A Primer on Acemoglu, AI and Automation

When AI Boosts Average Productivity but Reduces Marginal Productivity

Primer on Acemoglu

I sort of crashed last night during the game and I think it is because for two nights straight, I slept in this recliner beside my dad’s bed as I was worried I wouldn’t hear him if he needed me. I’m pretty sure I didn’t really sleep since I decided to go upstairs at halftime. I remember waking up an hour later in the bed around 9pm, but then I guess I fell back asleep until 3:30am.

I tend to sleep much better here than at home. I keep telling myself it’s because I haven’t yet bought blackout shades for the bedroom. But then when I get home, I start to believe that my sleep is always supposed to be this bad and just move on with my life operating at 40-60% rest. Anyway, though, this morning, I was up at 3:30am and started working on this substack because I decided that maybe I should try to write out my “big picture” goals for the AI class on Mondays and not just do a post mortem on Fridays. Or maybe I just do this today because I went to bed thinking about this.

One of my goals for the class is that we learn Acemoglu’s theoretical work on automation and AI in enough detail that they know the model, and understand its implications. That takes us not just through Acemoglu’s writings with Johnson, but it’s actually more about his theoretical work with Restrepo. But it’s a lot to translate this for my students, most of whom are undergrads, which is why this class is a “beta version 1.0” of the longterm class. So this week, I am going to try and help students understand two paragraphs from the Acemoglu and Johnson 2024 article on power looms and their effect on weaver employment and earnings. The first one is here:

In previous work (Acemoglu & Johnson 2023), we propose a framework that blends Ricardo’s and Thompson’s ideas to clarify when new technologies improve the lot of workers. For the benefits of growth to be shared, the right combination of technological and political conditions must exist.

New technologies can reduce the value of marginal productivity for workers even as they raise average productivity. Most saliently, automation—the substitution of machinery for tasks previously performed by workers—displaces workers and can reduce, rather than increase, the demand for labor. This is what happened to handloom weavers.

They then link to a footnote that is somewhat more dense which was actually what put me on this track to develop this out more for my students. It’s footnote 6 and it reads:

The short version of how this can happen is as follows. A technological improvement increases output for given quantities of factors of production and thus raises average labor productivity. Demand for labor, as well as employment and wages, is determined by labor’s marginal productivity (or, more precisely, by the value of the marginal product of labor). The general presumption is that average and marginal productivity of labor should move together, but there is no theoretical guarantee for this. They do so when the aggregate production function of the economy can be approximated by a Cobb-Douglas function, which imposes that these two quantities be proportional to each other. The same conclusion also applies when the aggregate production function exhibits constant returns to scale and the cost of capital remains constant even as the demand for capital increases. In general, however, there is no such guarantee, and automation—defined as machines taking over tasks previously performed by labor—expands the wedge between average and marginal productivity of labor (see Acemoglu & Restrepo 2018, 2019 for discussion).

So that’s the goal, and there’s a few moving parts to it that precede the headline conclusion that I have to figure out how to explain to students that haven’t necessarily taken intermediate micro. Those are:

1. Why should we care about productivity?

2. What’s average and marginal productivity and why should I care about those?

3. What’s Cobb-Douglas and can I show them the claim that average and marginal productivity of labor respond in the same direction to technology shocks?

4. What’s the intuition that AI, and automation more generally, can cause the opposite whereby marginal productivity falls while average productivity rises?

5. And if that happened, what is the ramifications it has on output, employment, wages and income?

But first the paywall coin flip. Best two out of three. Heads I paywall, tails I don’t.

Geez! Swear I don’t have my thumb on the scale. Cosmos wrote the code and ran it in python. Well, I guess if you subscribe, though, you will get to read more about this point in Acemoglu’s work and my effort at trying to make the lecture.

import numpy as np
import matplotlib.pyplot as plt

# Flip a fair coin three times (1 = Heads, 0 = Tails)
flips = np.random.binomial(1, 0.5, 3)
heads_count = np.sum(flips)
tails_count = 3 - heads_count

# Determine the winner
winner = "Heads" if heads_count > tails_count else "Tails"

# Plot the results
fig, ax = plt.subplots(figsize=(5,5))
bars = ax.bar(["Heads", "Tails"], [heads_count, tails_count], color=["orange", "blue"])

# Label the bars
for bar in bars:
    ax.text(bar.get_x() + bar.get_width()/2, bar.get_height() - 0.1, 
            str(int(bar.get_height())), ha='center', va='bottom', fontsize=12, color='white', fontweight='bold')

# Titles and labels
ax.set_title("Best Two Out of Three", fontsize=14, fontweight='bold')
ax.set_xlabel("Outcome", fontsize=12)
ax.set_ylabel("Count", fontsize=12)
plt.suptitle(f"Winner: {winner}", fontsize=12, fontweight='bold', color='red')

# Show the plot
plt.show()

---

Why should we care about productivity?

Keep in mind my definition of economics used in this class. Economics is the science that seeks to ask and answer three questions:

1. What will we make (production)

2. How will we make it (technology, factors of production)

3. How will we share it (distribution, structure of wages, income, consumption)

Artificial intelligence, in the writings of Acemoglu, Autor and other economists, impacts all of these. The magnitudes are in debate, as we are early in the process of integrating AI into production. We don’t know yet which “tasks” will and will not be automated, and we don’t know what will happen in the sectors that are automated with regards to demand for labor relative to the others that presumably will have growing demand for other tasks. In short, we know nothing other than historical case studies and theoretical models, both of which Acemoglu and Johnson are about.

We care about productivity because it is the key driver of economic growth, full stop. By making more out of less, we are able to spend that excess time on other productive tasks which in turn produce more. The more we produce, the more prosperity we have. Productivity is the story of the wealth of nations. And labor is the central driver of this productivity story — historically directly, then increasingly over time indirectly. With the rise of the importance of capital and technology, the role that labor has in driving productivity is less clear and can be harder to calculate, but even when the driver is capital and technology, since labor created them, historically economists still saw productivity as driven by labor. It is no wonder then that Adam Smith and others ascribed all value to labor, as it seemed it was logically impossible it could be anything else.