betaFeedback

Reading v1View latest version

WorksEconomics, Econometrics and Finance

The Turing Valley: How AI Capabilities Shape Labor Income

Abstract

There is concern that progress toward AI systems with strong capabilities across domains will reduce the importance of human input in production and thus wages. We show that when knowledge is tacit and multidimensional, making AI less jagged can instead raise labor's marginal product. Tacit knowledge makes sequential problem solving efficient because problems cannot be assigned ex ante to the agent best equipped to solve them. When organizations cannot fully integrate human and AI knowledge across dimensions, improving AI where humans initially have an advantage can remove from the referral pool problems that would otherwise consume human time and remain unsolved. By concentrating scarce human time on problems humans can solve, better screening can raise labor's marginal product even as fewer problems require human input. Our results imply that the human-AI versus AI-only performance gap used to measure human contribution to output need not track the marginal product of labor.

Introduction

The rapid advancement of Artificial Intelligence (AI) is set to transform the future of work,2 prompting widespread debate over its potential societal impact. While some contend that AI will boost productivity and drive economic growth, benefiting society as a whole, others worry that the technology could substitute human labor, reduce wages, and weaken the economic and political power of workers ([?]; [?]; [?]; [?]; [?]; [?]; [?]).

Motivated by this debate, this paper develops a model of a knowledge economy where production requires solving multidimensional problems, and firms organize production by matching humans with AI-powered machines in hierarchies designed to use knowledge efficiently. The novelty of our approach lies in placing the multidimensional nature of knowledge at the center of the analysis. This focus allows us to explore the possibility that AI may outperform humans in certain knowledge dimensions—such as pattern recognition or data analysis—while underperforming in others—such as reasoning and decision-making. This creates opportunities for human-AI collaboration to outperform both humans and AI alone.3

Using this model, we then seek to answer the following questions: Which types of AI improvements increase labor income, and which ones reduce it? Do AI advancements that allow humans to focus on more exceptional, less routine problems necessarily increase labor income? How does the AI that maximizes labor income compare to the AI that maximizes capital income (i.e., the income of machine owners)? How do these outcomes depend on the availability of computing power (or “compute”) and the user-friendliness of the AI technology?4

Formally, we consider a competitive economy where firms can employ humans or “machines” (i.e., units of compute) to produce. Production requires time and knowledge to solve problems, which vary in difficulty across two dimensions. Successful production occurs when the knowledge of the human or machine attempting a problem exceeds the problem’s difficulty in each dimension; in other words, both dimensions must be addressed for production to succeed. For example, in healthcare, patient care requires not only an accurate diagnosis but also clear and effective interaction with the patient.

Each human is endowed with one unit of time and an exogenous knowledge profile, which we assume to be identical for all humans. Machines, which are in fixed supply, also possess one unit of time each, and their knowledge is powered by an exogenous AI algorithm common to all machines. Hence, when the AI algorithm improves, the knowledge of all machines simultaneously improves. We impose no restrictions on the relationship between the knowledge profiles of humans and machines; machines may be inferior to humans in both dimensions of knowledge, superior in only one dimension, or superior in both.

Humans and machines can also communicate at a cost, enabling them to assist each other in solving problems. This interaction allows for the formation of two-layer organizations, where one factor specializes in production while the other specializes in problem-solving.5 Successful production in these organizations occurs when the combined human and machine knowledge in each dimension surpasses the problem’s difficulty. This interaction potentially creates knowledge synergies, meaning the set of problems that humans and machines can jointly solve may be larger than the union of the sets of problems each can solve alone.

As the baseline scenario, we examine the case where machines are sufficiently abundant so that the binding constraint in human-AI interactions is human time. In other words, not all machines can be matched with humans in two-layer organizations. This assumption is motivated by the exponential growth of computational capacity over the past two centuries [?]. In this case, the equilibrium rental rate of machines equals the fraction of problems they can solve independently, and labor captures the full value of human-machine collaboration. Furthermore, improvements in AI in any dimension of knowledge always increase total output, total capital income, and average labor productivity (defined as output per human). Our primary focus, however, is on understanding how advancements in AI impact equilibrium labor income, which is determined by the marginal product of labor.

