This is specifically AI in the terminal: programs where agents run on your computer, in a folder you pick to host a project. These are programs like Claude Code, OpenAI’s Codex, or Google’s Antigravity. If, as Andrej Karpathy argues, the first age was copying-and-pasting from the web browser, then this second age is working iteratively on projects this way.

(I’m getting used to the likely third age, where AI just exists alongside your workflow. Through Claude Cowork, I get a daily update of my emails, Slack channels, calendar, and projects. When it comes to next steps on projects I’m managing, I have it scan documents and the recorded transcripts of all my meetings to summarize what’s happening. I’m excited to continue to expand that.)

But for now let’s stick with projects. In that previous post, I talked about how I use this quite a bit for data analysis, for investigative dives into topics I want to explore. I’ve found, especially with the latest versions like Opus 4.8, that I can open a folder, talk through a problem for it to investigate, and after about 20 minutes it comes back with a pretty in-depth look at it. Like the people at Anthropic, I rely on the “unreasonable effectiveness” of HTML as an output to display the information the agents produce.

One example: I’ve been studying the Acemoglu & Restrepo 2018 and 2019 core papers on task-based automation, apparently we all need theories of AI economics these days, and these are the standard, though complicated, references in the literature. To understand it better as I read it, I had Claude Code create a supplementary reading guide you can read here. It walked through the math derivation of the labor share formula at a dramatically slower pace (sorry Daron) with extra steps and clarifications so I have a small hope of following it. I understand we’re excited to speed up math discovery with AI, but I’m here to use it to slow it down, by, for example, explaining how exactly to understand the shadow price in the Lagrange optimization of a task-based aggregate output function. I then had Claude create additional tabs accessible at the top explaining how the equation values change, how the paper hit the literature as it existed at the time, how these debates existed historically, and then a literature review retrospective on how the papers have been incorporated since then.

You can get ambitious with these dives. I wanted to learn more about the American Time Use Survey (ATUS), both in and of itself and because Derek Thompson and others who post stats from it go viral on Substack. So I used my time with Fable 5 to have it study when the ATUS goes viral, and use that to come up with 63 potential hot takes, each of which could be a Substack post. It then downloaded the data and ran the numbers on each of them, determining which ones passed correctly, and sorted the accurate ones by a potential virality score. You can see the full HTML with tabs at the top for multiple summary levels here:

There are probably more productive uses of the frontier models.1 I’ll say that if you are using terminal AI agents, you should have personalized your own CLAUDE.md file. Just encouraging it not to write like AI gets you pretty far. I have also loaded an anti-slop writing skill (GitHub here) that does a pretty good job of forcing it out of the AI voice. I highly recommend we all start using something like this, because some of the stock formulations are really starting to kill me.

That’s what I had been doing. So what’s new?

Writing

Setup and Fact-Checking

I think I have figured out my workflow with writing and AI. I’ll use the recent housing report, Building Affordability: The Policy Agenda for America’s Housing Crisis, which I co-wrote with Becky Chao and Ned Resnikoff, and the subsequent blog posts as a jumping-off point. (I’ll clarify upfront that none of the report text itself was written with AI.)

If you are someone who writes and is uncertain about how to incorporate AI, let me make two suggestions. The first is to get a baseline familiarity with GitHub and the practices of git. This is essentially a sophisticated version of track-changes for computer code, though you can use it for text as well. Getting that set up goes beyond this post (but you can use AI to help!), but when it’s set up you’ll have normal folders on your computer that you can quickly and easily back up, that track every single change. This is the standard computer programmers use, where they need to be able to track every comma change across hundreds of files; regular writers can use it too.

The second is to try working in text with a .md file, called Markdown. It’s just a plain text file you can read with some markings for formatting. It’s not proprietary: you can read it in anything that reads text (I like MarkEdit and my own Draftwatch). Like the best of the internet, the late Aaron Swartz played a role in its creation. AIs work really well with it, getting better accuracy and token usage than PDFs. For any of your Google Docs, you can go to File -> Download, and one of the formats you can save to, besides a Word doc or PDF, is a .md file.

So to start, once we had a draft, I took the entire text of the housing report (about 40 pages of text) from Google Docs and downloaded it as a .md, put it in a folder, and asked Claude Code to give a list of obvious typos and grammar errors. Here’s what it found 57 seconds later:

Some of them are a bit funny, and it was a good start, and after a few rounds that was clear. I then asked it to run a comprehensive fact-checking operation, after first coming up with a plan to tackle it in steps. It listed out how it would check the bibliography, and it also came up with a list of hundreds of facts that it proceeded to check and verify (we have .md links in this paragraph). This fact-checking works significantly better than if you were trying to do it in the browser, which I have done in the past, because it can write notes to itself, build an infrastructure in the folder, and go back-and-forth. It found a few potential errors or things it couldn’t confirm for us to check, citations it could confirm and those it couldn’t reach from the terminal web access (which we then just fed into the browser version of Claude which could). It was great to get this level of coverage, and to double-check them ourselves and fix them.

Editing

I wrote a few follow-up blog posts on the paper after it came out. My general process for writing is to start by dictating using Wispr Flow. I have a good sense of what I want to communicate in my head, but hate writing the first words, and speaking out the rough draft makes it much easier on me. I have AI clean up the transcription and then do my own pass of a full rewrite multiple times.

We want to be able to let AI rip but don’t want to destroy everything we spent time building. That’s true for civilization as a whole and also true for personal writing drafts. And that’s an awkward thing about editing with AI: if you ask it to make straightforward and targeted changes to something you’ve written, you can’t really be sure it isn’t also changing a ton of other stuff in the document too. Spot-checking is unreliable, and just manually making the suggested adjustments doesn’t save you much time.

Now here is where git pays off. I “commit” the draft so the current version is saved, then ask AI to edit the markdown. Here, below, for the draft of the blog post announcing the paper, I asked Claude Code “to fix typos and obvious errors. Also search the internet to add links to the cited papers and hunt down the missing factual TKs I don’t have here.” I then used git to look at the differences to see that that’s only what it did.2

I have used Claude Code as an editor but it’s generally been a pain, because you can never be sure what else it changed; this process has fixed that for me. The AI has nothing to do with git showing the differences. As you can see it did only what I asked it to do, and I have confirmation there’s no odd hallucinations I’m missing. You can then commit that version again and iterate.

For my post on the ROAD to Housing Act, I did the above steps, and I also then asked Claude Code to suggest line edits that would make it read better. Here you can see what it suggested in the terminal, and from the git diff we can confirm that it made only those changes, which I can go back and edit myself.

I’ve been professionally edited in plenty of settings, and this is what it feels like to have an editor go over your work at the line level and clean up awkward sentence structures. Coming from a computer science background originally, I view this as just the evolution of spellchecking and search, the way all computer processes have evolved in their sophistication and abstraction. But I can understand a variety of reactions to the technology and writing.

I decided to use my Fable 5 tokens to formalize this process into an IDE for writers that I’ve vibe-coded, which I’ll announce here in the future once it has been tested more. But for now I’ll mention it in a footnote if anyone wants to test it with me, I could use additional eyes.3

Side note: Has the writing peaked?

You may have caught that there was a lot of drama about Fable 5, the latest Claude model, which was released under tight export controls, briefly pulled to comply with a government directive, and then restored last week after the administration reversed course. During that initial period it was available, I put the .md file of my housing report in a folder, along with several supporting documents, press releases, and other prep materials, and asked it to write ten-minute remarks for a group of activists and campaigners, briefing them on the state of the ROAD to Housing Act from online sources, plus a summary of my report and what it covers. I gave it a minute-by-minute breakdown of how the remarks should flow, then told it to write. I then did the same exercise with Opus 4.8, the previous frontier Claude model, in a different folder.

You can see both versions here, where I also asked Fable to summarize the differences at the top. Both of these are unusable. (I didn’t use them.) And while Fable 5 strikes me as much more sophisticated at coding from some of the other work I’m doing, including that IDE project, for writing I did not think it was a step up. It’s possible Moore’s Law won’t apply to creative writing, and we’ll just be stuck at the slop level. I don’t know if that’s good or bad.

Iterating Ideas Themselves

The other substantial difference in my workflow is using it to push back on my ideas, and to give me context for how to debate things. I understand the serious worry AI will flatten our thinking, but there is a world where AI could lead to deeper understanding for those who seek it. I have definitely used it to summarize things I could have read for more depth. But I’ve also used it to force better understanding of arguments. Anything I write goes through several “what is wrong with this? / what would opponents criticize?” rounds and I get clear answers.

