Task-Based Models Show Where AI Substitutes for and Complements Human Labor

For artificial intelligence to fundamentally reshape the global economy, a specific mathematical condition must hold: the technology must either drastically reduce the cost of performing existing tasks or generate entirely new categories of human labor. If it does neither, its macroeconomic impact remains a rounding error. Currently, the empirical evidence suggests this constraint is only weakly satisfied. While large language models can write code and draft emails, the actual cost savings across the broader economy remain marginal, limiting AI's immediate disruptive potential.

To understand why, labor economists rely on the "task-based model," a framework that decomposes jobs into discrete, measurable activities rather than treating occupations as monolithic wholes. Originally formalized by economists Daron Acemoglu and Pascual Restrepo, this model has become the standard tool for predicting how automation affects employment.[1]

The model operates on a simple premise: a job is not a single action, but a bundle of tasks. A radiologist, for example, analyzes images, consults with patients, and updates medical records. AI does not replace the "radiologist"; it automates the specific task of image analysis. By breaking jobs down to this atomic level, economists can map exactly where technology substitutes for human labor and where it complements it.

According to the framework, automation introduces a "displacement effect." When an algorithm takes over a task previously performed by a human, labor demand in that specific area drops. Michael Webb's analysis of AI's labor market impact, which uses natural language processing to match AI patents to job descriptions, indicates that highly educated, white-collar workers face the highest exposure to this displacement.[2]

Displacement is only half the equation. The task-based model also accounts for a "productivity effect." When AI reduces the cost of a task, the overall cost of production falls. This cost savings can increase demand for the final product, which in turn drives up demand for the non-automated tasks within that same industry. If AI makes legal research 30% cheaper, law firms might take on more clients, increasing the demand for lawyers to perform client consultations and courtroom advocacy.

The Federal Reserve Bank of San Francisco highlights a third mechanism: job transformation and specialization. As AI absorbs routine cognitive tasks, workers are forced to specialize in areas where human judgment remains superior. Their November 2025 research indicates that this specialization can mitigate job losses, provided workers have the flexibility and training to transition into new roles.[3]

The most critical variable in the task-based model is the creation of "new tasks." Historically, technological revolutions—from the power loom to the personal computer—have generated entirely new activities that require human labor. If AI creates enough new, labor-intensive tasks, it can offset the displacement effect. The current debate centers on whether generative AI is primarily an automating technology or an augmenting one that will spawn new industries.

Quantifying these effects remains notoriously difficult because task frequencies are rarely measured in real-time. To bridge this gap, researchers at Anthropic developed a new measure to project AI's economic consequences between 2026 and 2030. Their framework maps future paths of AI capabilities into implied paths for GDP, wages, and unemployment.[4]

The results are surprisingly muted. In Anthropic's "modest change" scenario, the researchers note that "AI adds less than half a point to GDP growth by 2030 and raises unemployment by a tenth of a point." This aligns closely with Acemoglu's May 2024 macroeconomic assessments, which conclude that predicted TFP gains over the next 10 years are "modest and are predicted to be less than 0.53%."[4][5]

These projections sit in sharp contrast to the transformative claims made by leading AI developers, who often anticipate the automation of most cognitive work within years. The discrepancy stems from the task-based model's reliance on empirical data from "easy-to-learn" tasks. The model assumes that automating complex, context-dependent tasks—where human judgment is paramount—will be significantly harder and yield diminishing returns.[5]

Furthermore, the task-based model reveals a troubling dynamic regarding wage inequality. Even when AI improves the productivity of low-skill workers in certain tasks, it may widen the gap between capital and labor income. If the primary benefit of AI is cost savings for corporations, the financial gains will accrue disproportionately to capital owners rather than the workforce.[5]

The deciding factor over the next five years will be the ratio of task automation to new task creation. If AI developers can push the technology past the barrier of "hard-to-learn" tasks without simultaneously generating new roles for displaced workers, the displacement effect will overwhelm the productivity gains. Until that threshold is crossed, the macroeconomic data will continue to reflect a slow reallocation of labor rather than a sudden paradigm shift.