Our first main result states that labor income decreases when AI improves in dimensions where its knowledge is currently inferior to humans but increases when AI improves in dimensions where it is already superior to humans. Viewed through the lens of [?]'s famous test of machine intelligence,6 improvements in AI in dimensions where machines cannot pass the Turing Test reduce labor income, while improvements in AI in dimensions where machines can pass the Turing Test increase it. For this reason—and given the topography that arises—we refer to the function that maps AI’s knowledge to labor income, illustrated in Figure 1, as the “Turing Valley.”

The Turing Valley: How AI Knowledge Shapes Labor Income

Figure 1. The Turing Valley: How AI Knowledge Shapes Labor Income. Notes. The green and blue areas represent the regions of AI knowledge where, in equilibrium, machines are placed in the bottom and top layers, respectively. The red area around human knowledge represents the region of AI knowledge where two-layer organizations do not emerge in equilibrium.

For intuition, let us focus on the case where machines are placed at the bottom layer, so they seek assistance from humans when they cannot solve the problem independently. AI improvements in dimensions where machines are stronger than humans increase labor income because they enhance the value of human-machine collaboration. This occurs for two distinct reasons. First, every new problem machines can now solve on their own is one that the organization could not solve before. Hence, machines now pose fewer questions that humans cannot answer, effectively improving the pool of problems humans have to deal with. Second, these improvements expand the set of problems that humans and machines can jointly solve more significantly than the set of problems machines can solve alone, i.e., they increase the extent of knowledge synergies between humans and machines.

In contrast, AI improvements in dimensions where machines are weaker than humans reduce labor income by reducing the value of human-machine collaboration. Such improvements only broaden the set of problems that machines can solve independently without increasing those that humans and machines can solve together. As a result, machines ask more difficult questions, worsening the pool of problems humans have to deal with.

Our second main result characterizes the AI that maximizes total labor income. We show that when humans are sufficiently weak in both knowledge dimensions, labor income is maximized when AI is as good as possible in both dimensions. Otherwise, labor income is maximized when AI simultaneously performs as poorly as possible in the dimension of knowledge where humans are relatively strong and as well as possible in the dimension where humans are relatively weak.

The reasoning is as follows: Machines that perform well in both dimensions are always useful for humans because they enable them to solve a large fraction of problems. However, these machines are also relatively expensive due to their ability to solve many problems independently. In contrast, machines that excel in one dimension but underperform in another are less costly—because they lack autonomy—and their usefulness depends on whether humans can complement the knowledge they lack. Therefore, unless humans are sufficiently strong in at least one dimension, the AI that maximizes labor income is the most advanced AI possible.

Our two main results have several important implications. First, AI advancements that further automate routine problems that could already be solved by combining human and machine knowledge necessarily reduce labor income. This occurs despite the fact that more knowledgeable machines allow humans to assist more machines with exceptional problems, as each machine requires less assistance, thereby increasing average labor productivity. The decrease in labor income happens because these improvements lead to a larger portion of output being captured by machines, outweighing the additional output generated by humans who assist a larger number of more advanced machines.

Second, if machines initially underperform humans across all dimensions of knowledge and AI improvements are gradual, these advancements will inevitably decrease labor income before potentially increasing it. Politically, pursuing or allowing technological advancements that initially reduce labor income could be contentious, especially since these initial reductions may only be reversed by substantial subsequent AI advancements—a scenario that may or may not materialize. Therefore, temporary labor income support policies might be necessary to address these political challenges. Alternatively, AI developers could be required to meet discrete improvement thresholds before deploying new AI models.

Third, the AI that maximizes labor income may differ significantly from the one that maximizes capital income (before considering the costs of developing this technology). This discrepancy occurs when humans are relatively strong in at least one dimension of knowledge. In this case, labor prefers relatively cheap, non-autonomous machines that can complement humans in the dimension where they are relatively weak. In contrast, capital prefers autonomous machines that can perform as well as possible in both dimensions of knowledge.