So I won’t share all these materials, but during the drafting of the housing report I would have AI consistently push back on the arguments, summarizing a wide range of different opinions. Here, for example, is how Claude Opus quickly summarized one thing nine different relevant actors would find appealing about the housing report, and one thing they would push back on or be skeptical about. This was helpful for me building the argument, and helpful for my organization to get a sense of any incoming criticism that might be on its way.

I can share a more in-depth example. I gave congressional testimony earlier this year on the debt and deficit. You may not know this, but minority witnesses at these hearings basically get a day or two to put it all together. While I was prepping, we got the names of the three majority witnesses and their testimony, so I had AI build a full HTML document.

It read everything the three of them had ever written, then laid out the best arguments it thought they’d make against my testimony point-by-point, along with responses I could give. A second tab also had it go through all their testimony and pull out what they had in common. Then it went through each of their claims in detail and flagged things for me to be aware of and answers to have on hand. Here, for instance, is how the AI understood how the Republican witnesses would respond to my second substantive claim:

I am aware that some of my favorite uses of this technology mirror the plot of the Nathan Fielder show The Rehearsal. So be it. These examples, and really the whole thing, don’t substitute for experience and judgment, for knowing which points are relevant and what to emphasize when. But the raw factual material and how it’s presented can be automated pretty quickly. The base case is a world flooded with slop. But there is a path for better and more interesting things to happen.

1

When Fable was first announced and then banned, all the elder dads at the MoCo playground started talking about it. There was a lot of “what did you do with your 72 hours of Fable?”

One dad said “I used Fable to one-shot a full GitOps CI/CD pipeline, reducing deployment times from 45 minutes to 90 seconds.”

I responded that I used it to reverse engineer a time diary to try to go viral on Substack. The response from the dads was similar to this meme:

3

Is it too late to vibe code apps? That feels like a January 2026 idea. But either way, I finally understood the app I wanted to track changes on AI edits, and it wasn’t until Fable 5 that the vibe-coded architecture made sense. I’m going to do something more formal at a later date, but if you are interested in beta-testing a vibe-coded app, check out Draftwatch!

The diff comes from your local git, not from the AI vendor. You get independent verification of what the agent or script actually did.

Why this vibe-coded app? I designed it to be an Integrated Development Environment (IDE) for writers. Most IDEs show git diff at the line level, which is appropriate for coders (where each statement is generally its own line) but terrible for writers, who work at the paragraph level. This IDE is built around displaying git diff --word-diff=porcelain, which lets writers see the specific words being edited. Even the IDEs that do show this make it harder for writers to track edits, and they carry coding features and visual baggage that writers won’t need. Please leave comments if you use it!

“I said I’m not signing the housing bill […] it’s all about the interest rates. Lower the interest rates and you are going to have all the housing you want.” - President Trump, via Acyn

If the third time is the charm, the fourth time is pathological. And for the fourth time during the past year and a half, President Trump chose his own goals over affordability and everything else voters elected him to do.

Earlier this week, Congress passed the Road to Housing Act, the first serious effort in decades to get more housing built. While Republicans were promoting it, Trump announced he wouldn’t sign it unless the Senate passed a bill on voting.

Before we dig into that, there are many examples of Trump’s indifference to what voters expected, but three stand out:

The first was the destabilizing announcement of tariffs. Rather than a tariff strategy designed to lock in competitive advantage or check China, we got tariffs on things we don’t make, like bananas, tariffs as an aggressive wedge against long-time allies, and us selling high-end NVIDIA chips to China anyway. It’s a confusing mess, and people are paying for it.

The second was the budget. Severe cuts to SNAP and Medicaid and the expiration of the premium tax credits, all to fund the continuation and expansion of high-end tax cuts, while also blowing out the deficit. All together, according to the Yale Budget Lab, the bottom 80% of people are worse off between the tariffs and the budget. Again, reducing their incomes is not what people elected Trump to do.

Costly military adventurism in Iran is the third. Understanding the specific strategy here isn’t really my world, but just as a concerned citizen, it’s pretty obvious that there’s no plan. There was never a plan. Trump ran as the anti-war candidate and I’m sure people are remembering this betrayed promise as they cut back their summer travel plans over energy prices.

What’s Good in the ROAD Act

Which makes the housing action even more pathological. The one thing he could have done was sign anything on housing, and the Road to Housing Act is much more than symbolic. I think there’s a lot here that you probably don’t know. This bill has mostly been debated in fairly technocratic circles, and it bundles together provisions drawn from more than 60 separate bills, 36 of them with bipartisan sponsors. There’s no big headline item driving the conversation. Instead, it’s a genuine bipartisan bill, a good mix of what liberals like and what conservatives like, negotiated carefully over many rounds. That’s why it passed 358 to 32 in the House and 85 to 5 in the Senate.

A few things stand out on the affordability front. Much of the challenge of federal intervention in the housing market is that it is fundamentally local. But mobile home regulations are not. So the bill makes important, much-needed changes to mobile homes, where the federal government sets the standards. As Rachel Cohen at Vox has reported, the current rules require every manufactured home to be built on a permanent steel chassis. The bill removes that requirement, which alone takes $5,000 to $10,000 off the cost of a home. That also opens up floor plans, makes two-story manufactured homes possible, and helps these homes fit into more neighborhoods.

There are other baby steps toward industrial policy for housing, including clearing FHA financing barriers for modular developers, raising FHA loan limits for manufactured homes and ADUs, and commissioning studies of off-site construction’s cost-effectiveness, all of which can help lower the capital and regulatory friction around factory production. As more people are discussing (D’Amico et al 2024, Gupta & Teles 2026), the lack of factory-scale housing production is certainly a major reason we’ve seen declining productivity in construction, and these simple changes can start to push us in that direction. We discuss this element of boosting productivity in housing in the Economic Security Project’s Building Affordability report, which I recently co-authored with Ned Resnikoff and Becky Chao.

Building codes are another other target. The bill sets federal standards and pilot funding for some important reforms, notably single-stair buildings at six floors and under. As we note in our report, the stairs rule is especially hard on young families. Requiring a second staircase forces smaller apartments with fewer bedrooms, which is a big reason new construction skews toward studios and one-bedrooms while the family-sized two- and three-bedroom units are so scarce. And there’s little extra safety to show for it (Pew 2025). This bill doesn’t have the funding to compel areas to adopt these reforms. However, a coordination mechanism for local politicians, activists, and other policy entrepreneurs can be helpful, and a few well-studied pilots can open people’s eyes on these subtle costs. The bill lets the federal government and HUD name other focal points to coordinate around.

ROAD does more that we discuss in the report. It also loosens the timing of inspection rules that lead some landlords to turn away Housing Choice Vouchers (HCV). The obvious downside to this bipartisan bill is the lack of funding to push reforms more dramatically, and HCV would benefit from more funded pilots. But this does have upsides to build from.

The bill also bars large institutional investors, the funds that already own at least 350 single-family homes, from buying up any more. My honest read is that the evidence on the national effects is mixed, both on landlord characteristics (Gurun et al 2023) and renters-vs-buyers (Coven 2025). But there are places, notably the Sun Belt metros across the Southeast like Atlanta and Charlotte (especially in their entry-level tier where first-time buyers compete), where the concentration is high enough to matter. And they managed to write it without the negative impact of cutting off build-to-rent construction or potentially forcing evictions after a set number of years. This is overall a balanced solution to a concern that is shared by both parties and is genuinely widespread among the public.

A personal fave: everyone is trying to figure out where the “Occupy to DSA to sewer socialist YIMBY elected representative” pipeline is going to end up. Well, a person who pioneered this trend, Pennsylvania State Senator Nikil Saval (with a detour through small left-wing literary magazines), started a state program in 2022 called Whole-Home Repairs. It funded an office in every county where people could go to fix up and weatherize an aging home, cut their energy bills, or adapt it so an older relative could keep living there. Nearly 4,000 Pennsylvania households have used it, and Maine, Maryland, and Washington built their own versions. The Road to Housing Act takes that idea national as a pilot in Section 202. Keeping existing homes from falling into disrepair is one of the key channels for housing affordability, and it shows how the new YIMBY sewer socialist left is demonstrating projects that can scale.

Even if it becomes law, the fact that Trump’s first instinct was to use the central economic anxiety people face as leverage for other priorities is just another example of how badly his team has misunderstood the job.

