The economic history of the last two centuries has been shaped by the continuous interplay between technological progress and the accumulation of human capital. From the onset of the Industrial Revolution, the substitution of capital for labor has followed a predictable trajectory, automating manual tasks while leaving the domain of cognitive work largely untouched.
However, the emergence of generative Artificial Intelligence marks a structural break in this historical trend. We are currently witnessing a shift where machines are no longer confined to routine physical or codifiable cognitive tasks but are increasingly capable of performing sophisticated knowledge work—ranging from coding and research to complex problem-solving [Ide, 2025]Ide, E. & Talamàs, E.
Artificial Intelligence in the Knowledge Economy
Journal of Political Economy, 2025.
This development necessitates a fundamental re-evaluation of our economic models, akin to the re-evaluation of welfare criteria necessitated by the Stern Review in the context of climate change [Stern, 2006]Stern, N.
The Economics of Climate Change: The Stern Review
Cambridge University Press, 2006. Just as climate policy requires weighing the welfare of future generations against present consumption, the integration of AI into the knowledge economy requires a rigorous assessment of the trade-offs between aggregate output efficiency and the normative distribution of labor income.
Core Argument
The impact of AI is not deterministic but depends critically on two factors: its autonomy (whether it acts as a coworker or merely a copilot) and its capability level (whether it possesses basic or advanced knowledge). These choices create fundamentally different equilibrium outcomes for wages, inequality, and output.
The defining characteristic of the knowledge economy is that production know-how is predominantly tacit—or noncodifiable. It is developed through repeated observations of practical successes and failures and is inherently embodied in individuals [Polanyi, 1966]Polanyi, M.
The Tacit Dimension
New York: Doubleday, 1966. Unlike codifiable knowledge, which can be inscribed in manuals or software code, tacit knowledge resists easy transfer. Consequently, individuals' time and knowledge become critical bottlenecks in production, making firms and hierarchical organizations central to the efficient utilization of this scarce resource [Garicano, 2000]Garicano, L.
Hierarchies and the Organization of Knowledge in Production
Journal of Political Economy, 2000.
The fundamental innovation of modern AI lies in its ability to learn by example, uncovering patterns that cannot be codified into explicit rules. This capability allows AI systems to acquire tacit knowledge and perform cognitive, noncodifiable tasks, thereby opening the door to a fundamental reorganization of production.
Section 2The Knowledge Hierarchy Baseline
To understand the transformative potential of AI, one must first establish a robust baseline of the pre-AI economy. We utilize the canonical model of knowledge hierarchies introduced by Garicano (2000) and further developed by Garicano and Rossi-Hansberg (2004, 2006) as the "working horse" for our analysis [Garicano, 2000]Garicano, L.
Hierarchies and the Organization of Knowledge in Production
Journal of Political Economy, 2000.
The Mechanics of Tacit Knowledge
In the baseline economy, production requires two inputs: labor (time) and knowledge. Humans are endowed with one unit of time and are heterogeneous in their knowledge level, denoted by z. Production opportunities present problems of varying difficulty x, distributed uniformly. A solution is achieved when the individual's knowledge exceeds the problem's difficulty.The binary production function captures the essence of the knowledge economy: value is generated not by effort alone, but by applying specific expertise to stochastic challenges.
The crucial friction is the cost of communicating tacit knowledge. Because experts cannot write down solutions for every contingency in advance, help must be sought in real-time. If a worker encounters a problem they cannot solve, they consult a manager. This communication consumes a fraction h of the manager's time, regardless of outcome. This communication cost effectively limits the span of control of experts and necessitates the formation of hierarchies.
The AI Transformation of the Knowledge Hierarchy
Isometric representation of pre-AI (left) vs post-AI (right) organizational structures. Warm amber nodes represent human workers; cold white lattice represents AI compute layer.
Management by Exception
In equilibrium, agents organize themselves into hierarchical firms exhibiting "management by exception"—a structure where less knowledgeable individuals handle routine problems, while more knowledgeable individuals handle the exceptions. The hierarchy functions as a filter: workers shield solvers from routine tasks, allowing expertise to be leveraged across more problems.
The optimal team size is determined by the solver's time constraint. This reveals a fundamental complementarity: as workers become more knowledgeable, they solve more problems independently, reducing interruptions. This allows solvers to manage larger teams. Consequently, equilibrium exhibits positive assortative matching: the most knowledgeable solvers are matched with the most knowledgeable workers, establishing a convex wage structure where solvers earn premiums for leveraged expertise.
Section 3Modeling the AI Shock
The introduction of AI represents a discontinuity in the production function of knowledge. We model AI based on three key developments: the scalability of tacit knowledge, the rise of foundation models, and the emergence of autonomous agents [Ide, 2025]Ide, E. & Talamàs, E.
Artificial Intelligence in the Knowledge Economy
Journal of Political Economy, 2025.