Our baseline setting assumes that (i) machines are abundant and (ii) all humans have the same knowledge profile. Our two main extensions relax each of these assumptions. In the first extension, we consider the general case where there is an arbitrary amount of machines. In this context, we identify a new force that can cause improvements in AI to reduce labor income: AI improvements may increase the amount of machines each human can assist to such an extent that the equilibrium shifts from a state where not all machines are matched with humans to one where they are. This shift induces a discontinuous drop in labor income, as labor is then forced to share with capital the value created by human-machine collaboration.

This extension also reveals that making machines more user-friendly—by reducing the communication costs between humans and machines—can be a double-edged sword from labor’s perspective. Specifically, while lower communication costs always increase labor income when machines are abundant (by increasing the value of human-machine collaboration), the resulting rise in machine demand from a more user-friendly technology can lead to a discontinuous drop in labor income akin to the one described above. More generally, this extension suggests that policymakers concerned about labor income might want to ensure an adequate supply of machines (i.e., sufficient compute) to keep machines abundant as AI advances.

In our second extension, we examine a model with two types of humans, each with distinct knowledge profiles. We show that our first main result—regarding which types of AI improvements increase total labor income—generalizes as follows: Total labor income decreases when AI improves in dimensions where it is inferior to all types of humans (i.e., machines fail the Turing Test against any human in that dimension) but increases when AI improves in dimensions where it surpasses all types of humans (i.e., machines pass the Turing Test against all humans in that dimension).

Regarding our second main result—the type of AI that maximizes labor income—we show that this result generalizes on a type-by-type basis. However, the same AI may not maximize the labor income of both types of humans. This result indicates that the direction of AI development could become politically divisive not only along capital-labor lines, as previously discussed, but also within the labor force itself.

The remainder of the paper is organized as follows: After reviewing the most relevant literature, we introduce our baseline model in Section 2. Section 3 characterizes the competitive equilibrium, and Section 4 presents our main results. Sections 5 and 6 discuss our two main extensions. Finally, Section 7 concludes.

The main contribution of this paper is the development of a new model of human-AI collaboration with multidimensional knowledge to analyze how various AI improvements impact labor income. In doing so, we contribute to two distinct literatures: the literature on knowledge hierarchies and the literature on the effects of technological progress on labor income.

In the context of the literature on knowledge hierarchies, the most closely related paper is our own recent work ([?]). That paper is the first to introduce AI into a canonical model of a knowledge economy: The knowledge hierarchies of [?], [?], and [?].7 The present paper differs from [?] in two key respects. First, the research question is different. [?] focus on characterizing the effects of AI on organizational and labor outcomes by comparing the equilibrium before and after AI’s introduction. In contrast, the present paper seeks to understand the effects of different types of AI improvements once AI is already in use. This question is important because this technology is improving at the same time as it is being deployed, and there is an active debate on whether the current direction of AI developments will benefit society as a whole or concentrate economic power among a select few ([?, ?, ?, ?, ?, ?]).

Second, while [?], in line with the previous literature on knowledge hierarchies, examines a setting where knowledge is unidimensional, the present paper introduces multidimensional knowledge. In this context, we show that organizations can be valuable for reasons beyond the traditional specialization motive found in one-dimensional models. For instance, when machines struggle with dimensions of knowledge in which humans excel—and vice versa—communication between humans and machines can allow them to solve problems that neither could solve independently.8

This paper is also related to the literature examining the effects of technological progress on labor income (e.g., [?, ?, ?, ?, ?, ?]). A key insight from this body of work is that the conventional wisdom—that technological progress increasing average labor productivity is ultimately beneficial for workers—is flawed. As explained by [?] and [?], this is because wages are determined by marginal productivity, and there is no guarantee that average and marginal productivity move together.9

We contribute to this line of work by studying the effects of a specific type of technological progress—AI advancements—on labor income. To do this, and motivated by the fact that AI primarily impacts knowledge work, we develop a framework inspired by the theory of knowledge hierarchies. Through this framework, we identify which types of AI improvements benefit or harm labor and identify the AI that serves labor’s best interests as a function of the knowledge of the existing population, the availability of compute, and the user-friendliness of the AI technology.

The Model

We begin by formally describing the model, followed by a discussion of its main elements and assumptions.