Apparently he is opposed to all those as well. Trimming down a bloated FOMC policy statement is good.1 But it’s not just that he didn’t submit his estimate of appropriate future policy rates over the next few years (the “dot plot”) last week, he apparently didn’t submit any values for the Summary of Economic Projections, which has existed since 2007. During the press conference he conveyed very little information on what the Fed is thinking at all. When asked explicitly about his “reaction function” he deflected. See Nick Timiraos’s WSJ coverage on how confused it left everyone.

Maybe that’s fine if this was a boring meeting, but with inflation increasing, financial markets really wanted to know if hikes were more likely to be on the table. I’m trying to imagine the counterfactual where there was no dot plot and all the other information was also missing, with traders suddenly having to price in a ton of confusion over multiple possible Fed reactions. Rates would be basically indeterminate (remember that word, we’ll come back to it in a minute) in this scenario.

Modern macroeconomics is about expectations. What’s the theory of expectations here? When asked about rolling back communication, Warsh gave a very specific answer (transcript):

EDWARD LAWRENCE. Thanks. Welcome, Mr. Chairman. Edward Lawrence with Fox Business. So, if you don’t give a lot of ongoing forward guidance, won’t the market have more volatility, and shouldn’t Americans have more access into what you’re thinking going forward?

CHAIRMAN WARSH. So I think financial markets perform best when they react to incoming data. I think the financial markets work less efficiently when they ask a question. How will the Federal Reserve react to that incoming information? The more that markets are paying attention to what’s happening in the real economy, deciding what’s good data and what’s less good data, the more financial markets can price what they believe is the most likely and what are the tail risks.

Financial market prices are probably the most important source of information to guide central bankers. But when all the financial markets are doing is reflecting back what we’ve said, then we’re taking the most important source of information and we’re being blind to it. I’d like us to create a system where those blinders come off, where markets are following data that they efficiently think is reliable. And they’ll be watching data, we’ll be watching data. They’ll come with better information through market prices to us, we can make more informed decisions.

Read that again. This is not what I was expecting, which was an argument that forward guidance locks the central bank into paths it fears it cannot change (which is what he said during his Senate testimony). That argument relates to the recent inflation wave and I think it is important to debate it. (A fan chart replacing the dots might be nice.) It also wasn’t that forward guidance should only be done at the zero lower bound, or that its effects are overstated by the models (McKay, Nakamura, Steinsson 2016), or any of the other many things economists have debated over the past decade.

This answer is pure “Ketchup Economics.” It’s a theory that derives strong answers from prices without any idea of what is informing those prices in the first place. The problem compounds when this argument ignores the complex interplay between market prices and central bank action.

The idea that the Federal Reserve can simply target private-sector inflation information that exists independently of the Fed’s own forecasts, and that more public central bank information can have negative effects, is not new to Warsh here. These two propositions were debated by many from the mid-1990s to the mid-2000s.

The Indeterminacy of Targeting Forecasts

Take Warsh at his word that market prices are “probably the most important source of information to guide central bankers,” then ask what happens if the Fed actually sets policy off that forecast. In the 1990s many proposed precisely this: that the Fed set interest rates based solely on private-market inflation forecasts. Ben Bernanke and Michael Woodford took apart that argument in their 1997 paper Inflation Forecasts and Monetary Policy, which built on an earlier 1994 Woodford model.

You can do it all with their simple toy model:

This is straightforward. The inflation target is zero here (but it could be anything). Inflation in the next period equals inflationary pressures (s) with the Fed setting rates to offset them (-R) plus shocks. In this case, the Fed observes the economic pressures and sets rates to offset them.

Now what if the Fed follows private sector forecasts of inflation instead?

The Fed is looking at the private market’s forecast of inflation (f) and reacting with an intensity I. But once we come up with an equation for that forecast, the last line above, it includes that government policy choice I in its definition! This is the circularity everyone is feeling from this new direction.

Let’s go to the next step.

No value of I makes that last line true. There’s no solution. Or, in English, the forecast is worth watching only because it still moves when inflation pressure moves, so a Fed that reacts hard enough to cancel that pressure also flattens the forecast, and the better it gets at hitting its target, the less the signal it is chasing has left to say. The confusion everyone is feeling falls right out of this economic logic.

Bernanke and Woodford are able to draw the same result with more complicated methods (it’s a fun paper to read) we’ll have some fun in a footnote too, showing this confusion can also derive from microfounded economic theory too.23

Here is how Bernanke and Woodford summarize their overall findings:

An interesting possibility is that the central bank might target current private-sector forecasts of inflation [...] we show that strict targeting of inflation forecasts is typically inconsistent with the existence of rational expectations equilibrium, and that policies approximating strict inflation-forecast targeting are likely to have undesirable properties. We also show that economies with more general forecast-based policy rules are particularly susceptible to indeterminacy [Konczal note: there’s that word from above!] of rational expectations equilibria.

This dynamic is sometimes described as a “hall of mirrors.” More generally, the thing Warsh doesn’t like, the market reflecting the Fed back rather than bringing new information, becomes even more of a problem when the central bank targets the forecast. Around the same time, the field’s attention was turning to the determinacy of equilibrium under interest-rate rules more broadly (Clarida, Galí, and Gertler 2000; Woodford 2003), and the narrower question of whether to target market forecasts quietly dropped out of view.

When Can There Be Too Much Information?

So the central bank should communicate. But just a little, or a lot? The argument that central bank communication has an ambiguous effect on social welfare made its case in Stephen Morris and Hyun Song Shin’s 2002 paper Social Value of Public Information.

In a world with Keynesian Beauty Contest dynamics, traders try to minimize a loss, with the first term punishing them for being wrong about the fundamental θ, and the second punishing them for straying from what everyone else is doing (the Keynesian beauty contest). More precise public information sharpens everyone’s read on θ, which is good, but if the public signal is wrong it moves everyone the same wrong way at once, which is bad and can dominate the first.

A wave of work pushed back. As Lars E. O. Svensson showed, within this same model more public information is better in virtually every realistic case: the perverse result requires private information to be at least eight times as precise as the central bank’s public signal before it can even arise.

Michael Woodford then took the case apart in his 2005 overview, Central Bank Communication and Policy Effectiveness, calling the conditions Morris and Shin need “quite special.” But for our purposes here, Woodford calls out that the Fed credibly knows more than anyone else about the future rate path because it sets the rate path. Since policy works almost entirely through expectations, with “very little else” mattering, a Fed that goes quiet does not free markets to discover the fundamental. It just leaves them to guess at the one fundamental that only the Fed controls: the future path of the funds rate.

Maybe I am reading too much into a random answer? Given how buttoned-down his answers were on everything else, the fact that he had such a clear, detailed, and energetic response to an obvious question he was going to get makes me think this is substantive for him.

I am open to new arguments on communications here, as we all should be. But I do think central bank transparency on its actions is actually very good, for both economic and democratic reasons. If a central bank has political independence it needs to be able to articulate what it’s doing. And I have had a far better mental map of the Fed’s decision making before this presser than whatever awfulness is being cooked up in the Trump administration. If this information becomes opaque for those who can’t afford to attend Bank of America private dinners, we are worse off for it.

Let’s see who is on the Communications Task Force. If Warsh really wants to take apart the use of communications that’s evolved since the early 2000s, that’s going to be a hard slog. But on the other hand, a lot of economists take this so much for granted that I’m wondering if they’ll be able to muster the correct defense.

Inflation-targeting, expectations, and the surrounding communications infrastructure is so baked in at this point that people have probably forgotten the kind of debates mentioned above even happened (do they even teach them in grad school anymore?). Those that were active have often stepped back from public life and may not be able to communicate in the new digital and social era. I mean, where are we going to find an economist who was in the thick of debates about communicating expectations at the zero-lower-bound early on, even at Princeton during this time period, and also big on Substack today?

3

It builds on the Affordability Framework Becky and I wrote last fall, where we diagnose affordability and our more general economic challenges as deriving from a combination of broken markets and broken incomes. It struck me as a useful way to break out of the silos of how policy information is created, which tends to focus very narrowly. I also thought the framework was broad enough to cover a lot of ground, but had the depth to make substantive and useful claims. But journalists covering it, such as Rachel Cohen Booth at Vox, flagged that it really needed to be deployed somewhere to see if any of this was true, which I agreed with.

Enter housing. As you can see in the graphic above, we tackle it through missing supply, through broken markets, but also equally emphasize that even when housing markets work, incomes will be a challenge.

There are a lot of housing reports out there, so where does this add value? We think this report pulls from the major camps of thought around housing. But I think a useful way to approach it is to contrast it against the Abundance movement right now.

