It was a universally accepted fact that a young man or woman in 1800s Lancashire could find gainful employment as an apprentice weaver. In the pre-factory cottage industry, a weaving family will usually have one handloom. With the introduction of mechanized wool spinning, many jobs were available to young people who were willing to develop their skills.
A typical intern’s experience begins with frustration. A master weaver can do anything an apprentice can do, but twice as fast and better. Set up the loom quickly, spot flaws in the fabric quickly, produce twice the yardage a day. By all standards, the student is a poor worker. However, the king did not spend the morning fixing the bobbins. Every hour spent winding yarn is an hour spent on the loom, where only a master weaver can maintain the speed required by the merchant. A student takes a deep breath throughout the day, mostly not because he is wrong, but because his time is wasted that way.
The king has full profit in every work. The apprentice has a comparative advantage in bobbin spinning, because the opportunity cost of the apprentice’s time is low. This distinction, formalized by David Ricardo in 1817, is one of the most powerful results in economics. Even if one group is better at everything, both are better if each is specialized in terms of comparative advantage.
Can we replace the machine with the master?
Much of the panic surrounding AI hinges on revealing the full benefits. LLMs can write clearly and persuasively. They quickly summarize large texts. They generate fast Python scripts in seconds. In these various tasks, AI is a direct competitor. If a job is simply a collection of such jobs, 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 cost. For people, the binding bond is time. In AI, the limit is calculated. These are very different issues, and different enough to keep people in the picture.
Take the radiologists. Agarwal et al. (2024) showed that supervised algorithms outperformed radiologists in reading chest X-rays, even for rare diseases. Here, AI acts as a competitor for a particular task of image interpretation, and shows a comparative advantage, which is the opportunity cost that makes AI perform many pattern matching tests much lower than a human. However, the output of the algorithm does not represent a recommendation or decision for treatment. The radiologist still communicates with the patient, coordinates with the therapists, and uses contextual judgment as to whether the abnormality warrants intervention.
In this broader professional context, AI is more of a tool than a direct competitor. The opportunity cost of a radiologist to perform high-content tasks is low compared to the opportunity cost of AI, because the same computer can instead diagnose thousands of other scans. As machines replace humans in routine tasks, they increase human comparability in judgment. Proper division of labor involves permanent replacement. Machines take jobs where computers are cheap, leaving humans to specialize where human time is the most efficient input.
Should we be worried though?
Comparative advantage tells us that two agents benefit from the trade, but it does not say anything about how the benefits are distributed. If the computer becomes cheap enough, the wages of the workers go down as well. Restrepo (2025) develops a model that shows that wages are linked to the cost of computing needed to replicate human skills. If the cost of a digital worker falls to zero, the share of labor income in GDP falls with it.
That sounds scary, but ‘without limitation’ does a lot of work in that sentence. The Stanford HAI 2025 AI Index Report found that the cost of implementing a GPT-3.5 level system dropped 280 times between 2022 and 2024. But we may be approaching the physical and economic limits of cheap computing.
- Physical barriers. We are approaching the atomic limit of hardware. Today’s chips have gate spacings of about 48 nanometers. The smallest practical gate possible is about 0.34 nanometers, the width of a single carbon atom. The entire distance from current designs to the atomic limit yields about a 140-fold improvement in density, which is less than the cost reductions already achieved in the past two years.
- Power and demand side. There is no software genius that eliminates the need for earth, money, and electricity. And as unit costs fall, the total demand for computing increases rapidly, opening up new use cases that keep computing scarce compared to human labor.
Ultimately, the difference between AI as a competitor and AI as a tool is defined by the dynamic boundary of comparative advantage. While machines are moving us away from the mundane tasks where they hold the absolute edge, the physical and economic limitations of the computer force it to specialize, turning it into tools that augment human judgment.
By offering jobs where the machine is a superior competitor, we focus our time on high content roles where human understanding is always the most effective input: judgement, physical presence and creative improvement. We are still living the story of the Industrial Revolution. The modern worker maintains his value by redeployment within a growing labor class, except now when redeployment is occurring at a faster pace than ever before.