The Baseline Setting

The Basics.—There is a unit mass of identical humans, each endowed with one unit of time and exogenous knowledge h=(h1,h2)∈(0,1)2\boldsymbol{h}=(h_1,h_2)\in(0,1)^2. Additionally, there are μ>0\mu>0 units of compute, normalized such that one unit of compute is equivalent to one unit of time. An exogenous AI algorithm provides these units of compute with knowledge m=(m1,m2)∈[0,1]2\boldsymbol{m}=(m_1,m_2)\in[0,1]^2. For simplicity, we refer to one unit of compute as a “machine,” with each machine having one unit of “time.”

There are many identical competitive firms. Production occurs within these firms, which are the residual claimants of output. Time and knowledge are the only inputs. Firms have no fixed costs and a maximum of two layers.

Single-layer firms can be automated or non-automated. Single-layer automated firms hire a single machine to produce. This machine dedicates her full unit of time to a single production opportunity. Each production opportunity is associated with a problem with unknown difficulty x≡(x1,x2)\boldsymbol{x}\equiv(x_1,x_2) drawn from a distribution with full support on [0,1]2[0,1]^2, cumulative distribution function FF and probability density function ff. If the knowledge of the machine exceeds the problem’s difficulty in both dimensions, i.e., if m1≥x1m_1\geq x_1 and m2≥x2m_2\geq x_2, the machine solves the problem and produces one unit of output. Otherwise, no output is produced. Single-layer non-automated firms are identical to their automated counterparts, except that they use humans for production instead of machines.

Two-layer organizations combine humans and machines and can be bottom-automated or top-automated. Bottom-automated (“bb”) firms rent machines for production and employ a single human to assist with problems the machines cannot solve. Production occurs when the combined knowledge of the human and machine exceeds the problem’s difficulty, i.e., if max⁡{m1,h1}≥x1\max\{m_1,h_1\}\geq x_1 and max⁡{m2,h2}≥x2\max\{m_2,h_2\}\geq x_2. Otherwise, no production occurs. Assistance consumes c∈(0,1)c\in(0,1) units of the human’s time. Since each machine cannot solve the problem it encounters with probability 1−F(m)1-F(\boldsymbol{m}), bb firms optimally employ n(m)n(\boldsymbol{m}) machines to fully utilize the human’s time, where:

n(m)≡1c(1−F(m))n(\boldsymbol{m})\equiv\frac{1}{c(1-F(\boldsymbol{m}))}

Top-automated (“tt”) firms are similar to bb firms, except the roles are reversed: Humans engage in production, while machines specialize in problem-solving. In this case, consulting with a machine consumes c∈(0,1)c\in(0,1) units of the machine’s time. Consequently, tt firms always employ n(h)n(\boldsymbol{h}) humans to fully exploit the time of the machine that assists these humans.

Figure 2 depicts the four firm configurations that can potentially arise in equilibrium. Note that a two-layer organization never uses only humans (or only machines) in both layers of the organization, as the human (or machine) engaged in production would solve the same set of problems as the one specialized in problem-solving. A similar reasoning also implies that focusing on two-layer organizations is without loss of generality in this setting: organizations with more than two layers cannot solve any problems that two-layer organizations cannot solve.

The four possible firm configurations

Figure 2. The Four Possible Firm Configurations

For brevity, we refer to agents (humans or machines) in two-layer organizations as “workers” if they engage in production and as “solvers” if they specialize in problem-solving. We call agents in single-layer organizations “independent producers” as they engage in unassisted production work.

Wages, Prices and Profits.— Let ww be the wage of humans and denote by rr the rental rate machines. All agents in this economy are risk neutral and maximize their income. This implies, in particular, that labor supply is inelastic. We normalize the value of each unit of output to one.

The profit of a single-layer firm as a function of its type is given by:

Πs={F(h)−wif non-automatedF(m)−rif automated\Pi_s = \begin{cases} F(h) - w & \text{if non-automated} \\ F(m) - r & \text{if automated} \end{cases}

In other words, the profit of a single-layer non-automated firm is the expected fraction of problems that a human independent producer can solve, F(h)F(h), minus her wage, ww. The profit of a single-layer automated firm is obtained similarly, with the human replaced by a machine.