This paper takes the YIMBY arguments very seriously and I think it continues to push them, for instance by bringing building codes more centrally into the conversation. But there are two challenges it poses.

The first is that because supply matters so much, we should go further than zoning. We focus on public options in housing, which become easier to stand up if we do the YIMBY playbook and address construction costs and building codes. But we should also reexamine financing as well as industrial policy for innovation too. Reflecting on their book Abundance one year out, the writers Ezra Klein and Derek Thompson flagged financing as the big question mark once you get past zoning. That’s why I think there have been a number of proposals out there, such as Chris Hughes and Aaron Klein at Brookings, from the Center for Public Enterprise, and from Groundwork Collaborative, taking this seriously. We give an overview.

The other challenge is taking the income side seriously as well. As my co-author Ned Resnikoff puts it, increasing supply can do quite a lot: research suggests that in a healthy housing market, one where there’s enough market-rate housing to meet demand, people making as low as 60 percent of an area’s median income can likely find a decent place to live. But even under those circumstances, you still have people at 50 percent of area median income, or 30, or 20, or zero. That population is going to need some form of support no matter how much market-rate housing we build.

If you want to get your hands around urban inequality, and deep housing cost burdens and homelessness in particular, you need to do more than build housing. You also need to stabilize the incomes of the people the market alone can’t reach and, with thoughtfully designed tenant protections, insulate them against sudden shocks. These two focuses are not in conflict.

It seems like the politics of housing is ripe for this kind of convergence. Check out what New York City Mayor Mamdani said in late May as he launched his housing plan:

The conversation around housing was framed as a simple binary, a question reduced only to yes or no, a side you had to take between structures and people: “Do you believe in building, or do you believe [in] fighting for tenants? Should we build our way out of this, or should we organize to preserve what we have?”

For a long time, the two were framed as mutually exclusive. [...] As we made it easier for New Yorkers to stay in their homes, one thing became increasingly clear. There was no way to drive down housing costs without also building more housing. [...] And data from other American cities told a clear story of what that building could lead to. Between 2015 and 2024, 120,000 homes were built in Austin. In December of 2021, Austin's median rent was $1,546. By this past January, it had fallen to $1,296, even as the city's population continued to grow. I saw these same patterns in Minneapolis and Seattle. It was clear when it comes to housing, the binary we face is not between structures and people. It is not between building and organizing, nor is it between the tenants of today and the tenants of tomorrow.

It is a far more basic one. It is between a government that debates and a government that delivers. We can keep people in their homes, and we can build the homes that they need to live in.

I hope you check it out! We have a Zoom seminar this afternoon at 3pm ET if you are interested in watching us discuss this.

Two weeks ago I gave testimony to the House Committee on Oversight and Government Reform, Subcommittee on Economic Growth, Energy Policy, and Regulatory Affairs. It was for a ‘roundtable’ titled Reducing America’s National Debt: Rooting Out Federal Waste, Fraud, and Overregulation.

It wasn’t clear in advance if the focus was the debt, waste, or regulation, so my opening remarks (presented below) quickly touch on each. I took the opportunity to be critical of the Trump administration’s second term.

But to give you a sense of how President Trump is warping the Republican Party and making Congressional Republicans vestigial, it wasn’t actually a hearing. The testimony from me and the three Republican witnesses wasn't even entered into the record. This is because hearings give members the ability to request subpoenas and otherwise act by majority vote.

Some Democrats had been teaming up with Republicans to use Oversight subpoena powers to actually govern; this is how the Epstein files got subpoenaed. But rather than having hearings potentially result in bipartisan votes for accountability that might impact Trump, Republican leadership is just replacing hearings with fake ones called roundtables. (You can read more about this from Arthur Delaney here as well as The Hill here.)

The roundtable is on YouTube here. It was so small we could all sit across from one another. The closeness made it less likely I’d get hit with personal attacks or gotcha questions that have become an occupational hazard of testifying these days.

I will say, having done (actual) hearings during peak inflation in 2022, when everyone hated the economy, it was much more fun to point out people hate the economy even more out here in 2026. Of course consumer sentiment is lower than in 2024; inflation is higher, the health care cuts are brutal, and the President is on a war tour rather than dealing with anything people face. It was a blast to point this out multiple times!

My statement is below. You will notice it draws from many peer think tank groups; this gives you a sense on where the current thinking is on many of these topics.

Chairman Burlison, Ranking Member Frost, and members of the Subcommittee, thank you for the opportunity to testify. My name is Mike Konczal. I am Senior Director of Policy and Research at the Economic Security Project, where we advocate for ideas that build economic power for all Americans. Before that, I was Special Assistant for Economic Policy to the President at the National Economic Council during the Biden administration.

Debt

If you are concerned about the trajectory of the debt and deficit in this country, the One Big Beautiful Bill Act (OBBBA) was a generational mistake. It is impossible to reconcile being worried about the national debt with supporting the actions taken with that law.

At a moment of near full employment with elevated interest rates, there was an opportunity to think seriously about reconciling the medium-term trajectory of revenues and spending. The Trump administration did the exact opposite. CBO scored OBBBA as adding $3.4 trillion to deficits over the next decade (Congressional Budget Office, 2025a).

To partially pay for the tax cuts driving that deficit, the bill imposed severe cuts to healthcare, food, and income support, layered on top of tariffs that American consumers are now bearing in higher prices. You can’t take the debt seriously by having Americans sacrifice so the already wealthy can pay even less.

The CBO finds that under OBBBA itself, the bottom 20 percent of American households are worse off as a result of this law, losing a total $24 billion each year while the top 10 percent gain $190 billion. Once you also account for the administration’s tariffs, the Yale Budget Lab estimates that households in the bottom 70 percent on average lose ground. The bottom decile loses more than $2,000 a year, over 5 percent of their income, while the top decile alone gains nearly $10,000 (see Figure 1) (Congressional Budget Office, 2025b; The Budget Lab at Yale, 2025). This constitutes a major shift to the wealthy from everyone else, and these outcomes represent a striking failure of economic policy: they manage to shrink paychecks for the majority while simultaneously exploding the deficit.

Figure 1. OBBBA plus Trump tariffs drain income for the bottom 70 percent of households, while the top 10 percent net a sizable gain. Combined average annual change in household resources by income decile, dollars per household. Source: The Budget Lab at Yale, Combined Distributional Effects of the OBBBA and of Tariffs, December 2025 update (tariff policy as of November 17, 2025).

The reason we are on this trajectory is because of the repeated tax cuts that hollowed out our ability to raise reasonable revenues consistent with the promises we’ve made: for the social insurance that gives families income security, for dignified retirement, for the public investment that keeps us a world leader. Twenty-five years ago, there was no structural budget deficit, as revenues were expected to keep up with the rising costs associated with the retirement of the Baby Boom generations.

In 2012, before most of the Bush tax cuts were made permanent and before the 2017 Trump tax cuts were even on the table, CBO projected the 2025 deficit at 1.8 percent of GDP. In reality, it came in at 5.8 percent. The entire gap is a revenue story. Programmatic spending in 2025 came in slightly below what CBO had projected back in 2012. Revenue came in more than four percentage points of GDP below it (see Figure 2). Estimates from the Center on Budget and Policy Priorities find that without the Bush and Trump tax cuts, debt-to-GDP ratios would be around 55 percent, not the 100 percent it currently is (Duke, 2026; Bernstein & Kogan, 2026). Today’s deficits come from 25 years of tax cuts eroding the revenue base, not from new programs or runaway spending. The obligations Congress already made are arriving on schedule.

Figure 2. Permanent tax cuts lifted 2025 deficits well above 2012 projections. Actual 2025 vs. 2012 CBO projection — before Trump tax cuts, with permanent Bush cuts — as a share of GDP. Chart replicated from the Center on Budget and Policy Priorities; CBPP analysis of 2012 CBO projections and CBO Economic and Budget Update, February 11, 2026.

Waste

This committee is tasked with tackling government waste. The OBBBA is designed to create a giant, wasteful time tax on families who rely on Medicaid and other programs. Forcing eligible people to continuously refile paperwork, document their hours, and re-verify their income is every bit as wasteful and ineffective as the federal projects this committee has long criticized. It just shifts the waste onto working families. CBO’s own analysis finds that more than 10 million Americans will lose health coverage under this law (Georgetown Center for Children and Families, 2025). CBO has also found that under state-level work requirements, many participants lose coverage because they were unaware of how these bureaucratic processes work or found it too difficult to demonstrate compliance (KFF, 2025). Most of the people pushed off these programs are not ineligible. They are simply unable to bear the time-tax of paperwork, especially those taking care of children, seeking higher-paying employment, or working multiple jobs to make ends meet.

