It was a widely recognized fact that in Lancashire in the 1800s, a young man or woman could find gainful employment as a weaving apprentice. In pre-factory cottage industries, weaving families usually owned one handloom. The dawn of mechanized wool spinning created many jobs for young people looking to upskill.
The typical apprenticeship experience begins with a setback. A master weaver can do everything an apprentice can do, but twice as fast and better. Set up your loom faster, find fabric defects faster, and double your production yards per day. No matter how you look at it, the intern is an inferior worker. But the master did not spend the morning preparing the bobbin. All the time spent winding thread is time not spent on the loom, and only a skilled weaver can maintain the pace demanded by the tradesmen. Apprentices wind bobbins all day long, not because they’re bad at it, but because it wastes less time.
Masters have absolute superiority in every task. Apprentices have a relative advantage in bobbin winding because the opportunity cost of their time is low. This distinction was first formalized by David Ricardo in 1817 and is one of the most powerful achievements in economics. Even if one party is better at everything, both will be better off if each specializes according to their comparative advantage.
Can I replace the master with a machine?
Much of the panic surrounding AI still revolves around pointing out the absolute benefits. LLMs can write clearly and persuasively. Quickly summarize large documents. Generate flawless Python scripts in seconds. For these individual tasks, AI becomes a direct competitor. Human workers have a problem when a job is just a collection of such tasks.
But Ricardo’s challenge is to identify where AI has a comparative advantage and whether it manifests itself at the job level. Comparative advantage is determined by opportunity cost. For humans, the binding constraint is time. For AI, the constraint is computing. These are completely different constraints, and they are different enough to keep humans in perspective.
Get a radiologist. Agarwal et al. (2024) showed that a self-supervised algorithm outperformed human radiologists in reading chest radiographs, even for rare diseases. Here, AI acts as a competitor for the specific task of image interpretation and demonstrates a comparative advantage. This means that the opportunity cost of having an AI perform many pattern matching exercises is much lower than it would be for a human. However, the output of the algorithm does not provide recommendations or treatment decisions. Radiologists still communicate with patients, coordinate with clinicians, and make context-specific judgments about whether an abnormality warrants intervention.
In this broader professional context, AI is more of a tool than a direct competitor. The opportunity cost of performing high-context tasks for radiologists is low compared to the opportunity cost of AI. This is because the same compute can diagnose thousands of other scans instead. Even as machines replace humans in daily tasks, they amplify humans’ comparative advantage in decision-making. Correct division of labor involves continuous redistribution. Machines will take on tasks where computing is cheap, and humans will specialize in areas where human time can be most efficiently used.
Should I be worried anyway?
Comparative advantage tells us that two agents benefit from trade, but it tells us nothing about how the profits are divided. When the cost of computing becomes low enough, the minimum wage for human workers will fall with it. Restrepo (2025) develops a model showing that wages converge to the computational cost required to reproduce human skills. As the cost of digital labor falls toward zero, the share of labor income in GDP will fall with it.
As scary as it sounds, “unlimited” plays a very important role in this sentence. According to the Stanford HAI 2025 AI Index report, the cost of running a GPT-3.5 level system has fallen by a factor of 280 between 2022 and 2024. But we may be approaching the physical and economic limits of cheap computing.
Physical constraints. We are approaching the atomic limits of hardware. Current chips have a gate pitch of about 48 nanometers. The smallest physically possible transistor gate is about 0.34 nanometers, which is the width of one carbon atom. Over the remaining distance from the current design to the atomic limit, density will improve by a factor of about 140, which is less than the cost savings already achieved over the past two years. Energy and demand side. No amount of clever software can eliminate the need for land, capital, or electricity. And as unit costs fall, aggregate demand for computing will grow faster, unlocking new use cases where computing continues to be scarce relative to human labor.
Ultimately, the difference between AI as a competitor and AI as a tool is defined by the shifting boundaries of comparative advantage. Machines will replace us in routine tasks where they have an absolute advantage, but the physical and economic scarcity of computing will force them to specialize, turning them into tools that amplify human judgment.
By abandoning tasks for which machines are superior competitors, we focus our time on high-context roles where human intuition is the most efficient input, such as judgment, physical presence, and creative improvisation. We are still living the story of the Industrial Revolution. Modern workers maintain their value by relocating within an increasingly fluid division of labor, and this relocation is now occurring faster than ever before.