AI as Scalable Tacit Knowledge
The primary economic distinction between human intelligence and Artificial Intelligence is scalability. Human knowledge is rivalrous in application; an expert can only solve one problem at a time. AI, however, decouples knowledge from biological time constraints. Once an AI model acquires knowledge (encoded in its parameters), that knowledge can be applied across all available computational resources in the economy.
Human Capital
- • Rivalrous in application
- • Fixed supply, heterogeneous
- • Constrained by biological time
- • Knowledge transfer is costly
AI Capital
- • Non-rivalrous once trained
- • Infinitely replicable
- • Constrained only by compute
- • Knowledge encoded in parameters
The Putty-Putty Nature of AI Capital
To understand the economic character of AI capital, it is useful to draw on the distinction between "putty-clay" and "putty-putty" technologies. Putty-clay capital is flexible ex-ante but rigid ex-post—a welding robot cannot easily become a painting robot. Putty-putty capital remains flexible even after investment.
AI and compute behave as putty-putty capital. The underlying resource—GPUs/TPUs—is general-purpose. A foundation model trained for coding can be fine-tuned to analyze legal contracts or diagnose medical images. This implies that the transition dynamics of the AI economy will be faster and more volatile than previous technological revolutions.
Autonomy: Coworkers vs. Copilots
A critical contribution of the Ide and Talamàs framework is the distinction between Autonomous and Nonautonomous AI [Ide, 2025]Ide, E. & Talamàs, E.
Artificial Intelligence in the Knowledge Economy
Journal of Political Economy, 2025.
Autonomous AI
AI agents perform the same functions as humans. They can operate as coworkers (pursuing opportunities autonomously) and as solvers (providing advice).
Represents the technological frontier where AI acts as an independent economic agent [Hadfield, 2025]Hadfield, G.
AI and the Future of Work
Allied Social Science Associations Annual Meeting, 2025.
Nonautonomous AI
AI agents are restricted to acting solely as copilots. They provide advice but cannot independently pursue production opportunities.
Reflects "human-in-the-loop" regulatory environments and current technological constraints.
Section 4Equilibrium Analysis: Autonomous AI
We first analyze the benchmark case where AI is fully autonomous. Firms can hire AI agents (by renting compute) to replace humans in either the worker or solver layer of the hierarchy. The equilibrium effects depend critically on the AI's knowledge level relative to the human population. We identify two distinct regimes.
Basic Autonomous AI: The Displacement of Routine Labor
Consider the case where AI knowledge corresponds to that of pre-AI workers—the less knowledgeable segment. In this regime, AI agents handle routine production tasks but lack deep expertise. Because AI is scalable and cheap, firms deploy it as workers, fundamentally altering the demand for human knowledge.
This creates a "hollowing out" effect. The least knowledgeable humans face direct competition from AI in routine production. Furthermore, they lose access to the best human solvers, who now manage AI fleets. The most knowledgeable individuals benefit significantly, leveraging their expertise across massive scales of AI agents, decoupling their income from human labor supply constraints.
The Divergence of Fortunes
Dual-panel distribution showing wage changes by knowledge level. Left: Basic AI (polarization). Right: Advanced AI (U-shaped benefits).
Advanced Autonomous AI: The Democratization of Expertise
Now consider AI knowledge corresponding to pre-AI solvers—the more knowledgeable segment. Here, AI agents solve complex problems previously requiring human experts. Firms deploy AI as solvers, and because it's scalable, these AI solvers assist human workers infinitely.
Surprisingly, the least knowledgeable benefit most. Their productivity is no longer bottlenecked by the scarcity of human experts. Access to superior AI problem-solving increases their marginal product. The very top humans also benefit, using Advanced AI as workers for super-specialized production. Only the marginal human solvers—those just below the top tier—face displacement into routine work.
Non-Monotonic Relationship
Basic AI polarizes the market, harming the low-skilled and helping the high-skilled. Advanced AI, by commoditizing expertise, lifts the low-skilled. In both autonomous scenarios, overall output is maximized because the economy uses the most efficient resource for every task.
Section 5Nonautonomous AI: The Normative Trade-off
This regime is particularly relevant for policy discussions, reflecting environments where AI is permitted only as a "copilot" and forbidden from executing tasks independently—the "human-in-the-loop" paradigm.
The Copilot Mechanism
When AI agents are restricted to the solver role, competitive dynamics fundamentally change. Because AI cannot act as a coworker, it does not substitute for human labor in routine production. The least knowledgeable humans face no competition for their primary role. AI acts purely as a complement, relaxing the time constraint on problem-solving.
Nonautonomous AI primarily benefits the least knowledgeable individuals. By providing cheap, scalable advice, it allows low-skill workers to solve problems they otherwise could not. Since AI cannot replace them in production, these workers capture a significant share of productivity gains. This creates a "leveling up" effect where the performance gap between low-skill and high-skill workers compresses.
The Normative Frontier
The balance between aggregate output (efficiency) and distributional equity. Autonomous AI maximizes the former; Nonautonomous AI favors the latter.
