📈 economics

Data-Driven Automation

This paper develops a dynamic model where heterogeneous, endogenously accumulating data with cross-task spillovers drives automation, leading to rich short-run dynamics and explosive growth but ultimately resulting in a slow, power-law decay of labor's share and stagnant long-run wages due to generic market inefficiencies.

Original authors: Maryam Farboodi, Andrew Koh, Anchi Xia

Published 2026-06-10

📖 6 min read🧠 Deep dive

CC0 1.0

Original authors: Maryam Farboodi, Andrew Koh, Anchi Xia

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Learning Loop"

Imagine the economy is a giant kitchen with thousands of different recipes (tasks). In the past, robots (capital) could only cook simple, repetitive dishes like chopping onions. But modern AI is like a super-chef that can learn to cook anything, provided it has enough cookbooks (data).

This paper argues that we are entering a new era where the act of cooking creates the cookbooks.

The Loop: Every time a human or a robot cooks a dish, it generates a "recipe log" (data).
The Upgrade: The AI reads these logs to get better at cooking that specific dish.
The Spillover: Sometimes, learning to cook a steak helps the AI learn how to cook a burger, because the skills overlap.

The authors built a mathematical model to answer three big questions:

Will robots eventually do everything?
What happens to human wages?
How fast will this happen?

1. The Race Between "Getting Better" and "Getting Bored"

The speed of automation depends on two main factors:

How much the dishes are alike (Substitutability): If a robot learns to cook a burger, does it instantly get good at cooking a sandwich? (High substitutability). Or is cooking a burger totally different from cooking a salad? (Low substitutability/Complementarity).
The Law of Diminishing Returns: The first 1,000 cookbooks make the AI amazing. The next 1,000 make it slightly better. The next 1,000 make it barely any better.

Scenario A: The "All-In" Automation (Low Substitutability)

If the tasks are very different from each other (like cooking a salad vs. coding a website), the AI has to learn each one separately.

The Result: The AI eventually learns everything. It becomes fully automated.
The Catch: Even though the AI gets infinitely better at cooking, human wages stop growing. Why? Because as the AI gets better at the "hard" tasks, humans are pushed into doing the "easy" tasks that the AI hasn't mastered yet. But since the AI is getting so good so fast, it eventually takes over those easy tasks too. Humans end up training their own replacements.

Scenario B: The "Stuck" Automation (High Substitutability)

If the tasks are very similar (like writing a legal contract vs. writing a marketing email), the AI can easily transfer its skills.

The Result: The AI gets obsessed with the tasks it is already good at. It pours all its resources into mastering those few tasks and ignores the rest.
The Catch: The economy becomes partially automated. The AI dominates a few "core" sectors, but a "periphery" of tasks remains stuck with humans because the AI doesn't see the point in learning them (they aren't similar enough to the core tasks to be worth the effort).
The Good News for Humans: In this scenario, human wages can actually keep growing forever because the AI never fully takes over the whole economy; it leaves a niche for humans.

2. The Speed of Change: The "Fat Tail"

You might think that if AI is learning so fast, it will take over everything overnight. The paper says: Not so fast.

Even in the best-case scenario where the AI eventually does everything, the process is surprisingly slow.

The Analogy: Imagine a runner sprinting the first 95% of a marathon. They are fast. But the last 5%? They slow down to a crawl.
The Math: The paper shows that the number of tasks done by humans shrinks according to a "power law." This means there is a "fat tail" of tasks. Even after the AI is super-smart, there will always be a small, stubborn group of jobs that humans do for a very long time. It's not because humans are better; it's just that the AI is so busy getting perfect at the other 99% that it takes a long time to bother with the last 1%.

3. The "Contagion" Effect

What if the AI learns to cook Italian food, and that helps it learn French food?

The Network: The paper uses a map (a graph) to show how tasks are connected. If every task is connected to every other task (even indirectly), the AI will eventually learn everything, no matter how different the tasks seem.
The Core-Periphery Problem: Imagine a city with a "Core" (tech hubs) and a "Periphery" (rural towns). If the Core shares data with itself but rarely talks to the Periphery, the Core gets super-automated very fast. The Periphery gets left behind. The AI might take over the Core in a few years, but it could take decades to reach the Periphery because the "data bridge" is weak.

4. Is the Market Doing the Right Thing? (Inefficiency)

The paper argues that the free market is bad at planning how to collect data.

