Can AI Crash the Economy?

We often assume that innovation leads to economic growth. But that's not universally true. What kinds of innovation are the right kind that grow prosperity for everyone? Let's look at a very simple economic model to understand what kinds of innovation lead to economic growth and what types of innovation have to potential to shrink growth or even cause an economic crash.

With the widespread adoption of large language models (LLMs) and other types of generative AI, there has been speculation about what the impact on the economy will actually be. Anthropic, an AI vendor, has claimed that annual GDP growth rate will at least double. Some more outlandish predictions (without any evidence) suggest we might see annual doubling (i.e. >100% annual growth) of GDP by 2030. A recent article in Ars Technica takes another look at that work by Anthropic; needless to say there were some problems.

In this article, I'll provide some intuition regarding what conditions are required to achieve economic growth.

What we'll see is that wage growth is a key indicator. A new technology cannot yield economic growth overall without also yielding average wage growth. Looking at preliminary data, wages seem to be falling in response to the advent of large language models. This is a very worrying sign.

The models used below are highly simplified. None of this is intended to make a prediction about the future. But looking at these simple models and understanding their assumptions can help us form an intuition about the impact of innovation.

Let's dive in.

Classical economic growth

In classical economic growth models, it is often assumed that consumption will just always grow. That is, all good produced can also be sold. Indirectly, this assumes that wages also grow, otherwise, there is not enough money to buy those goods.

The reason for this wage growth is assumed to be innovation: If new industries form to create new products, consumption will shift towards the new superior product. To produce the new superior product, people are needed and the new industry can afford to pay people more.

But what if innovation is purely an improvement on the (labor-) efficiency of production in existing industries and does not lead to wage growth?

An extreme case

Let's first examine an extreme case to form a bit of intuition. Suppose that all goods can be produced without labor. This is the full automation scenario: everything is made by robots/AI. In this scenario humans have no form of income and therefore cannot buy anything. Therefore there is no demand, no revenue for producers, no returns on any investments, the economy as we know it would cease to exist.

It's a possibility that prices would drop to zero and we all live in a utopia, but that's a topic for another day.

If, one day, automation can increase production beyond the ability or desire to consume, it is not clear how a capitalist system can survive. Aggressive redistribution of wealth will be necessary.

At least we can see from this example that extreme efficiency increase would, counterintuitively, lead to economic collapse.

The simplest model for labor and capital

If you want to follow along with the computations you can see the code for the models and the plots in this GitHub repo.

Let's see if we can quantify this with a basic model. We will start with the Solow model for GDP growth. In a nutshell, this model assumes that production requires both labor and capital. Capital in this context should be read as all "stuff" including factories, machines, raw materials etc.

The model makes the following assumptions:

1. Increased labor increases production, so does increased capital.
2. Diminishing returns: the increased productivity is smaller for each consecutive increase in either capital or labor.
3. No economies of scale: Doubling both labor and capital doubles production.
4. The economy is closed, no external inflows of capital or labor.
5. Capital depreciates at a constant rate
6. The fraction of production that is "saved" to increase capital is fixed.

The above assumptions yield what's usually called the Cobb-Douglas function for production.

Figure 1: Two side-by-side plots. Left: Economic output as a function of Capital. The plot shows increasing output as a function of capital but at a diminishing rate. Right: a similar relationship is shown between output and labor.

An interesting finding that follows from the above assumptions is that, in the absence of growth in labor or efficiency, production eventually equals capital depreciation and an equilibrium is reached. The equilibrium is independent of the starting capital as we can see in the figure below

Figure 2: The figure shows economic output as a function of time. The output converges to the same equilibrium regardless of starting capital.

A shock to equilibrium

What happens if a change is introduced suddenly to an economy that was in equilibrium before. For example, a new automation technology is invented and introduced suddenly over a few time-periods. The figure below shows a really interesting behavior.

The figure shows that production increases, first fast and then more gradually as capital increases to match the new labor efficiency. Notably wages go up in tandem with the increased productivity. That is, workers are paid for the increased value they provide. Interestingly, wages slightly decrease after the initial increase because some revenue is needed to offset the higher cost of capital as capital increases.

Increased efficiency without wage increase

Up to now, we've assumed that wages will grow in line with the increased production per worker. But what happens we we remove that assumption. What we see in the real world at the moment is that firms claim efficiency increases but real wages are stagnant. Layoffs seem to indicate lower labor requirements at least in some industries.

We can model this by making a slight modification to the Solow model: we assume that production at time t is capped by the aggregate income of all consumers at time (t-1). That is, firms must sell to buyers with money and buyers cannot spend what they do not have. The results are starkly different from before.

The figure shows near-total economic collapse. Lower labor requirements lead to lower employment, which leads to lower demand which leads to even lower labor requirements. A runaway spiral.

There is a lot more to say about this but the simple message is this: The kind of innovation that spurs growth is the kind of innovation that yields new, better quality products. Because it's consumer preference for new product that spurs wage growth.

Wage inequality is not included in this model. It would be interesting to ask if an economy with high inequality can mask this effect. Is it possible for the measured average income to go up while the majority of people experience contraction?

Are LLMs special?

The current situation with LLMs does seem different from previous waves in automation. At least the claim by vendors is that LLMs can be used for nearly any task and therefore the potential impact is greater. But importantly, LLMs are equally available to all participants in the market. That is, it's hard to create a competitive advantage with AI because all your competitors have access to the same tech at the same price. Companies feel a fear of missing out, but creating new successful products is really hard. That's why we're seeing so many companies just focus on automating existing processes.

What does this mean for the future?

The conclusions of this very simple model are broadly similar to the findings of the Nobel Prize winning team of Daron Acemoglu that the "wrong" kind of automation can lead to a decrease in social welfare. The findings also broadly align with efficiency wage theory.

All this means we should not assume that AI will be beneficial for all, regardless what kind of tech we develop and how we use it. We're not on an unchangeable path. We should not assume that LLMs and other automation is used in ways that benefit society. We can make better choices.

---

Next Up: my next article we'll look at the same problem from the perspective of individual firms. We'll be able to quantify the limits of Jevons paradox in the context of AI. Surprisingly, we'll see that increased efficiency can shrink a company instead of growing it.