Worse, research from The Hamilton Project finds that the majority of SNAP and Medicaid recipients subject to work requirements are already in the labor force, but more than a third will periodically fall below a required hours threshold. This means they will lose coverage not because they stopped working, but because their hours dipped below a requirement in a given month. This is not by their own choice. Low-wage workers will lose coverage simply because of the volatile, unpredictable scheduling common in their industries, like service-sector work. Workers shouldn’t lose health care because their employers cut their hours (Ananat et al., 2025; Bauer et al., 2018).

If we are serious about waste, fraud, and abuse, we already know where it is. We know that the tax gap, taxes legally owed but not paid, runs at roughly $700 billion every year, with the top 1 percent of earners alone accounting for more than $160 billion (Internal Revenue Service, n.d.; Sarin, 2021). We know that every $1 invested in IRS enforcement against high-income taxpayers returns at least $5 in revenue, with returns running even higher at the top of the income distribution (Internal Revenue Service, 2024). Rather than address these clear sources of unfair waste, this administration has gone in the opposite direction, gutting the enforcement capacity the Inflation Reduction Act built up (Rubin, 2026; Bryant et al., 2026). We know that Medicare Advantage plans are overpaid by an estimated $80 to $140 billion every year through upcoding and favorable selection. The Committee for a Responsible Federal Budget projects $1.3 trillion in overpayments over the coming decade (Committee for a Responsible Federal Budget, 2026).

Regulations

Now when it comes to regulation, we are living through the worst regulatory environment for business on record under this administration. Companies dislike bad regulations, but they hate regulatory uncertainty. The Economic Policy Uncertainty Index hit its highest readings on record in 2025, with the average value during this term matched only by the months immediately following the COVID outbreak (see Figure 3) (Baker et al., 2016).

Figure 3. Average policy uncertainty under Trump rivals only the COVID shock. Legacy Three-Component Economic Policy Uncertainty Index, monthly, 2000–2026. Source: Baker, Bloom, and Davis, policyuncertainty.com. Dotted red line = average of monthly values across 2025–2026.

This administration has personalized regulation on behalf of its own interests, using federal power as leverage against specific companies and the people who run them. The President pressured Intel into giving the federal government a 10 percent equity stake, bragging publicly that the company’s CEO “walked in wanting to keep his job, and he ended up giving us $10 billion for the United States” (NPR, 2025). The FCC approved Paramount’s $8.4 billion merger with Skydance only after the company paid the President $16 million to settle a personal lawsuit, with the President then publicly claiming an additional $20 million in advertising commitments from the new owners (Newsweek, 2025; Variety, 2025). The Defense Department designated the American AI company Anthropic as a national security supply-chain risk, which is normally reserved for firms tied to foreign adversaries like China and Russia, over disagreements about what its models could do for the government (Fortune, 2026). The Department of Energy abruptly canceled $7.6 billion in already-awarded funding for 223 clean energy projects in October 2025, with all but a handful located in states that voted against the President, signaling to any developer that federal commitments can be revoked based on the politics of the zip code (Washington Post, 2025; PBS NewsHour, 2025).

No CEO making a capital expenditure decision today can be confident that next year’s deal won’t be the subject of tomorrow’s investigation.

Conclusion

Finally, if we are serious about affordability and the national debt, we cannot stand by while the President exerts personal political pressure on members of the Federal Reserve, including criminal investigations against a sitting governor and the [then current] Fed Chair, alongside sustained public campaigns demanding the rate cuts the President wants. This is a fiscal concern as much as a democratic one. Political pressure on the Fed creates a political instability premium. That premium raises rates on the national debt and on the credit cards, mortgages, auto loans, and small-business loans Americans use every day (The Budget Lab at Yale, 2024).

It does not surprise me that Americans are far angrier about this economy than they were in 2024. Consumer sentiment and economic approval polling have both deteriorated sharply this year, especially as inflation has accelerated. Americans see large-scale cuts to programs they rely on, and new limits on proven economic boosts for families that we should instead be expanding, like the monthly Child Tax Credit. They see an administration that is not prioritizing their kitchen-table issues or the long-term health of the government’s finances. There is still time to put the fiscal trajectory on a sustainable footing. But every month we delay, and every disastrous funding cut and revenue loss of OBBBA that gets locked in further, makes the work harder.

Thank you. I look forward to the discussion.

References

Ananat, E., Gassman-Pines, A., & Howard, O. (2025, May). Work requirements penalize workers in volatile occupations. The Hamilton Project. https://www.hamiltonproject.org/publication/post/work-requirements-penalize-workers-in-volatile-occupations/

Baker, S. R., Bloom, N., & Davis, S. J. (2016). Measuring economic policy uncertainty. Quarterly Journal of Economics, 131(4), 1593–1636.

https://www.policyuncertainty.com

Bauer, L., Schanzenbach, D. W., & Shambaugh, J. (2018, October). Work requirements and safety net programs. The Hamilton Project. https://www.hamiltonproject.org/publication/paper/work-requirements-and-safety-net-programs/

Bernstein, J., & Kogan, B. (2026, February). President Trump’s “Big Beautiful Bill” raises the fiscal gap to 2.4 percent. Center for American Progress. https://www.americanprogress.org/article/president-trumps-big-beautiful-bill-raises-the-fiscal-gap-to-2-4-percent/

Bryant, K., DeBot, B., & Leiserson, G. (2026, February 24). The bipartisan budget deal rewards tax cheats and sets up the IRS to fail. Tax Law Center at NYU Law. https://taxlawcenter.org/blog/the-bipartisan-budget-deal-rewards-tax-cheats-and-sets-up-the-irs-to-fail

Committee for a Responsible Federal Budget. (2026, March). New data suggests MA overpayments of $1.3 trillion over the next decade. https://www.crfb.org/blogs/new-data-suggests-ma-overpayments-13-trillion-over-next-decade

Congressional Budget Office. (2025a, July). Estimated budgetary effects of Public Law 119-21. https://www.cbo.gov/publication/61570

Congressional Budget Office. (2025b, August). How the 2025 reconciliation act (Public Law 119-21) will affect the distribution of resources available to households. https://www.cbo.gov/interactive/2025-reconciliation-act

Duke, B. (2026, March). Costly tax cuts increase our nation’s fiscal challenges. Center on Budget and Policy Priorities. https://www.cbpp.org/research/federal-tax/costly-tax-cuts-increase-our-nations-fiscal-challenges

Fortune. (2026, February 27). Trump orders U.S. government to stop using Anthropic but gives Pentagon six months to phase it out while Hegseth adds supply-chain risk designation. https://fortune.com/2026/02/27/trump-us-government-anthropic-claude-pentagon-6-months-phaseout-ai-standoff/

Georgetown Center for Children and Families. (2025, August 14). New CBO health coverage estimates of budget reconciliation law. https://ccf.georgetown.edu/2025/08/14/new-cbo-health-coverage-estimates-of-budget-reconciliation-law/

Internal Revenue Service. (n.d.). Tax gap projections for tax year 2022 (Publication 5869). https://www.irs.gov/pub/irs-pdf/p5869.pdf

Internal Revenue Service. (2024, February). Return on investment: Re-examining revenue estimates for IRS funding (Publication 5901). https://www.irs.gov/pub/irs-pdf/p5901.pdf

KFF. (2025, July). A closer look at the work requirement provisions in the 2025 federal budget reconciliation law. https://www.kff.org/medicaid/a-closer-look-at-the-work-requirement-provisions-in-the-2025-federal-budget-reconciliation-law/

Newsweek. (2025, July 24). CBS, Nickelodeon, Paramount Pictures head to Skydance after Trump lawsuit settlement. https://www.newsweek.com/paramount-merger-skydance-approved-fcc-trump-settlement-2103793

NPR. (2025, August 22). Intel will give the U.S. government a 10% stake, Trump says. https://www.npr.org/2025/08/22/nx-s1-5509673/trump-says-us-government-will-take-stake-intel

PBS NewsHour. (2025, October 2). White House cancels nearly $8B in clean energy projects in blue states. Associated Press. https://www.pbs.org/newshour/politics/white-house-cancels-nearly-8b-in-clean-energy-projects-in-blue-states

Rubin, R. (2026, April 12). America’s new tax mantra: “The IRS isn’t going to catch me.” Wall Street Journal. https://www.wsj.com/politics/policy/irs-staffing-tax-enforcement-1a18e33f