The Tyranny of Efficiency vs. Equity
While Nonautonomous AI is favorable for equity, it imposes a penalty on aggregate efficiency. By restricting AI from performing routine production—where it might be more efficient—the economy operates inside its production possibility frontier. Overall output is strictly higher with Autonomous AI [Ide, 2025]Ide, E. & Talamàs, E.
Artificial Intelligence in the Knowledge Economy
Journal of Political Economy, 2025.
| Regime | AI Role | Winners | Losers | Output | Inequality |
|---|---|---|---|---|---|
| Nonautonomous | Copilot only | Least Knowledgeable | Most Knowledgeable (relative) | Lower | ↓ Reduces |
| Basic Autonomous | Coworker (routine) | Most Knowledgeable | Least Knowledgeable | Higher | ↑ Increases |
| Advanced Autonomous | Solver / Manager | Both extremes | Marginal Solvers | Highest | ∪ U-shaped |
Source: Adapted from Ide & Talamàs (2025)
Section 6Empirical Reconciliation
The theoretical framework offers a powerful lens for reconciling seemingly contradictory empirical evidence regarding AI's labor market impact.
The "Jagged Frontier" of Evidence
Current studies present a bifurcated picture. On one hand, experimental evidence from controlled settings suggests AI reduces performance inequality. Studies by Dell'Acqua et al. (2023) and Noy and Zhang (2023) find that AI tools help low-performing consultants and writers "catch up" to high performers [Dell'Acqua, 2023]Dell'Acqua, F. et al.
Navigating the Jagged Technological Frontier
Harvard Business School Working Paper, 2023. Our analysis identifies this as the signature effect of Nonautonomous AI—where humans remained primary agents, using AI for assistance.
On the other hand, non-experimental labor market data suggests different trends. Berger et al. (2024) document significant declines in demand for low-skill freelancers following ChatGPT's release, while demand for high-skill editors remained robust [Berger, 2024]Berger, P.G. et al.
Employer and Employee Responses to Generative AI
Working Paper, 2024. This aligns with our Basic Autonomous AI regime, where AI substitutes for low-skill human labor.
Reconciling the Evidence
The confusion in empirical literature arises from conflating different modes of AI deployment. When AI is used as a tool (copilot), it compresses inequality. When used as an agent (coworker), it tends to displace labor and widen inequality.
Comparison with Globalization
AI differs from previous labor supply shocks like offshoring in a fundamental way: homogeneity and scalability. Offshoring involves integrating heterogeneous human populations constrained by biological limitations. AI introduces homogeneous agents scalable infinitely, subject only to compute costs. This doesn't just shift the labor supply curve—it fundamentally alters the elasticity of substitution, making competitive pressure from AI more elastic and potentially more disruptive than human migration shocks.
Section 7Conclusion
This analysis has examined AI's integration into the knowledge economy, focusing on the interplay between distributional norms and the development of new productive knowledge. We have shown that AI's introduction is fundamentally different from expanding human labor force due to the unique properties of scalable tacit knowledge and the "putty-putty" nature of computational capital.
First
There is an inherent trade-off between maximizing economic output and preserving labor income equality. Autonomous AI maximizes efficient resource use but tends to concentrate gains among the most knowledgeable while displacing the least knowledgeable.
Second
Nonautonomous AI—often advocated under "Human-in-the-Loop" regulations—serves as a potent equalizing force. It disproportionately benefits the least knowledgeable but comes at the cost of aggregate efficiency.
Third
The evolution from "Basic" to "Advanced" AI capabilities may naturally resolve some distributional concerns. Advanced Autonomous AI has potential to benefit a broader population base by commoditizing high-level problem solving.
Fourth
AI's impact is not deterministic but a function of institutional choices regarding autonomy. As Hadfield (2025) suggests, we must shift from viewing AI as a tool to viewing it as a novel economic agent [Hadfield, 2025]Hadfield, G.
AI and the Future of Work
Allied Social Science Associations Annual Meeting, 2025.
The theoretical considerations presented here suggest we are standing at an "Archimedean point" in economic history. Future empirical work must move beyond aggregate productivity measures to carefully disentangle specific modes of AI deployment—autonomous versus nonautonomous—and their heterogeneous impacts on the workforce.
References
Ide, E. & Talamàs, E. (2025). Artificial Intelligence in the Knowledge Economy. Journal of Political Economy.
Garicano, L. (2000). Hierarchies and the Organization of Knowledge in Production. Journal of Political Economy.
Polanyi, M. (1966). The Tacit Dimension. New York: Doubleday.
Stern, N. (2006). The Economics of Climate Change: The Stern Review. Cambridge University Press.
Hadfield, G. (2025). AI and the Future of Work. Allied Social Science Associations Annual Meeting.
Dell'Acqua, F. et al. (2023). Navigating the Jagged Technological Frontier. Harvard Business School Working Paper.
Berger, P.G. et al. (2024). Employer and Employee Responses to Generative AI. Working Paper.