The Problem: Companies only care about the data they use right now. They don't care that their data might help a different company later.
The Fix: A "Planner" (like a government) would look at the big picture.
- If tasks are very different, the Planner would force robots to work on the "boring, data-poor" tasks first to speed up learning. The market won't do this because it's not profitable today.
- If tasks are very similar, the Planner would double down on the "data-rich" tasks to make the AI explode in power. The market might be too cautious.
Takeaway: The market naturally drifts in the wrong direction, making automation either too slow or too lopsided.

5. The "Explosive" Future (With Savings)

So far, we assumed the amount of robots (capital) is fixed. But what if we can build more robots?

The Singularity: If the AI gets better at building robots, and building robots creates more data, which makes the AI even better... you get a feedback loop.
The Result: The economy could experience explosive growth where output and consumption become infinite in a finite amount of time. It's like a snowball rolling down a hill that gets bigger faster than it can roll.
The Human Cost: Even in this explosion of wealth, human wages might still stagnate because the robots are doing almost everything.

Summary in One Sentence

Data-driven automation creates a powerful loop where work generates the intelligence to do more work, but depending on how similar tasks are, this could lead to a future where robots do everything (stagnant wages) or just the "easy" stuff (growing wages), and in either case, the transition will be surprisingly slow with a stubborn "tail" of human jobs left behind.

Technical Summary: Data-Driven Automation

1. Problem Statement

The paper addresses the trajectory and economic implications of modern AI-driven automation, characterized by a distinct feedback loop between data generation and capital productivity. Unlike previous waves of technological change that primarily displaced routine tasks, modern AI systems rely on vast, heterogeneous, and task-specific data to improve performance. The authors identify a critical gap in existing literature: while data is generated as a byproduct of economic activity, it simultaneously augments the productivity of capital, thereby expanding the frontier of automated tasks.

The central problem is to determine how this data-automation feedback loop shapes the speed, direction, and long-run limits of automation. Specifically, the paper investigates:

Under what conditions the economy achieves full versus partial automation.
How the network structure of data spillovers (transfer learning) influences automation dynamics.
The implications for long-run wages and economic efficiency.
Whether data-driven automation, combined with capital accumulation, leads to explosive growth or stagnation.

2. Methodology

The authors construct a dynamic general equilibrium model based on the task-based framework (Zeira, 1998; Acemoglu and Restrepo, 2018), extended to incorporate endogenous data accumulation.

Core Model Features

Task Space: A continuum of tasks $X = [0,1]$ aggregated via a Constant Elasticity of Substitution (CES) production function with elasticity $\sigma$ .
Factors of Production: Capital ( $K$ ) and Labor ( $L$ ). In the baseline, these are fixed; later, capital accumulation is introduced.
Data as a State Variable: Data ( $D_{it}$ ) is task-specific and accumulates endogenously as a byproduct of production: $\dot{D}_{it} = y_{it}$ . Data does not depreciate.
Capital Productivity: The productivity of capital in task $i$ , $\psi^K_i$ , is a function of effective data $A_i(D)$ :
$\psi^K_i(A_i(D)) = f_i \cdot (A_i(D))^\eta$
where $\eta \in (0,1)$ captures diminishing returns to data, and $f_i$ represents task-specific heterogeneity.
Cross-Task Spillovers: Modeled using a graphon $W(i,j)$ , a measurable function representing the intensity of data spillovers from task $j$ to task $i$ . Effective data for task $i$ is $A_i(D) = \int W(i,j) D_j dj$ . This captures transfer learning and skill overlaps.
Automation Boundary: An equilibrium threshold $\gamma_t$ separates tasks produced by capital ( $i \le \gamma_t$ ) from those produced by labor ( $i > \gamma_t$ ).

Analytical Approach

The authors analyze the model in three stages:

Data Autarky: No spillovers ( $W(i,j) = \mathbb{I}_{i=j}$ ). They derive conditions for full vs. partial automation based on the relationship between task substitutability ( $\sigma$ ) and data returns ( $\eta$ ).
Connected Economies: Introducing spillovers to analyze "contagious" automation and core-periphery dynamics.
Efficiency and Growth: Comparing the decentralized equilibrium to a social planner's problem and introducing endogenous capital accumulation to study growth limits.

3. Key Contributions and Results

A. Long-Run Automation Limits

The degree of automation depends on the interplay between cross-task substitutability ( $\sigma$ ) and the diminishing returns to data ( $\eta$ ).