The profit of a two-layer organization as a function of its type (i.e., bb or tt) is given as follows:

Πb=n(m)[F(m∨h)−r]−w\Pi_b = n(m)[F(m \mathbin{\vee} h) - r] - w
Πt=n(h)[F(m∨h)−w]−r\Pi_t = n(h)[F(m \mathbin{\vee} h) - w] - r

where m∨hm \mathbin{\vee} h denotes the maximum dimension per dimension between mm and hh.10 In both cases, the profit of a firm is its expected output minus the cost of the resources it uses. For instance, in the case of a bb firm, its total expected output is n(m)F(m∨h)n(m)F(m \vee h), as such organization attempts n(m)n(m) problems and successfully solves a fraction F(m∨h)=F(max⁡{m1,h1},max⁡{m2,h2})F(m \vee h) = F(\max\{m_1,h_1\},\max\{m_2,h_2\}) of them. In turn, the cost of the resources is w+n(m)rw+n(m)r, as the firm employs one human solver and n(m)n(m) machine workers.

Competitive Equilibrium.—Let αs\alpha_s, αb\alpha_b, and αt\alpha_t be the mass of humans hired by single-layer non-automated, bottom-automated, and top-automated firms, respectively. Similarly, denote by μs\mu_s, μb\mu_b, and μt\mu_t, the mass of machines rented by single-layer automated, bottom-automated, and top-automated firms.

Definition 1 (Competitive Equilibrium). An equilibrium consists of a set of non-negative allocations (αs,αb,αt,μs,μb,μt)(\alpha_s,\alpha_b,\alpha_t,\mu_s,\mu_b,\mu_t) and a set of non-negative prices (w,r)(w,r), such that:

  1. Firms optimally choose their structure (while earning zero profits).

  2. Single-layer non-automated firms hire a mass αs\alpha_s of humans.

  3. Single-layer automated firms hire a mass μs\mu_s of machines.

  4. bb firms hire a mass αb\alpha_b of humans and rent a mass μb=αbn(m)\mu_b=\alpha_b n(m) of machines.

  5. tt firms hire a mass αt\alpha_t of humans and rent a mass μt=αtn(h)−1\mu_t=\alpha_t n(h)^{-1} of machines.

  6. Markets clear: αs+αb+αt=1\alpha_s+\alpha_b+\alpha_t=1 and μs+μb+μt=μ\mu_s+\mu_b+\mu_t=\mu.

Machines are “Abundant”.—As our baseline, we focus on the case where machines are sufficiently abundant so that the binding constraint in human-AI interactions is human time, not machine time. More precisely, we assume that:

μ>max⁡{1c(1−F(1,h2)),1c(1−F(h1,1))}(1)\mu> \max\left\{\frac{1}{c(1-F(1,h_2))},\frac{1}{c(1-F(h_1,1))}\right\} \tag*{(1)}

This condition ensures that, regardless of machine knowledge m∈[0,1]2m \in[0,1]^2, there are always more machines available than those demanded in equilibrium by two-layer organizations, i.e., μ>αb∗n(m)+αt∗/n(h)\mu> \alpha_b^*n(m)+\alpha_t^*/n(h).11 In Section 5, we consider the general case where condition (1) does not necessarily hold.

Discussion of the Model

Before proceeding with the analysis, we briefly comment on the main assumptions underlying our model.

First, our framework emphasizes the multidimensional nature of knowledge. This focus enables us to explore the possibility that AI may outperform humans in certain knowledge dimensions while underperforming in others. This creates opportunities for human-AI collaboration to outperform both humans and AI alone. Moreover, it allows for the possibility that machines may excel in certain dimensions of knowledge but still lack “autonomy” because they struggle in other dimensions.

Second, we assume that communication between humans and machines allows them to combine their knowledge, potentially creating knowledge synergies—meaning the set of problems that humans and machines can jointly solve may be larger than the union of the sets of problems each can solve alone. In the Online Appendix, we show the sense in which our results hold when humans and machines cannot combine their knowledge (so solving a problem requires that one of them knows how to solve it independently).

