In the early 19th century Lancashire, it was a widely accepted truth that young individuals could secure jobs as weaving apprentices. During this era of cottage industry, families typically operated a single handloom. However, the advent of mechanized wool spinning ushered in a plethora of opportunities for eager youngsters prepared to enhance their skills.
The journey of an apprentice often begins in frustration. A master weaver can accomplish tasks at double the speed and quality of an apprentice. They set up looms swiftly, detect faults in fabric more readily, and produce twice the output daily. By all metrics, the apprentice appears to be the less competent worker. Yet, the master seldom spends time preparing bobbins. Each hour dedicated to winding yarn is an hour lost in the loom, where only a master can keep pace with the merchant’s demands. The apprentice winds bobbins throughout the day—not due to incompetence, but because their time is simply less valuable when spent in this manner.
The master possesses an absolute advantage in every area. Conversely, the apprentice has a comparative advantage in the art of bobbin winding, as the opportunity cost associated with the apprentice’s time is lower. This insight, which David Ricardo first articulated in 1817, remains a cornerstone of economic theory. It posits that even if one party excels in every aspect, both parties benefit when they specialize according to their comparative advantages.
Can we substitute machines for the master?
The anxiety surrounding AI often stems from its absolute advantages. Large Language Models (LLMs) can produce clear and persuasive writing, swiftly summarize extensive documents, and generate competent Python scripts in mere seconds. In these specific tasks, AI poses a direct threat to human workers. If a job is viewed merely as a collection of such tasks, the future for human employees looks bleak.
However, the Ricardian dilemma invites us to explore where AI holds a comparative advantage and if this translates to the job market. Comparative advantage hinges on opportunity costs. For humans, the limiting factor is time; for AI, it is computational power. These constraints are distinct enough to ensure that humans still play a crucial role in the workforce.
Consider the case of radiologists. Research by Agarwal et al. (2024) indicates that self-supervised algorithms have outperformed human radiologists in reading chest X-rays, including for rare diseases. Here, AI competes in the specific realm of image interpretation, showcasing its comparative advantage—its opportunity cost of performing complex pattern-matching tasks is significantly lower than that of humans. Nonetheless, the algorithm does not provide treatment recommendations or decisions. A radiologist still plays a vital role in patient communication, coordination with clinicians, and exercising contextual judgment regarding necessary interventions.
In this larger professional landscape, AI serves more as an enhancement than a competitor. The opportunity cost for a radiologist engaged in high-context tasks is relatively low when compared to that of AI, which could instead be diagnosing thousands of other scans. As machines take over routine tasks, they enhance human comparative advantages in judgment. The optimal allocation of labor involves a continual reallocation, where machines handle the tasks that benefit from lower computational costs, leaving humans to excel where human time is a more efficient resource.
Should we be concerned regardless?
While comparative advantage suggests that both parties gain from trade, it does not address the distribution of these gains. If computational power becomes exceedingly inexpensive, the wage floor for human workers could diminish correspondingly. A model developed by Restrepo (2025) illustrates that wages could converge with the cost of the computational resources required to replicate human skills. If the expense of digital labor approaches zero, the labor income share of GDP could decline as well.
This scenario sounds alarming, but the phrase ‘without limit’ carries significant weight. The Stanford HAI 2025 AI Index Report notes that the operational costs for a GPT-3.5-level system plummeted 280-fold between 2022 and 2024. Nevertheless, we may be nearing the physical and economic limits of cheap computation.
- Physical constraints. We are approaching the atomic limits of hardware capabilities. Present-day chips possess gate pitches around 48 nanometers. The smallest physically possible transistor gate measures approximately 0.34 nanometers, equivalent to the width of a single carbon atom. The remaining gap from current designs to this atomic limit suggests a potential 140-fold improvement in density, which is less than the cost reductions already achieved in the preceding two years.
- Energy and demand-side factors. No level of software innovation can eliminate the fundamental need for land, capital, and electricity. As unit costs decline, the overall demand for computing resources tends to escalate, unlocking new applications that keep computational power relatively scarce compared to human labor.
Ultimately, the boundary between AI as a competitor and AI as a tool is shaped by the evolving landscape of comparative advantage. While machines may replace humans in routine tasks where they excel, the physical and economic limitations of computation compel them to specialize, rendering them instruments that enhance human judgment.
By relinquishing tasks where machines hold an upper hand, we can redirect our focus toward high-context roles where human intuition remains the most effective resource: judgment, physical presence, and creative improvisation. We find ourselves in a modern narrative reminiscent of the Industrial Revolution. Today’s worker retains their worth by adapting within a rapidly shifting division of labor, albeit at an unprecedented pace.
