In the 1800s, a young person in Lancashire faced a significant opportunity: they could secure employment as a weaving apprentice. During this time, the cottage industry dominated, with families typically owning a single handloom. The rise of mechanized wool spinning opened the door to numerous jobs for ambitious individuals eager to enhance their skills.
The apprenticeship journey often begins with a sense of frustration. A master weaver can complete all necessary tasks not only faster but also more efficiently than the apprentice. They can set up a loom and identify defects in the fabric with remarkable speed, producing double the yardage each day. By all traditional measures, the apprentice appears to be the less capable worker. However, the master weaver does not spend their time winding bobbins each morning. Time spent on this task detracts from their ability to keep up with the production rate required by merchants. Thus, the apprentice is tasked with bobbin winding throughout the day—not due to incompetence, but because their labor is less costly in that context.
The master possesses an absolute advantage in every aspect of the work, while the apprentice holds a comparative advantage in bobbin winding due to the lower opportunity cost of their time. This key distinction, first articulated by David Ricardo in 1817, is a cornerstone principle in economics. Even when one party excels at every task, both benefit when they each engage in work where they hold a comparative advantage.
Can we replace the master with the machine?
Much of the concern surrounding artificial intelligence (AI) centers on its absolute advantages. Large Language Models (LLMs) can draft text with clarity and persuasiveness, summarize extensive documents efficiently, and generate acceptable Python scripts in moments. In these discrete functions, AI acts as a direct competitor to human workers. If jobs consist solely of such tasks, human workers may be at risk.
However, the real challenge is to pinpoint where AI possesses a comparative advantage and whether this is observable at the job level. Comparative advantage hinges on opportunity costs. For humans, the critical constraint is time, while for AI, it is computational power. These differing constraints are significant enough to ensure that humans remain relevant in the workforce.
Consider radiologists. Research by Agarwal et al. (2024) demonstrated that self-supervised algorithms can outperform human radiologists in reading chest X-rays, even in the case of rare diseases. Here, AI functions as a competitor solely in the domain of image interpretation, showcasing its comparative advantage due to the lower opportunity cost associated with AI’s capacity to perform numerous pattern-matching tasks simultaneously. However, the AI’s output does not provide treatment recommendations or clinical decisions. A radiologist still has crucial roles that involve patient communication, clinician coordination, and contextual judgement to ascertain whether an anomaly requires intervention.
In this broader context, AI serves more as a tool than a direct competitor. The opportunity cost of radiologists performing high-context tasks is comparatively low versus that of AI, which could instead process thousands of additional scans. Even as machines take over routine tasks, they enhance human comparative advantages in judgement. Optimal task allocation requires a continual reallocation of responsibilities: machines handle the tasks best suited for their computational efficiency, while humans focus on areas where their time offers greater value.
Should we worry anyway?
While comparative advantage illustrates how two parties can benefit from trade, it does not address how those benefits are distributed. If the cost of computation becomes sufficiently low, it can drive down the wage floor for human workers. Research by Restrepo (2025) proposes a model showing that wages can converge to the cost of computing needed to replicate human skills. If the expenses of a digital workforce approach zero, the share of labor income in GDP is likely to decrease.
Although this scenario is alarming, the phrase ‘without limit’ in that context requires careful consideration. The Stanford HAI 2025 AI Index Report found that the costs associated with operating a GPT-3.5-level system fell by 280 times between 2022 and 2024. Nonetheless, we are nearing the physical and economic limits of inexpensive computation.
- Physical constraints. We are approaching the atomic limits of hardware. Today’s chips are built with gate pitches around 48 nanometers. The smallest possible transistor gate is around 0.34 nanometers, equivalent to the width of a single carbon atom. The remaining gap from current designs to this limit offers approximately a 140-fold increase in density, which is less than the cost reduction we have achieved in recent years.
- Energy and demand. No amount of clever software can eliminate the necessity for land, capital, and electricity. As unit costs decrease, total demand for computational power tends to rise faster, unearthing new applications that keep computation relatively scarce compared to human labor.
Ultimately, the distinction between AI as a rival and as a tool lies in the evolving boundaries of comparative advantage. While machines may replace us in routine tasks where they excel, the physical and economic limitations of computation compel them to specialize, enhancing their role as tools that amplify human judgement.
By relinquishing responsibilities where machines perform better, we can dedicate our time to high-context roles where human intuition is the most effective resource—such as judgement, personal interaction, and creative problem-solving. We continue to navigate the narrative established during the Industrial Revolution. Today’s worker adds value by adapting within an increasingly dynamic division of labor, with the rate of that adaptation growing faster than ever before.
