It seems like every week, there is a new post proclaiming that eventually AI will be better than humans at everything. This is generally followed up by a supposed corollary that this means everyone will lose their job and the machines will do everything. These arguments come in a variety of forms. Sometimes those making these arguments suggest that this will bring about a sort of utopia, in which humans can free ride off of the machines, enjoy the abundance, and dedicate their time to art and golf and reading great books. Other times, those making these arguments are describing a type of dystopia. The machines take over. People don’t have jobs and cannot survive without government assistance.
Should we fear such scenarios?
No. There are some very fundamental misunderstandings of price theory buried in these claims. Today, I would like to work through them and help you, the loyal readers of Economic Forces, navigate these claims.
Comparative Advantage
The first problem that I have with these types of “the machines will do everything” arguments is that they rely on the idea of absolute advantage. The machines will be better than humans at everything. That is an absolute advantage. But the very basis for specialization and trade is not absolute advantage.
What we really care about when determining who should produce what are opportunity costs. The concept of comparative advantage says that you should specialize in the type of production you can do at the lowest opportunity cost. What does this mean? It might help to work with a particular example.
Imagine that there are two people, Adam and Betty. There are two possible goods that they can produce: carrots and bananas. Suppose that if Betty devotes all her time to carrots, she can grow 10 carrots. If she devotes all her time to bananas, she can grow 5 bananas. Adam, on the other hand, can only grow 8 carrots if he devotes all his time to carrots and 2 bananas if he produces only bananas.
Clearly, Betty has an absolute advantage. She is better at growing both carrots and bananas. If absolute advantage is what mattered, then Betty would produce everything. However, think about opportunity costs. If Betty devotes all her time to bananas, she is giving up 2 carrots for every banana she grows. If Adam devotes all his time to bananas, he is giving up 4 carrots for every banana he grows. In other words, the cost of growing bananas is higher for Adam because the opportunity cost (what he is giving up) is much higher. Thus, Adam should grow carrots and Betty should grow bananas.
Now, imagine that Adam discovers some new technology that makes him 10 times more productive at growing everything. Now, he can grow 80 carrots or 20 bananas. He now has an absolute advantage over Betty. Nonetheless, the opportunity cost of growing 1 banana is still 4 carrots. He should still produce carrots and Betty should grow bananas.
We can generalize this to the example to think about machines and workers. Let the cost of labor be the wage given by w and the cost of compute be given by the per unit cost c. Furthermore, suppose that for any task j, the productivity of workers is given as L(j). Correspondingly, the productivity of AI is given as A(j). It follows that workers should perform the task whenever w/L(j) < c/A(j), or when they have a comparative advantage. As long as this condition holds for some tasks, humans will not be completely replaced by machines. Nonetheless, there is more to the argument that I need to discuss.
AI, the Number of Tasks, and Prices
Okay, we have established that specialization depends on the relative costs of production. When it is cheaper to use humans, you use humans. When it is cheaper to use AI, you use AI. That is comparative advantage. But we still need to think about what happens as things change. In particular, we need to think about what happens as AI gets more productive.
One argument that people seem to make is as follows. Suppose the productivity of AI continues to rise. If A(j) rises for all j, then AI will become cheaper and will take over tasks that otherwise had been done by humans. Carried to its logical conclusion if this continues indefinitely, humans will eventually do nothing.
A related argument that you hear people make is based on the number of tasks. In our original example, we had a two-person, two-task model. Adam and Betty could produce carrots and bananas. They specialize in what they can do at the lowest opportunity cost. But imagine there is a third person, Chris, and yet there are still only two tasks. Suppose that Chris doesn’t have a lower opportunity cost of producing either of the two goods. Doesn’t that imply that Chris has nothing to do? (Hint: No! Wages adjust.)
Similarly, in my generalized example, if the productivity of AI is rising in every single task, this would seem to imply that eventually AI could do every task at a lower cost. But is that correct? If not, what are we missing?
There are reasons to doubt both of these arguments.
The first is that the number of tasks in my examples has been fixed by assumption. In reality, the number of tasks is endogenous. There is a famous saying in economics that “the division of labor [specialization] is limited by the extent of the market.” In the case of rising AI productivity, society would be producing more output and generating more aggregate income. That tends to produce incentives to split existing tasks into finer tasks. Inventions and innovations lead to the production of new goods, which require new tasks to complete.
But there is another reason to doubt these claims. To this point, I have assumed that wages and the cost of compute remain constant as the productivity of AI rises. Of course, we know that this isn’t the case. As AI models improve and AI gets better at particular tasks, this will tend to increase the demand for compute and reduce the demand for labor. The net result is the relative cost of labor will decline (i.e., wages will fall relative to compute).