Third, we assume that humans and machines cannot identify the specific dimensions in which they fail to solve a problem. Consequently, they cannot selectively choose which problems to seek assistance with among those they cannot solve. If, however, humans and machines could identify exactly which dimensions they know how to solve and which they do not, firms would optimally develop complex protocols—which could, in principle, be contingent on the knowledge of humans and machines, as well as the difficulty of the problem at hand—dictating when production workers should escalate unresolved issues. While this extension is intriguing, the complexity it introduces is beyond the scope of this paper and is left for future research.

Fourth, to focus on the matching between humans and machines rather than between humans, our baseline model assumes that all humans share the same knowledge profile. However, in Section 6, we show how our main insights generalize to the case in which there are two different types of humans. The analysis of this multidimensional model with more general distributions of human knowledge is beyond the scope of this paper and is left for future research.

Fifth, we assume that all machines are equipped with the same AI and that when AI improves, the knowledge of all machines simultaneously improves. This assumption is motivated by the fact that algorithms, as digital information, are non-rival and have nearly zero marginal reproduction costs [?]. Consequently, even if multiple AI algorithms exist, all firms will choose to equip their “machines” (i.e., the units of compute they rent) with the best available AI.

Finally, in the baseline model, we assume that machines are sufficiently abundant, making human time the limiting factor in human-AI interactions.12 This assumption is motivated by the exponential growth of computational capacity over the past two centuries ([?]). However, some commentators are beginning to caution that constraints on energy and chip supply could cause the availability of computational resources to lag behind rising demand.13 For this reason, in Section 5 we consider the general case in which machines may or may not be abundant.

Equilibrium Characterization

Our main goal is to characterize how equilibrium labor income depends on machine knowledge. Towards this end, this section characterizes the competitive equilibrium of the economy. In Section 4, we present our main results.

The abundance of machines relative to humans implies that a positive mass of machines must be rented by single-layer automated firms. Hence, by the zero-profit condition of these firms, r∗=F(m)r^*=F(\mathbf{m}). In other words, the equilibrium rental rate of machines is equal to the fraction of problems they can solve on their own.

Define wˉi\bar{w}_i as the highest wage an organization of type i∈{s,b,t}i \in\{s,b,t\} can offer, given that r∗=F(m)r^*=F(\mathbf{m}). The wage wˉi\bar{w}_i is determined by the zero-profit condition of a firm of type ii:

wˉs≡F(h)wˉb≡n(m)(F(m∨h)−F(m))wˉt≡F(m∨h)−F(m)/n(h)(2)\begin{aligned} \bar{w}_s &\equiv F(\mathbf{h}) \\ \bar{w}_b &\equiv n(\mathbf{m})\left(F(\mathbf{m}\vee\mathbf{h})-F(\mathbf{m})\right) \\ \bar{w}_t &\equiv F(\mathbf{m}\vee\mathbf{h})-F(\mathbf{m})/n(\mathbf{h}) \tag*{(2)} \end{aligned}

Finally, let Rs={m∈[0,1]2:wˉs≥max⁡{wˉb,wˉt}}R_s=\{\mathbf{m}\in[0,1]^2:\bar{w}_s\ge\max\{\bar{w}_b,\bar{w}_t\}\}, Rb={m∈[0,1]2:wˉb>max⁡{wˉs,wˉt}}R_b=\{\mathbf{m}\in[0,1]^2:\bar{w}_b>\max\{\bar{w}_s,\bar{w}_t\}\}, and Rt=[0,1]2∖(Rs∪Rb)R_t=[0,1]^2\setminus(R_s\cup R_b). We then have the following proposition:

Proposition 1. The competitive equilibrium maximizes total output. The equilibrium rental rate of machines is r∗=F(m)r^*=F(\mathbf{m}) and the equilibrium wage is w∗=max⁡{wˉs,wˉb,wˉt}w^*=\max\{\bar{w}_s,\bar{w}_b,\bar{w}_t\}. Moreover:14

References

  1. [1]Carter, Michael. 2001. Foundations of Mathematical Economics. The MIT Press.

Paper details

vlnc:W_krzws7fa
Published
Added
License
CC BY 4.0
Usage

views · 0

downloads · 4

A view counts once per browser per day, a download each time the PDF is requested.

  • Economics, Econometrics and Finance

Not recorded for this paper

View statement
Looking for related papers…
Contents