Sarin, N. (2021, September). The case for a robust attack on the tax gap. U.S. Department of the Treasury. https://home.treasury.gov/news/featured-stories/the-case-for-a-robust-attack-on-the-tax-gap

The Budget Lab at Yale. (2024, May). Political risks to the U.S. safe harbor premium. https://budgetlab.yale.edu/news/240502/political-risks-us-safe-harbor-premium

The Budget Lab at Yale. (2025, August). Combined distributional effects of the One Big Beautiful Bill Act and of tariffs (December 2025 update). https://budgetlab.yale.edu/research/combined-distributional-effects-one-big-beautiful-bill-act-and-tariffs-0

Variety. (2025, July 31). Trump claims Skydance to give $20 million in ads with Paramount merger. https://variety.com/2025/tv/news/trump-unconfirmed-claim-skydance-20-million-advertising-psas-paramount-deal-1236467234/

Washington Post. (2025, October 2). Trump officials cancel $7.6 billion in clean energy projects. https://www.washingtonpost.com/climate-environment/2025/10/02/clean-energy-grants-canceled/

Questions about vibes and AI are at the end, but also check out three more technical ones I asked: (1) was UI actually better than keeping people on payrolls in 2020? (2) how do we understand the relationship between the 2021-2023 wage compression and inflation? And (3) as AI research reorients us to thinking of the economy as an aggregation of tasks, instead of workers or firms, what can we learn from the idea of monopsony?

I plan on trying to do these more often; if you have a book you’d like to discuss please reach out yourself or through your publisher.

AI-generated transcript. Confirm via video before any quotations.

Introduction

Mike Konczal: Alright, we’re live. New television era. I’m very excited to be here talking about the book The Wage Standard with the economist and friend of the blog, Arin Dube.

I want to start by saying I love this book. I thought it was really good, and I highly recommend it. Three things about it. One: if you don’t know very much about the current debates around the labor market and wages and the minimum wage, it catches you up fast. It’s very accessible. Anyone who’s generally interested but doesn’t have background would learn quite a bit. Two: I try to keep up on this, and I learned a bunch. I got refreshed on things I’d forgotten, weird side debates I wasn’t aware of. So even if you feel like you know a lot of the stuff in this book, even if you know what monopsony looks like on a supply and demand curve, you’ll still learn from the book, which is impressive. Third: even if you don’t like these arguments, even if you’re very conservative or very neoclassical, or you just don’t like left-liberal economics, you’ll still get the best version of these arguments. They’re presented fairly, with technical parts in the appendix, but you’ll be caught up on how people in the broader center-left coalition are thinking about this.

So congratulations on a great book. It reflects a whole career of excellent work on this. I want to start by asking: what led you to write it? Economists don’t often do books.

Arin Dube: Thanks for having me, and thanks for the conversation.

Mike Konczal: You should do an introduction too. I forgot to introduce you more properly. Talk about where you teach and what you study.

Arin Dube: I’m a professor of economics at UMass Amherst, here in Massachusetts. That’s where I’m talking from.

The book is a culmination in some ways of a lot of work I’ve done on the topic. I’ve written on what minimum wages do. I’ve written about market power in the labor market, monopsony, as the word goes. But most of that is geared toward academics, particularly other economists; I’m trying to publish in economics journals. More and more it’s struck me that ultimately, if we want to have a conversation about wages, it needs to be done in a broader context, with people who aren’t studying economics. And yet these ideas are important to understand and wrestle with.

So I really wanted to have a broader conversation, and this book is an attempt at that, about a really simple point at a certain level. That simple point is that most American workers deserve a raise.

What’s the basis of the argument? The starting point is that between 1980 and 2019, overall productivity in America grew by something like 73%. People at the top, their wages grew by something comparable. But the median wage grew only by 23%. And at the bottom, the 10th percentile, even less. So the question becomes: why did this happen, and what can we do to make it better?

That’s really the core of the book. Let’s try to understand how the labor market works, and when it works well for people, because sometimes it does. But other times, potentially for long stretches of time, it doesn’t. Then try to unlock that mystery and actually make the market work better, with supporting institutions that can ensure more broad-based prosperity.

In many ways, I started writing this book early in the pandemic. I started writing, for example, about when the market can produce broad-based prosperity, because we actually had, for extended periods, broad-based wage growth in the United States during the 50s and 60s, and even more recently, in the late 1990s and the late 2010s. So the question becomes: what’s going on in those periods that makes us grow together, when at other times we just grow apart?

While I was writing this, we went through a period of a very different sort of labor market, a very tight one, with really strong wage growth at the bottom. It was uncanny because I was writing about ideas that were actually happening right outside, which made it exciting but also challenging, because I was trying to figure out what was going on in real time. The aim is really to engage with this broader set of topics, and hopefully with a broader set of readers, about what we can do to make sure there’s better pay in this country.

How Wages Are Set: Skills, Firms, and Monopsony

Mike Konczal: That’s great. You mentioned the pandemic. We have some viewers, so thank you for joining us; feel free to throw questions in the live chat. We’re definitely going to talk about all things pandemic, the Beveridge curve, inflation, and the vibecession. And the last question, in about 40 minutes, will be on AI. So if you’re excited to hear Arin on AI, you have to wait until the very end.

The book is great because it presents things as a series of economic ideas and debates, and then walks through the literature, how it evolves, and the tools, technology, and data you have access to. There are two big conceptual debates whose evolution you explain. One is what happens with minimum wages, a view that sees markets as very competitive and thus assumes a minimum wage would kill off a lot of low-wage jobs. You introduce the audience to the notion of monopsony and explain that. On the other hand, there’s this idea of human capital. People may have heard of it in an education context: the idea that wages for most people are basically set as a function of personal skills. You bring in the idea of superstar firms, and the idea that firms really matter. I’d love for you to talk through both of those as a set of economic ideas and debates.

Arin Dube: Let me start with the idea of how wages are set, because there’s a really interesting historical arc here. One view, like you said, is basically that wages are rewards for skills, compensation for what skills people bring to the market. That’s certainly a very important part of what wages are, but it’s not the only part.

The idea that employers actually have some degree of choice, some discretion in setting pay, is at the heart of what the word monopsony means. Literally, monopsony means a single buyer of something, in this case, labor. But that’s not really what the theory of a monopsonistic labor market is about. There are many employers. It’s just that employers have some degree of choice, some degree of market power, where they choose: are we going to go for a high-wage strategy, or a low-wage strategy? That choice can lead to different people getting paid differently, and it can also explain certain mysteries about the labor market. I’ll come back to those.

The basic idea that employers have some choice in setting pay goes back in some ways to Adam Smith. But the term monopsony was coined by Joan Robinson in 1933, a British economist well known for many things. In this case she laid out the theory. In the mid-20th century there was actually a good amount of empirical work supporting it, much of it done by more institutionalist labor economists who went and surveyed factories. Sumner Slichter, for example, surveyed a bunch of factories in the Boston area and documented that, look, there are actually different wages being paid in the same labor market for fairly narrow occupations.

Those ideas had some currency at the time, but they weren’t fully convincing because some economists were concerned that maybe these workers weren’t really the same. Even within narrow occupations, there are differences in skills, effort, and so on, and those are what drive the wage differences. So it took some time, and it took the advent of better data to settle the question. The better data took the form of matched employer-employee data, administrative data that allows us to track nearly everyone in the labor market as they hop from one job to another, and see what they actually get paid.

With that, you can see what happens if someone changes jobs. For example, we know Walmart wages are lower than wages paid by Target, a similar retailer. But is it the case that the same worker gets paid differently if they actually work at Target versus Walmart? The answer turns out to be: yes, they do. This reflects that it’s not just a matter of the skills you bring to the table. It’s also a choice employers have about what sort of wages to pay.

That’s a big part of what I try to highlight in the book. In many ways, the wage stagnation that has happened in different periods over the last 40, 50 years can be thought of as reflecting a set of choices: choices by employers about what wage strategies to pursue, and choices by policymakers about what supporting policies and institutions to provide.

Even though these ideas have been around for a while, we’ve seen a huge increase in research on these topics in the last 10 years. So it’s a particularly good time to tell these stories, because they’re also very active and vibrant research topics in academic economics. Connecting that to broader policy-relevant questions, with a broader audience, is one of the aims of the book.

The Minimum Wage: Natural Experiments and Channels of Absorption

Mike Konczal: That’s great. You do a great job of summarizing the minimum wage literature. Again, for people without background, but also for people with it, you have a good sense of how those debates evolved, what evidence was brought in at different points, and where it stands now. You talk about how for some people, if the minimum wage goes up, obviously something has to fall apart. And you talk about three Ps, productivity, prices, and profits, and what happens, and what we should think of it, especially in the broader center-left.

