These tweets caught my eye:
I suspect that it would be possible to create some sort of argument that the AI boom is hurting the job market, but at the risk of being unserious I don’t find this one to be particularly persuasive. Suppose I made the following argument:
Interest rates would be lower if we went into recession, which would help employment. “Economic booms don’t hurt the job market” is an unserious view.
Long time readers have probably guessed that I’d view this hypothetical claim as an example of the fallacy of “reasoning from a price change.”
In some respects, it is surprising that the labor market is so strong. We had a period of high inflation during 2021-23, and unemployment often rises sharply when the Fed uses a restrictive monetary policy to bring inflation back down. Why has unemployment merely edged up from 3.4% to 4.2%? I’m not certain, but perhaps because the disinflation policy was gradual, and even today inflation remains above the Fed’s 2% target. Nonetheless, if the labor market is currently a bit subpar, it is probably due to the lingering effects of the Fed’s anti-inflation policy, not the AI boom.
I certainly agree with claims that unemployment might rise if the Fed pushed interest rates above their natural rate. But an AI boom tends to raise the natural rate of interest. Other things equal (including the Fed’s target interest rate), a higher natural rate of interest is actually expansionary—likely to lead to faster NGDP growth. Of course other factors such as the lower rate of immigration tend to reduce the natural rate of interest, so I’m agnostic on the question of whether monetary policy is currently too tight. (As an aside, TIPS markets are currently pricing in about 2.5% inflation over the next 5 years, which doesn’t suggest that money is particularly tight.)
Perhaps there’s an argument that AI spending crowds out more labor intensive industries, although in principle the Fed should offset that effect. Of course monetary policy is not perfect, but the internet boom of 1999-2000 doesn’t seem to have hurt the labor market. In 2000, unemployment fell to the lowest level since the 1960s. We eventually did have a mild recession, after the Fed engineered much slower NGDP growth in 2001.
READER COMMENTS
David Seltzer
Aug 19 2025 at 5:14pm
Scott, I thought about this. I think the fallacy here, as I infer it, a more efficient way of producing, as I assume AI is, destroys jobs. A following proposition; a less efficient way of producing creates more jobs. Why focus on employment or the labor market instead of production? High employment can come with people digging holes with spoons and then filling them up. I don’t see real productivity here even with low unemployment. If economies experience substantial increases with a strong labor market, would the the price of loans, interest rates, increase with an increase in demand for money? I hope my thinking isn’t too confused.
Robert Ferrell
Aug 19 2025 at 6:23pm
On a business trip to Beijing in the early 1990s I saw something like this. At one location there was a man stationed in the bathroom who handed me a very nice cloth towel to dry my hands. At another location there was a stack of paper towels. The man stationed in the bathroom handed me a paper towel from the stack. At a third location the paper towels came from a wall-mounted dispenser. There was a man in the bathroom who dispensed a length of paper towel and handed it to me. And at a fourth location they had a heated blow dryer for hands. The man stationed in the bathroom rushed ahead of me and pushed the button on the dryer for me. Full employment, but no increase in productivity from the towel dispensing technology.
I expect we will go through a transition period during which somebody will take requests from a “client”, input it into an AI chat bot and then supply the response to the client. No jobs lost, but no gain in productivity from all the CapEx either.
Don Geddis
Aug 19 2025 at 7:55pm
That’s straying into the Lump of Labor Fallacy. Efficiency can reduce the number of jobs in a given industry (such as agriculture in the US dropping from about 97% of the population to about 3% of the population over the last few centuries). It can affect the total amount of value produced by the economy (better efficiency allows for more total value creation). But it doesn’t affect the overall employment or unemployment rate. People are a resource, and jobs are what turn people into products (or services) with value.
For sure, individuals might lose a job they had held for awhile (over the short term). But the economy as a whole does not decline in the total number of jobs (over the long term). The Luddites may have been right about themselves, but were wrong about the economy.
David Seltzer
Aug 20 2025 at 9:04am
Don, I should have been more clear. Fallacy of of increased efficiency destroys jobs assumes there is only a fixed amount of work available in an economy. Thanks for pointing that out.
Knut P. Heen
Aug 25 2025 at 11:22am
The simplest way to analyze these questions without getting confused is to think about the opportunity cost. If someone does X now, they cannot do Y. If AI takes over task X, someone will have time to do Y instead. Thus both X and Y get done and we are better off.
Thomas L Hutcheson
Aug 19 2025 at 6:09pm
If AI investors were less bullish with the Fed inflation target unchanged, it would presumably reduce interest rates and there would be more investment by investors in other sectors. Would this be good? It depend on the actual returns on investment in AI vs other sectors.
In the meantime it seems like the tariffs and fall in net immigration are big and sectorally concentrated enough shocks to call for some additional inflation to facilitate relative price adjustment, even if the fall in the dollar has mitigated some of the relative price shock for exports and substitutes of non-tariffed imports.
