📈 economics

Pricing Intelligence: Task-Based Learning and Labor Displacement in the AI Economy

This paper analyzes how the interplay between profit maximization and endogenous AI learning on complex tasks creates a tension that determines whether labor displacement follows a convex path or a learning trap, with market competition acting as a key accelerator of this process.

Original authors: Carl-Christian Groh

Published 2026-08-12

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Carl-Christian Groh

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ 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 Great AI Learning Curve

Imagine the economy as a giant, bustling kitchen where millions of people are trying to get dinner ready. Some tasks are simple, like chopping carrots or boiling water; others are tricky, like deciding the perfect spice blend for a new recipe or handling a customer who is furious about a burnt meal. For a long time, humans have been the only chefs capable of handling the tricky stuff, while machines were mostly good at the simple, repetitive chopping. But now, Artificial Intelligence (AI) has entered the kitchen. It's getting incredibly fast at the simple tasks, but it's still learning how to handle the complex, judgment-heavy ones.

The big question economists are asking isn't just "Can AI do the job?" but "How fast will it get better?" This depends on a concept called learning-by-doing. Just like a human chef gets better at chopping onions the more onions they chop, an AI gets better at a specific type of task the more it performs that task. If an AI only ever does simple tasks, it will never learn how to handle the complex ones. This paper explores a fascinating tension: Who actually buys the AI? If the AI is sold only to people who need simple tasks done, the AI gets faster at simple tasks but stays clumsy at complex ones. But if it's sold to people with complex needs, it learns faster, but those customers might be harder to convince to pay. The paper asks: How do the prices AI companies charge and the competition between them change the speed at which AI learns to take over human jobs?

The AI Price Tag and the Learning Trap

In this study, the author, Carl-Christian Groh, builds a mathematical model to see how AI companies (like the giants we know today) decide who gets to use their technology. The story revolves around two types of workers: those who mostly need help with easy tasks (like writing a standard email) and those who need help with complex tasks (like drafting a legal argument for a weird new case).

Here is the twist: In the current "low-tech" state of AI, it is a superstar at easy tasks but still a bit clumsy at complex ones. Because of this, workers with mostly complex tasks don't want to pay much for the AI—they'd rather do the hard thinking themselves. Workers with mostly easy tasks, however, are eager to pay because the AI is already great at what they need.

This creates a tricky situation for the AI company. If the company wants to make the most money right now, it should only sell to the people with easy tasks. But if it does that, the AI never gets the data it needs to learn how to do the complex tasks. It's like a cooking school that only accepts students who want to learn how to boil water; the school makes money, but the students never learn how to cook a gourmet meal.

The paper finds that this tension leads to two very different futures for the economy:

The Convex Takeoff (The Happy Path): If the AI learns very quickly from doing complex tasks, the company eventually realizes it's worth lowering prices to attract the complex-task workers. As more of these workers join, the AI gets better at complex tasks even faster, which attracts even more workers. This creates a "convex" curve—a slow start that suddenly shoots up, leading to rapid job displacement as the AI masters the hard stuff.
The Learning Trap (The Stuck Path): If the AI is better at learning from easy tasks, or if there are very few people with complex tasks, the company will keep ignoring the complex-task workers to maximize short-term profits. The AI gets amazing at boiling water but never learns to cook the gourmet meal. The result? AI adoption stays slow, and humans keep doing the complex jobs for a long time.

The Competition Effect: More Rivals, Faster Learning?

The paper then asks: What happens if there isn't just one AI company, but two or more fighting for customers?

You might think that splitting the customers between two companies would slow things down because each company gets less data to learn from. And in some cases, that's true. If the competition is weak (like two companies that don't really fight hard on price), they both end up charging high prices. This keeps the complex-task workers away, and learning stays slow.

However, the paper shows that if the competition is intense—meaning the companies are fighting hard to lower prices—the story changes. Lower prices make it affordable for the complex-task workers to buy in. Even though each company gets less total data, the type of data they get changes. They get a much higher share of complex-task data. This "task-composition effect" can actually make the AI learn faster than if there were only one monopoly. The fierce price war forces the companies to serve the very users who are most valuable for teaching the AI how to do the hard stuff.

The Specialization Twist

Finally, the paper looks at what happens if the two competing companies start with different strengths. Imagine Company A is naturally better at easy tasks, and Company B is naturally better at complex tasks.

The model suggests that these small initial differences can grow into a massive gap. Company B, being slightly better at complex tasks, will naturally attract the complex-task workers. As it serves them, it gets even better at complex tasks, which attracts even more of those workers. This is called endogenous specialization. It's a self-reinforcing loop where the company that starts with a tiny edge in the "hard stuff" ends up dominating that entire market, learning incredibly fast, and accelerating the displacement of human labor in those complex fields.

The Bottom Line

This paper doesn't just say "AI will take jobs." It suggests that the speed and direction of that takeover depend heavily on business decisions.

Monopoly vs. Competition: A single AI company might get stuck in a "learning trap" if it's too greedy, ignoring complex tasks to milk easy-task profits. But if competition is fierce, it can force companies to serve the complex-task users, speeding up the AI's evolution.
The Learning Trap: If the AI learns slowly from complex tasks, or if the market is dominated by easy-task users, the AI might never learn to do the hard jobs, leaving humans in charge of the most difficult work for a long time.

The author uses simulations to show these paths. The results aren't a guarantee of the future, but they highlight a critical trade-off: The drive to make money today can sometimes slow down the learning that would make the AI useful for the hardest problems tomorrow. Whether we get a rapid, convex takeoff or a slow, stuck learning trap depends on how the market is structured and how well the AI learns from the difficult tasks it faces.