As the relative cost of labor declines, this ensures that labor will still have a comparative advantage in some tasks. It is also likely that new labor-intensive tasks will pop up since the declining relative cost of labor provides an incentive for people to find new labor-intensive tasks in production.
In addition, arguments that as AI productivity rises one can just spin up as many AI agents as there are tasks don’t seem to hold up. New AI agents increase the demand for compute, which also tends to bid up the cost of compute relative to wages, which reinforces the mechanism that I just described.
Thus, it certainly seems like prices will adjust to ensure that labor still has a comparative advantage in some tasks. Rising AI productivity alone is not enough to wipe away the comparative advantage of humans. Prices have an important role here.
Nonetheless, there is one remaining issue to consider. In this example of rising AI productivity in all tasks, the stabilizing force for comparative advantage is the relative decline in wages. Thus far, I have limited this to the decline in wages relative to the cost of compute. However, we also need to consider what is happening to labor income relative to output, or total income.
The Elasticity of Substitution and the Labor Share of Income
One argument that people worried about AI taking over all of the jobs might make is that labor might retain a comparative advantage, but if this occurs because wages are declining, one might reach a point where labor income reaches the subsistence level. Comparative advantage is hardly something to hang one’s hat on if everyone is living at subsistence.
This means that we need to think carefully about the endogeneity of the tasks. We also need to think about things like balanced growth paths and the conditions necessary for that.
Consider the investment decisions of firms. Suppose that firms have two choices. They can either invest in automation or they can invest in new labor-intensive tasks. Firms will tend to invest in automation when the rate of return to automation is higher than the rate of return for new labor-intensive tasks and invest in labor-intensive tasks when the opposite is true. These alternate investments can lead to a balanced growth path. For example, investments in automation might lead to fewer tasks being done by labor. The decline in the demand for labor reduces the relative wage. However, as the relative wage declines, this increases the rate of return to investments in labor-intensive tasks. Under certain conditions, this can lead to a balanced growth path in which labor’s share of income is constant. That means that even with greater levels of automation driven by AI, the dystopian future in which AI takes all the jobs and wages fall to subsistence levels or humans get completely driven out of production doesn’t actually happen. This is a dramatically simplified version of the paper by Daron Acemoglu and Pascual Restrepo.
We can also think about this in terms of the elasticity of substitution, as I’ve discussed in a previous post. We need to think about the distinct concepts of production and tasks. Sometimes these can be the same thing. Production is a single task. However, production isn’t typically limited to one particular input or one particular task. Production typically consists of a lot of complementary tasks. Framing everything as AI or labor ignores the complementarity of inputs in the production process. As I detailed in that post, when some tasks can’t be automated and tasks are complements in production, automation doesn’t cause growth to explode or the labor share of income to decline toward subsistence, or even zero. What happens is that the labor tasks become the limiting factor, or the weak link, in the production process and the total amount of production
What matters is if the elasticity of substitution is sufficiently small such that automated tasks and labor-intensive tasks can be considered complements. As long as this is the case, even in a world in which the use of automation explodes, labor-intensive tasks will be the limiting factor. You don’t get explosive growth, and the labor share of income remains constant along the balanced growth path.
Concluding Thoughts
Many of the arguments suggesting that AI is going to take all of the jobs (driving the labor share of income to zero), or at least drive labor income to subsistence levels, are motivated by appeals to the absolute advantage of AI. If AI is better at everything, then there is no need for humans, or so it is claimed.
However, my point in writing this post is that we rarely care about absolute advantage. What we care about in economics is comparative advantage. As I’ve argued, rising AI productivity is unlikely to wipe out labor’s comparative advantage in many tasks because wages and compute costs adjust to changing demand for compute and labor. Furthermore, to the extent to which rising AI productivity leads to lower wages, this provides an incentive for firms to invest in new, labor-intensive tasks.
Nonetheless, comparative advantage alone isn’t necessarily enough to save us. If comparative advantage is maintained through decline labor costs, that could potentially lead to scenarios in which labor has a comparative advantage but wages are near subsistence level.
When we consider these things within the context of general equilibrium and we recognize that production isn’t just one task, but involves a series of complementary tasks, then it becomes possible not only to make the case for the continued comparative advantage of labor, but also to say that labor’s share of income can remain stable as long as new tasks arise in which humans hold a durable comparative advantage. These are likely to be tasks that rest on institutions and preferences rather than technology.
Finally, it is also worth clarifying the differences between the utopian and dystopian scenarios associated with AI taking all the jobs that I presented in the introduction. Those scenarios take for granted that AI will take all of the jobs. The only difference between them is whether households own the compute. If some of the other points I raised in the post are true, like labor retains a comparative advantage, that some things can’t be automated, and that human and automated tasks are complements in production, then this is largely a moot point.