Garrett
Aug 19 2025 at 6:49pm
It’s a bit amusing to imagine Scott scrolling Twitter on his phone, seeing this tweet, half grinning, and saying to himself, “ah yes, more NRFAPC content.”
Scott Sumner
Aug 20 2025 at 10:56am
I only use my phone for phone calls and photos. I scan the internet on my desktop.
Garrett
Aug 19 2025 at 7:14pm
Here’s a more serious comment: I was playing around with GPT5 and had it formalize NRFAPC:
Minimal structural setup
Let inverse demand and supply be linear:
Demand: p = A – B*q + eps_d, with B > 0, eps_d a demand shifter (income, tastes).
Supply: p = C + D*q + eps_s, with D > 0, eps_s a supply shifter (costs, technology).
Equilibrium (Q*, P*) solves A – BQ + eps_d = C + DQ + eps_s.
Solve:
Q* = (A – C + eps_d – eps_s) / (B + D)
P* = (AD + CB)/(B + D) + [D/(B + D)]*eps_d + [B/(B + D)]*eps_s
The identification fact
Comparative statics (holding the other shock fixed):
dP*/d(eps_d) = D/(B + D) > 0
dP*/d(eps_s) = B/(B + D) > 0
Either a positive demand shock or a positive supply (cost) shock raises the equilibrium price. Observing Delta P > 0 alone does not identify whether demand rose, supply tightened, or both.
Quantities move in opposite directions:
dQ*/d(eps_d) = 1/(B + D) > 0
dQ*/d(eps_s) = -1/(B + D) < 0
Diagnostic sign table (standard slopes):
P up and Q up -> demand up
P up and Q down -> supply down (cost up)
P down and Q up -> supply up (cost down)
P down and Q down -> demand down
What the fallacy is, precisely
“Reasoning from a price change” = inferring the sign (and implications) of a demand or supply shift from Delta P alone, then jumping to the downstream story that would follow if that inferred shift were known to be true. Formally, the mapping from (eps_d, eps_s) to Delta P is many-to-one; any rule that maps sign(Delta P) directly to sign(eps_d) or sign(eps_s) is misspecified.
A Bayesian gloss: with symmetric priors over shocks, observing Delta P > 0 alone does not change the posterior odds for “demand vs supply.” Price isn’t a sufficient statistic for that decomposition.
Why it matters (the “jump to implications”)
Demand and supply shifts have different second-order implications (welfare, incidence, policy, forecasting). Misreading a price rise as “demand is hot” leads you to predict higher quantities and prescribe cooling policy, when a negative supply shock implies lower quantities and a different policy mix. Sumner’s macro version: reading low nominal interest rates as “easy money” confuses tight money (weak NGDP, low natural rates) with an accommodative stance—same identification problem.
How to avoid it (minimal fixes)
Use quantities: pair Delta P with Delta Q (see sign table).
Use exogenous shifters/instruments: income/tastes for demand; costs/technology/weather for supply.
Exploit timing and cross-sections: who moved first; which markets share demand but differ in costs.
Check forward prices/term structure: supply (cost) shocks vs demand shifts often show different forward patterns.
Quick summary
A price change is the net effect of multiple shocks filtered through slopes. Treating its sign as a diagnosis of “demand” or “supply” is an identification error—hence: never reason from a price change.
Alan Goldhammer
Aug 20 2025 at 10:47am
The “AI Boom” is concentrated among very few companies and a number of them are still privately held. We have chip suppliers, data center construction and maintenance, and programmers/data scientists. As I am an AI skeptic (disclosure: I do own MSFT in my portfolio), I have not followed the flow of funds into private companies who are developing AI tools. I assume that this is a combination of venture capital and private credit loans. Certainly, major established companies can fund things out of existing cash flows from product sales but this is not the case for Open AI, Anthropic and several others. My question is whether this is really crowding out borrowing for companies that make up the rest of our economy. If it does not, why should this have any impact on interest rates?
A bunch of AI stocks got hit hard yesterday and that continues this morning (10:30 AM EDT). It will take more than a couple of days of downturn to see if this is really a trend but it is interesting.
Matthias
Aug 23 2025 at 10:59pm
Going on a tangent: I’m a skeptic of a lot of things, but I still hold all the companies in my portfolio. That’s what index investing is like.
About your comment: financing out if retained earnings crowds out other investments just as much or as little: look at the opportunity costs; the retained earnings could have been returned to investors to reinvest or consume or invested by the company directly in something else.
Scott Sumner
Aug 20 2025 at 10:59am
Everyone, Some people do engage in the “lump of labor” fallacy when analyzing AI. To be clear, I don’t see Conor Sen doing that.
Comments are closed.