human beings themselves.

This may sound like a humanist slogan, but it is actually a straightforward inference from price theory. Resources, efforts, and talents will flow toward wherever abilities become relatively scarce and expensive. AI makes certain cognitive labor cheap, so the attributes of humans that cannot be automated, compressed, or replicated become valuable. The focus of economic growth may thus gradually shift from “producing more things” to “cultivating better people.”

If this judgment holds, the traditional GDP accounting system will face increasing pressure.

The current national economic accounting system is a product of the industrial age. It excels at recording commodities, services, wages, investments, and market transactions but struggles to capture human development within families, non-monetized experiences, improvements in personality capabilities, enhancements in health quality, the accumulation of psychological resilience, and the formation of social trust. Many truly critical productive activities either occur outside market transactions or are underestimated as mere consumption expenditures.

For example, a family investing substantial time in reading, exercising, self-expression, socializing, and exploring with their children is barely reflected in traditional GDP. Yet, from the perspective of the future economy, this may represent the production of the scarcest assets. If a society systematically enhances children’s health, psychological resilience, expressive and creative capacities, the future productive capacity it creates will far exceed many short-term commodity transactions fully recorded in the books.

More subtly, as AI drives down the marginal costs of traditional digital goods, textual products, and some services, the economic scale under old metrics may even be underestimated or distorted. A society’s true welfare may have risen sharply, yet monetary transaction growth remains limited; it may invest more resources in human development, only to have these counted as consumption rather than investment.

For this reason, over the next thirty years, the national economic accounting system must inevitably expand its boundaries to include human capabilities. Education, health, psychology, family, caregiving, sports, aesthetics, and personality development should no longer be treated merely as welfare or consumption sectors but understood as sectors where humans reproduce themselves.

However, in this domain, price accounting becomes exceptionally difficult. A “healthier, more expressive, more responsible, and more risk-taking” person cannot be precisely priced by a centralized algorithm like a machine, software, or a ton of steel. Human development is highly heterogeneous, its value often only becoming apparent years later, and is shaped by families, communities, cultures, and institutional environments.

For this reason, the cost approach will regain importance.

When outputs are difficult to price directly, measuring the true resource inputs a system expends is a pragmatic economic method. We can record how much time, labor, capital, and organization a society invests in child development, educational quality, mental health, sports, chronic disease management, elderly care, family support, and community building. Future economic growth will largely manifest as increased systemic investment in human development.

Such growth may accelerate macroeconomic performance somewhat compared to the past. Human self-reproduction has vast expansionary potential, perfectly positioned to absorb resources released by AI. Yet it will certainly not soar vertically as technological utopians imagine. The reason is simple: human physical development, emotional resonance, trust formation, personality shaping, and intergenerational reproduction all follow incompressible biological and sociological cycles.

A child cannot mature into a complete human in three months. A person’s character cannot be forged through a single model upgrade. A family’s production function cannot be rewritten by administrative decree. A society’s trust and aesthetics cannot be uniformly generated by a central server.

AI can accelerate tools but cannot eliminate growth.

One of the greatest dangers in the AI age is a new impulse toward centralized planning. Because AI seems omniscient—large models can process vast information, algorithms can generate plans, predict behaviors, assess performance, allocate resources—many people will develop a new utopian fantasy: since markets have friction, since humans err, since machines are smarter, let a central system decide resource allocation, let algorithms plan industries, let platforms dictate education, let metrics define excellence, and let administrative power shape the “future human.”

This is a dangerous, fatal conceit.

The more we enter an era of “elevating humanity itself,” the less we can rely on centralized planning. Human development depends critically on local knowledge. Every child’s talents differ, every family’s constraints differ, every community’s culture differs, every individual’s psychological structure differs, and every market opportunity differs. No central institution can pre-know which abilities, lifestyles, professions, aesthetics, or organizational forms will truly hold value over the next thirty years.

Public institutions remain essential. But their proper role is not to dictate the ideal human form, not to uniformly mold personalities through administrative metrics, and certainly not to turn AI into a new planning tool. Public institutions should instead focus on providing support, correcting imbalances, and empowering: helping vulnerable families improve their human development production functions, providing basic education and healthcare, protecting children from extreme disadvantage, flattening inequality of opportunity, and establishing fundamental rules and responsibility boundaries.

Markets, of course, are not perfect. They can be short-sighted, create anxiety, amplify inequality, and reward superficial charm over deep competence. Yet as a distributed experimental apparatus, the market remains irreplaceable. It allows countless families, schools, enterprises, communities, and individuals to experiment within their local knowledge; it permits new activities to be discovered, new needs expressed, new professions named, and new lifestyles validated. What it safeguards is the openness of human potential.

The revival of price theory does not mean returning to old-fashioned market fundamentalism but rather reunderstanding why markets matter. Markets matter not just because they are efficient but because, in a world of unknown futures, diverse abilities, and unnamed needs, only markets allow a society to conduct distributed exploration.

AI pushes the boundaries of the feasible set, but real-world frictions determine its diffusion speed. Technology opens the door to a new world, but human self-reproduction determines whether we truly step inside.

The next thirty years will not be a myth of vertical macroeconomic takeoff but a marathon for a generation. Old capabilities will depreciate; new ones will emerge. Old sectors will contract; new ones will expand. Old statistics will fail; new accounting will form. Old education will transform painfully, while family, health, psychology, expressiveness, aesthetics, and responsibility will become new productive cores.

Preserving markets for the new world means preserving space for unnamed activities. Preserving choice for families and individuals means preserving paths for unrealized human potential.

AI can generate text, images, code, and plans. But what will remain truly scarce in the future are humans who can grow up healthy, dare to take responsibility, deeply understand others, keenly discern opportunities, calmly organize cooperation, and conscientiously create life.

This thirty-year marathon has only just begun.