Technical Summary: Pricing Intelligence: Task-Based Learning and Labor Displacement in the AI Economy

Problem Statement
This paper investigates the determinants of the speed and direction of Artificial Intelligence (AI) learning, specifically focusing on how the strategic pricing incentives of AI providers and market competition shape the trajectory of AI capabilities and subsequent labor displacement. The central economic tension identified is that users with a high share of complex tasks are the most valuable for advancing AI's frontier capabilities (via task-specific learning signals) but are the least willing to pay for AI access in the current technological state, where AI holds a comparative advantage only in easy tasks. Conversely, users with primarily easy tasks have a high willingness to pay but provide less valuable data for improving complex-task performance. The paper asks how profit-maximizing providers navigate this trade-off and how market structure (monopoly vs. competition) influences the resulting learning dynamics.

Methodology
The author employs a dynamic microeconomic model of task-based labor displacement, extending the framework of Acemoglu & Restrepo (2018). The model features:

Agents: A unit mass of users with private information regarding their share of complex tasks ( $\theta \in [0,1]$ ) and an AI provider (or providers in duopoly settings).
Tasks: Tasks are binary (easy or complex). Users complete tasks using either labor or AI.
Technology & Learning: AI capabilities ( $b_{e,t}$ for easy, $b_{d,t}$ for complex) evolve endogenously. Learning is task-specific: the AI improves at a task category proportional to the volume of that specific task type completed by its user base. The efficiency of learning is governed by parameters $\rho_e$ and $\rho_d$ .
Pricing: Providers set access prices (or cutoffs) to maximize static, myopic profits. Users self-select based on willingness to pay, which is decreasing in $\theta$ under the assumption that AI is currently superior at easy tasks but inferior at complex tasks.
Market Structures: The analysis proceeds through three benchmarks:
1. Monopoly: A single provider sets prices.
2. Symmetric Competition: Two identical providers compete with switching costs.
3. Asymmetric Competition: Two providers with initial comparative advantages in different task categories.

Key Contributions
The paper contributes to the literature on the economic effects of AI by integrating strategic pricing with endogenous, task-specific learning. Unlike prior works that treat AI capabilities as exogenous or focus solely on screening and welfare, this model explicitly links provider pricing decisions to the composition of the data pool, thereby determining the direction of AI learning. It contrasts with consumption-based learning models where learning depends only on transaction volume, not on the heterogeneity of the consumer base.

Key Results

Monopoly Benchmark: Convex Takeoff vs. Learning Traps
In a monopoly setting, the tension between static profit maximization and complex-task learning generates two distinct dynamic paths:
- Convex Labor Displacement: If the efficiency of complex-task learning ( $\rho_d$ ) is sufficiently high relative to easy-task learning ( $\rho_e$ ), the AI provider eventually expands access to complex-task-intensive users. As the AI improves at complex tasks, the willingness to pay of these users increases, creating a positive feedback loop that accelerates adoption and labor displacement in a convex manner.
- Learning Traps: If easy-task learning is relatively efficient or the distribution of user types is highly right-skewed (few complex-task users), the provider optimally rations access to extract rents from easy-task users. This excludes the very users necessary for complex-task learning, causing the AI's comparative advantage in easy tasks to widen further. This results in a "learning trap" where complex-task learning and labor displacement remain slow or stagnate.
Symmetric Competition: Data Fragmentation vs. Task Composition
Moving from monopoly to symmetric competition introduces two countervailing effects:
- Data Fragmentation: Each provider receives less total data, which slows overall learning.
- Task Composition: Competition lowers prices, expanding the user base to include more complex-task-intensive users. This shifts the composition of data toward complex tasks.
- Net Effect: If competition is sufficiently intense (low switching costs), the task composition effect dominates. The increased relative rate of complex-task learning accelerates the overall speed of AI adoption and labor displacement compared to the monopoly benchmark. If competition is weak, the data fragmentation effect may dominate, slowing adoption.
Asymmetric Competition: Endogenous Specialization
When providers have initial asymmetric capabilities (one better at easy tasks, the other at complex), small initial differences can amplify over time. The provider with an initial advantage in complex tasks attracts a disproportionate share of complex-task-intensive users. This generates a feedback loop where the provider learns faster at complex tasks, further increasing its attractiveness to this segment. This endogenous specialization accelerates the rate of complex-task learning and labor displacement more than in symmetric scenarios.

Significance and Claims
The paper claims to demonstrate that the speed and direction of AI learning are not merely functions of technological potential but are endogenously determined by market incentives. It argues that:

Profit maximization can inadvertently slow the development of AI capabilities in high-value, complex domains if the immediate monetization of easy tasks is more lucrative.
Market structure plays a critical role: intense competition can paradoxically accelerate the displacement of labor by forcing providers to serve the very users (complex-task intensive) who are currently the least profitable but most critical for frontier learning.
The dynamics of AI adoption are non-linear; they can exhibit convex takeoffs or get stuck in learning traps depending on the relative efficiency of learning across task types and the distribution of user needs.

The analysis suggests that the path of AI-driven labor displacement is contingent on the interplay between pricing strategies, the composition of the user base, and the specific learning efficiencies of different task categories.

The Great AI Learning Curve

The AI Price Tag and the Learning Trap

The Competition Effect: More Rivals, Faster Learning?

The Specialization Twist

The Bottom Line

More like this