It was a fact universally acknowledged that a young man or woman in 1800s Lancashire could find gainful employment as a weaving apprentice. In the pre-factory cottage industry, a weaving family would typically own one handloom. With the dawn of mechanised wool spinning, plenty of jobs became available for the young and willing to upskill.
The typical experience of an apprentice begins with frustration. A master weaver can do everything the apprentice can do, but twice as fast and better. Set up the loom faster, spot faults in the cloth sooner, produce twice the yardage per day. By every measure, the apprentice is the inferior worker. Yet the master never spent the morning preparing bobbins. Every hour spent winding yarn is an hour not spent at the loom, where only a master weaver can maintain the pace the merchant demanded. The apprentice winds bobbins all day, specifically not because they are bad at it, but because their time is less costly spent that way.
The master has an absolute advantage in every task. The apprentice has a comparative advantage in bobbin winding, because the opportunity cost of the apprentice’s time is lower. This distinction, first formalised by David Ricardo in 1817, is one of the most powerful results in economics. Even when one party is better at everything, both are better off when each specialises according to comparative advantage.
Can we replace the master with the machine?
Much of the panic around AI rests on pointing out absolute advantages. LLMs can write clearly and convincingly. They summarise large documents quickly. They generate passable Python scripts in seconds. In these discrete tasks, AI is a direct competitor. If a job is merely a collection of such tasks, the human worker is in trouble.
The Ricardian challenge, however, is to identify where AI has a comparative advantage and whether this manifests itself at the job level. Comparative advantage is determined by opportunity costs. For humans, the binding constraint is time. For AI, the constraint is compute. These are very different constraints, and they are different enough to keep humans in the picture.
Take radiologists. Agarwal et al. (2024) showed that self-supervised algorithms have surpassed human radiologists at reading chest X-rays, even for uncommon diseases. Here, AI acts as a competitor for the specific task of image interpretation, and it demonstrates a comparative advantage, namely the opportunity costs of making the AI carry out numerous pattern-matching exercises are much lower than a human’s. However, the algorithm’s output does not yield a recommendation or a treatment decision. A radiologist still communicates with the patient, coordinates with clinicians, and exercises contextual judgement about whether an abnormality warrants intervention.
In this broader professional context, AI is more of a tool than a direct competitor. The radiologist’s opportunity cost of performing high-context tasks is low relative to AI’s opportunity cost, because the same compute could instead be diagnosing thousands of other scans. Even as machines substitute for humans in routine tasks, they amplify human comparative advantage in judgement. The correct division of labour is involves constant reallocation. The machine takes the tasks where compute is cheap, leaving humans to specialise where human time is the more efficient input.
Should we worry anyway?
Comparative advantage tells us that two agents benefit from trade, but it says nothing about how the benefits are distributed. If compute becomes sufficiently cheap, the wage floor for human workers drops with it. Restrepo (2025) develops a model showing that wages converge to the cost of the compute needed to replicate human skills. If the cost of a digital worker falls toward zero, the share of labour income in GDP falls with it.
That sounds terrifying, but ‘without limit’ is doing a lot of work in that sentence. The Stanford HAI 2025 AI Index Report found that the cost of running a GPT-3.5-level system fell 280-fold between 2022 and 2024. But we may be approaching the physical and economic boundaries of cheap compute.
- Physical constraints. We are nearing the atomic limit of hardware. Today’s chips have gate pitches around 48 nanometres. The smallest physically possible transistor gate is about 0.34 nanometres, the width of a single carbon atom. The entire remaining distance from current designs to the atomic limit yields roughly a 140-fold improvement in density, less than the cost reduction already achieved in the past two years.
- Energy and the demand side. No amount of software cleverness eliminates the need for land, capital, and electricity. And as unit costs fall, total demand for compute rises faster, unlocking new use cases that keep compute scarce relative to human labour.
Ultimately, the distinction between AI as a competitor and AI as a tool is defined by the shifting boundary of comparative advantage. While machines displace us in routine tasks where they hold an absolute edge, the physical and economic scarcity of compute forces them to specialise, turning them into instruments that amplify human judgment.
By surrendering the tasks where the machine is a superior rival, we focus our time on high-context roles where human intuition remains the most efficient input: judgment, physical presence and creative improvisation. We are still living the story of the Industrial Revolution. The modern worker maintains their value by repositioning within an increasingly fluid division of labour, except now that repositioning occurs at a faster speed than ever before.