I also want to bring up an argument you hear sometimes. I’ve seen this on Econ Twitter. There’s a paper by Anna Stansbury and co-authors saying, yeah, there’s no job loss with the minimum wage, but the jobs might get worse. In their paper, there might be more injuries. I don’t know how robust that is. How should people who support higher wages and support minimum wages think about the likely consequences across these different dimensions?

Arin Dube: I’m going to go back a little bit and then come back to the question about the channels. So just so everyone’s clear: we used to set the minimum wage at the federal level and adjust it fairly regularly. Between 1938 and 1980 or so, the minimum wage would go up periodically and it kept up with overall productivity of the economy, for the most part, and with wages generally. That was the basic pattern.

Then something happened. Starting with President Reagan, who did not raise the federal minimum wage during his two terms, we started going for really extended periods without any increase in the federal minimum wage. We’ve now gone over a generation. In fact, we’ve gone something like 17 years without raising the federal minimum wage, which is a first.

In general, this is a terrible way to set policy. Whatever you think the right level of the minimum wage is, going for 15 years or more and then suddenly increasing it is not a great way to do it. It’s a dysfunctional way of setting policy. But there’s a silver lining to a dysfunction like that: it becomes a natural experiment that allows us to actually study what happens. Because when the federal government stopped raising the minimum wage, some of the states stepped in. States as the laboratory of democracy in the United States is a long tradition, and the minimum wage became one such thing.

Suddenly you have these natural experiments: New Jersey just raised its minimum wage in 1992, but neighboring Pennsylvania did not. What happens? This, of course, is the very well-known study by David Card and Alan Krueger that I talk about in the book, which really shook up the discipline of economics and pushed us to think hard about the impact of the policy. Because they found that it surprisingly did not lead to fewer jobs in fast food in New Jersey compared to Pennsylvania. That was a big shocker, because most economists thought the opposite would happen. In the mid-80s, The New York Times had an editorial saying the right level of the minimum wage was zero, relying on the then-common sense that this is something with a large negative impact, including reduced jobs.

At this point we have a bunch of work that builds on this original Card and Krueger study. Some studies differ. Some find positive employment effects, some negative. But overall the body of evidence suggests employment effects are very small. One important takeaway: if we look at the last 15 to 17 years, we’re running a particularly profound natural experiment, where you have 30 states that have raised their minimum wage and 20 that have not. These differences have now been around for over 12 years. These are not short-term effects. You might think, oh, maybe in the short run things are muted but in the long run things matter more. I don’t know what the long run is, but 12 years is a pretty long time.

Anyone can do this. It’s very easy. Just look at restaurant employment per capita in those 20 states versus the 30 other states and track it. You don’t have to do anything sophisticated. It’s very clear: wages have grown a lot more in the 30 states than in the 20 others, and employment looks pretty much the same. It becomes really hard to argue that the kind of minimum wages we’ve actually enacted in the United States have had a large employment effect.

So then the question is: if it’s not employment, what is it? What are the channels of absorption? Here too we have a good amount of evidence. One thing that happens is that as these jobs become better, fewer workers quit and fewer leave. So it reduces turnover. The monopsony theory we talked about also helps us understand why employment may not fall. Simply put, when employers are choosing to pay a lower wage, they’re going to have more vacancies. They’re okay with that because, even though they don’t like vacancies, they also don’t like to pay higher wages, so they’ll live with vacancies more than raise the wage when they’re trying to maximize profit and have some monopsony power. But what happens if the government steps in and says you have to raise wages? Then your vacancies may actually fall, because you’ll have an easier time recruiting and retaining workers. So you may end up killing vacancies and not killing jobs to a certain extent. That’s in fact consistent with a lot of the evidence we find.

This also has a knock-on productivity effect, one of the three Ps. We see it in lower turnover. We also see it more directly in, for example, the retail sector, where value-added per worker rises beyond anything caused by changes in prices. The third channel is prices, because prices do go up somewhat when the minimum wage rises. But the way to think about it is that this is a fairly small increase in price compared to the much larger increase in wages for low-wage workers. Overall, higher-income consumers end up paying a little more for a burger to substantially raise wages at the bottom. And finally there’s some reduction in profits, which again is consistent with the idea that employers have some degree of market power.

Now, on your question about other potentially negative impacts on job quality: the Stansbury et al. paper you mentioned used city minimum wage differences across California and found that higher minimum wages are associated with more injuries. I take the paper seriously. It will be important to see if it replicates in other contexts. But it’s important to keep in mind the broader issue. The question is, are these jobs overall better? The answer is overwhelmingly yes. Even in their estimates, only a small fraction of the job value is offset by injuries.

We know these jobs are getting better because, as I said, you see very sharp reductions in separations. A great example comes from a paper I just put out a couple of weeks back, looking at the California fast food minimum wage. It’s a particularly high sectoral minimum wage affecting large fast food chains. Because these jobs are substantially better now than other jobs, turnover reduction is just off the charts. It fell a huge amount. That’s basically people voting with their feet, or in this case, not voting with their feet, not moving, which reflects that these are much, much better jobs.

So to me it’s pretty clear that a large body of evidence suggests job quality rises. Are some offsets possible? Yes. But, and I think the authors would agree, the implication, if there were greater injuries from work intensification in some settings, isn’t that you shouldn’t raise wages, but that you should also have better OSHA enforcement and stronger workplace standards. Higher pay and safe work aren’t substitutes. They can be a joint policy problem.

Full Employment, Wage Compression, and the Post-COVID Inflation

Mike Konczal: That’s great. I think it’s good to talk about COVID and the business cycle associated with it now and do some technical stuff. As a broad thing, your book has a great discussion about full employment and how, especially since 1980, basically the only time real wages adjusted for inflation actually go up over longer periods is the handful of years in which unemployment is below 5% or close to whatever is estimated as the natural rate. In periods where unemployment is quite high, as it’s been since 1980 and especially during the Great Recession, real wages don’t grow. They may even fall for many people outside of that tightness.

Obviously we had a period of very tight labor markets following the reopening in 2021 and 2022 in particular, with lots of job openings available and the unemployment rate reaching basically its lowest level since the 1960s, 3.5% for a brief period. Even now, with all the chaos, it’s still about 4.3% over the past year and a half.

Let’s approach the first question in a couple of steps. There was also very significant wage compression. You document this with co-authors and it’s in the book as well. Wages at the bottom grew much faster, maybe even 10% adjusted for inflation, compared to the top, where they grew much slower (more negative at the beginning, then picking up). So you have significant wage compression, wages at the bottom growing faster than those at the top, when in general over the last 50 years, especially since 1980, it’s been the opposite. That’s the definition of rising inequality, and that’s what we’d seen.

Should we understand that wage compression to have been an important component of the inflation wave that we saw during that period? CPI growth peaks at about 10% in 2022. A lot of that is the war in Ukraine, gas prices. But we also saw huge shifts basically across the board, including in what we often think of as wage-intensive, labor-intensive services. How do you see the relationship between the two? Did we have to pay an inflation cost to get that wage compression?

Arin Dube: It’s useful to think about what drives the compression. What drives the compression is a tight labor market. The question is, does a tight labor market necessarily cause inflation? Of course, higher wages: if I suddenly increased everyone’s wages by 10%, that would tend to have an impact on prices. But the question isn’t that. The question is, when you have a low unemployment rate, does that always produce inflation?

In a normal environment, without a major supply-side disruption, the answer often is no, not necessarily. One of the things I talk about in the book is how the Federal Reserve learned in the late 1990s, and then again in the late 2010s, that in fact the unemployment rate could fall a lot more than was the conventional wisdom, and that did not produce an inflationary boost. There was no big inflation in the late 1990s. There was no big inflation in 2019. In fact, the inflation rate was undershooting the target. But those periods also produced some degree of compression and broad-based wage growth, particularly strong wage growth at the bottom.

One of the things I argue is that if you look at the period between 1979 and 2019, pretty much if you just look at the seven years with a really tight labor market, and some evil-genius supervillain snapped their fingers and made those seven years disappear, what would happen? Well, if you’re at the top, the 90th percentile, your real wage growth averaged over this period would go from about 1.1% to 1%. Barely anything. But if you’re at the bottom, the 10th percentile, your real wage growth would go from an already small 0.3% to zero. Nothing. Nada.