Full Automation ( $\sigma \le 1/\eta$ ): If tasks are sufficiently complementary (or weak substitutes), the economy converges to full automation ( $\lim_{t\to\infty} \gamma_t = 1$ ). Capital allocation biases toward data-poor tasks, balancing data accumulation across the economy.
Partial Automation ( $\sigma > 1/\eta$ ): If tasks are strong substitutes and data exhibits diminishing returns, the economy may remain partially automated. Capital concentrates in a small set of "data-rich" tasks, creating a feedback loop where data-rich tasks attract more capital, while data-poor tasks are neglected. The ratio of data between rich and poor tasks diverges to infinity.

B. Wages and the "Training Own Replacement" Mechanism

Wage Stagnation: In the full automation regime ( $\sigma \le 1/\eta$ ), long-run wages stagnate. While data-augmented capital increases productivity, the displacement of labor onto a shrinking set of tasks increases the effective labor supply, offsetting productivity gains.
Unbounded Wage Growth: In the partial automation regime ( $\sigma > 1/\eta$ ), wages grow unboundedly because the automation boundary is bounded, preventing the complete displacement of labor.
Mechanism: The paper highlights a paradox: cross-task complementarities, which typically raise wages in partial equilibrium, lead to wage stagnation in this dynamic setting because they induce the production of "data-poor" tasks, accelerating the accumulation of data that eventually automates those very tasks.

C. Speed of Automation

Even when full automation is inevitable, the transition is slow.

Power Law Decay: The share of tasks produced by labor ( $1-\gamma_t$ ) decays as a power law in time ( $t^{-\eta/(1-\eta)}$ ).
Fat Tail: There remains a "fat tail" of tasks performed by labor for a long time, not because capital cannot perform them, but because the marginal cost of capital (due to limited data) exceeds the wage.
Spillovers and Speed: While cross-task spillovers ensure full automation regardless of $\sigma$ (Proposition 5), they do not fundamentally alter the long-run power-law speed. However, in the short run, network structures (e.g., core-periphery) can create imbalanced dynamics where a "core" of sectors automates rapidly while the "periphery" lags due to bottlenecks in data flow.

D. Efficiency

The decentralized equilibrium is generically inefficient.

Direction of Inefficiency:
- When $\sigma < 1/\eta$ , the market under-invests in data-poor tasks relative to the planner. The planner would tilt capital further toward data-poor tasks to accelerate convergence.
- When $\sigma > 1/\eta$ , the market under-invests in data-rich tasks. The planner would tilt capital further toward data-rich tasks to exploit the data feedback loop.
Cause: Firms do not internalize the positive externalities of their production on the aggregate data stock (both for their own task and for others via spillovers).

E. Explosive Growth (The Singularity)

When endogenous capital accumulation is introduced:

Data-Driven Singularity: The combination of data accumulation (non-depreciating, diminishing returns) and capital accumulation (depreciating) leads to explosive growth. Output and consumption become unbounded in finite time.
Distinction from Nordhaus (2021): Unlike the "supply-side singularity" which relies on exogenous exponential growth in capital productivity, this result stems from the endogenous feedback loop. Crucially, this result holds even if tasks are gross complements ( $\sigma < 1$ ), as the automation boundary expands endogenously to make capital and labor effectively perfect substitutes at the task level.

4. Significance and Claims

The paper claims to provide a rigorous theoretical framework for understanding the unique dynamics of AI-driven automation, distinguishing it from previous models of technological change.

Data as a Distinct Factor: The authors argue that data is not merely a prediction tool (as in Jones and Tonetti, 2020) but a capital-augmenting input that accumulates endogenously, creating a feedback loop that drives the automation frontier.
Reversal of Intuition: The paper challenges standard intuitions regarding wages and efficiency. It shows that cross-task complementarities can lead to wage stagnation and that market forces can distort the direction of data accumulation in ways that depend non-monotonically on substitutability.
Slow Transition: Contrary to narratives of rapid, immediate displacement, the model predicts a slow, power-law transition with a persistent "fat tail" of human labor, driven by the time required to accumulate sufficient task-specific data.
Network Effects: The introduction of graphons allows for the analysis of how the structure of knowledge (transfer learning) dictates the speed and balance of automation, explaining potential "core-periphery" disparities in the short run.

The authors conclude that while data-driven automation has the potential for explosive growth and full automation, the transition path is characterized by complex dynamics, potential inefficiencies, and a prolonged period where labor remains essential for generating the data required for its own replacement.