So tight labor markets are really the periods when you actually see real wage growth, and lots of those have happened without producing an inflationary burst. We did have an inflationary burst, which was very painful for many people, following the reopening after COVID. The question is, was that inflation a necessary part of the compression? I would argue not. There are a few ways to see why.

One is to look across the U.S. and ask what happened in states that became particularly tight versus other states that did not. Using that cross-state variation, you can fit what’s called a wage Phillips curve. Florida was a particularly tight labor market. Did wages rise more there than in Massachusetts? The answer is yes. In tighter labor markets, wages grew faster, particularly at the bottom. That’s step one.

Step two: did prices also rise particularly strongly in those states where the bottom wages rose? You can use regional price differences and say, let’s deflate the wage using real regional prices and not a national-level price index. If it turned out that bottom wage growth really fueled most of the inflation, then once you use regional prices, the bottom wage growth would look a lot more meager. It would partly undo itself. Is that what happened? We’re in the process of revising the paper, and we show: no, it makes a very tiny difference. Regional prices did not sharply pick up in the same way that bottom wages grew in tighter labor markets. That makes me think wage growth at the bottom was not a major driver of the overall inflation.

In fact, wages were struggling to keep up with price growth. So yes, wages grew, but in some ways they had to grow. Otherwise you’d have major negative real wage growth. And we did have negative real wage growth for many people in 2021 before wages actually picked up. It’s a messy situation. Learning the lessons of what a tight labor market does when you’re in the midst of a major supply disruption is not the ideal way to learn. But that’s what happened.

We can also look at other time periods, which is why it’s important to compare what happened when we had a tighter labor market that was not concurrent with a major supply-side shock. That helps us understand the broader lesson. And the broader lesson is really important, because we’ve actually spent more time closer to full employment in the last 10 years than we did in the 40 prior. Not surprisingly, we’ve actually had better wage growth in the last 10 years than we did in the 40 years prior. But if we fail to understand that lesson and say, “don’t bother with full employment, it’s going to lead to inflation,” even though I don’t think that’s true when it’s not coupled with a major supply disruption, then we will potentially lose that. And if we can’t extract the signal from the noise, it’s going to be hard to defend gains that we’ve actually made. And we have made gains, as I discuss in the book.

Mike Konczal: Yeah, absolutely. As you said, overall wages were negative at peak inflation in 2021 and 2022. If you run it under more sophisticated Fed modeling, in general prices are leading wages during this period. Wages don’t predict prices, they don’t come first and cause businesses to raise prices in response. What you see in sophisticated models is that nominal wages (the number on the paycheck) jump but the real value falls during peak inflation, and then afterwards nominal wage growth slows. So you’re getting less of a big number on the check, but you can buy more. That fits a supply-shock story, where workers are trying to catch up to all the stuff that’s happened and they follow this mimic pattern as the supply shock comes off. That’s kind of my basic model of the story, and I think it fits pretty well.

Also, if it really was wages causing it, lots of things can happen in the world. Models aren’t concrete, right? But you’d have to expect a real labor market disruption in order to bring down inflation, which we did not see. Unemployment went from 3.5 to 4.3, but basically 4.1 by the time the disinflation was over. That’s a zero increase, when most people thought you’d need a big jump if wages were really the central driver. And the inflation was everywhere, but the wage growth wasn’t.

Unemployment Insurance, the Great Reshuffle, and the Job Ladder

Mike Konczal: I have a more technical question, but I hope people stay on. There was a really big debate early in the COVID recovery about what to do with the fact that all these businesses had shut down. Speaking at a very high gloss: a lot of peer countries, especially in Europe, basically paid employers to keep people on payroll, “air-quote” working, even though they weren’t, to keep that employment relationship going. In the United States, we basically allowed the severing of that relationship by giving money to people in the form of unemployment insurance, very stepped-up UI, which is a whole other conversation, but very generous for the United States and I think quite revolutionary. Crucially, people were unemployed for a period.

A lot of people at the time were very nervous: oh no, being employed is important for people, for their skills, for their sense of purpose; thus when we try to reopen, these people will be dislocated and it’ll be hard for them to find work. My read, and I’m curious where the labor economics literature is coming in on this, is that because people had money in their pockets, demand was strong enough that there were lots of job openings. It was very easy for people to find them. Those job openings skyrocketed and then fell mechanically as we reopened, which I think threw off a lot of people. That resorting allowed people to upskill into better jobs. If you look at the aggregate data, and you talk about this in your book, a lot of the wage growth was from job switching, and a lot of the switches were into higher-paying, better jobs. I call it the great upskilling; people used to call it the “great quitting,” but that wasn’t accurate, and you have a great history of this in the book.

Obviously a lot of things are in play here, and we were more unequal so there’s more room for wage compression. But I’m curious what you make of that initial decision to go to unemployment insurance instead of preserving matches. Because there’s a Fed study I’ll link to when we put this online that I think links it to the unique productivity growth and wage growth we saw.

Arin Dube: Look, I’m on the record in early 2020, right at the very beginning of the pandemic, on policy calls with other economists saying we have to preserve the matches. Because it’s really costly if you don’t, to build good matches again. In Germany, for example, they use work-sharing. Instead, we’re just throwing people out of jobs and onto unemployment benefits. But apparently that’s the only way we can do it, then we need to provide a generous UI benefit. But gee, wouldn’t it have been better if we could have done what Germany was doing and preserved the matches? That was my understanding, and I think it was a common one. It wasn’t unreasonable.

But here’s the thing. What if I tell you those matches weren’t that good? Maybe they were crappy matches. And there’s a reason for that: we spent over a decade after the Great Recession where the job ladder had just stopped working. For listeners, the job ladder is basically the process through which people move from worse jobs to better jobs. Generally you’re going to have this churn, and generally upward movement in terms of job quality. That just stopped after the Great Recession. Maybe by 2019 it had really kind of started picking up. Quit rates had risen. But that meant we had this huge stock of pretty poorly paid positions, more so than even at other times in our country’s history.

So it was ripe for what I call the great reshuffle, where people basically move out of these positions and find better ones. And what better time to do that if you’re already one step removed from the work? And yes, if you have some money in your pocket, you may be a lot more willing to look around.

Now, the part that’s not accurate is that some people felt more generous UI benefits would lead people not to work. There’s a good amount of research on this, and while there’s some disagreement, the general set of findings is that it had fairly modest impact on whether you were working or not. But it may have played a role in how likely people were to actually look and search for better jobs.

The other interesting thing: there’s evidence that people who are at the worst-paid or really poorly paid positions often don’t realize there are better options out there. This has been shown in the data. So maybe suddenly, if people start moving and you learn that half your co-workers quit, presumably they got some job, you think, maybe I should look around more. There may be a multiplier effect. This is hard to know for sure. There is, of course, people who quit and put it on TikTok, “how I quit,” and those things may have accelerated some of it.

But whatever the reason, what is absolutely true is that this led to a reallocation of work, a reshuffling of people up the job ladder. That in turn put pressure on those low-paying positions to upgrade their pay. That was important. It also means people were more likely to be at more productive companies and workplaces. That actually has what we call in my work with Autor and McGrew a “double dividend.” You’re actually increasing productivity in the economy by reallocating workers to higher-productivity locations.

What’s also interesting is that something like this also happened, a bit before COVID, in Germany. We see a reduction in wage inequality from reallocation and a tighter labor market in the 2010s in Germany. The authors of that paper said they saw what was going on in the U.S. that we were documenting, went back, and found evidence consistent with it. So I think we’re gaining a general understanding of why a tighter labor market is not just good for workers, and of course it’s good for people looking for a job when there are more jobs. It actually makes the market function more efficiently, in some ways more competitively, than if you have a very slack labor market. Which highlights, again, the importance of a full-employment economy.

The Vibecession

Mike Konczal: We’ll start to wrap up with two more questions. Up to this point you’ve been very grounded in the evidence. Everything you’ve alluded to or mentioned has peer-reviewed studies behind it, and you’ve really dotted the i’s and crossed the t’s. The next two questions we can be a little speculative on: the vibecession and AI.

We had this great recovery in terms of GDP and jobs. It’s not so much that consumer sentiment and negative opinions on the economy are negative because there are good reasons to be concerned. Housing markets are very tough right now, among other things. But sentiment readings are at record lows and they’re much more pessimistic than we’d expect given the fundamentals. Various studies try to put more things in to predict why consumer sentiment and opinions on the economy are so low. They tend to point to the fact that prices went up, but sometimes (I’m not a hardcore rational expectations guy, but I always want a little bit more than that), I want more. What do you make of the way consumer sentiment has evolved over the last couple of years, given the labor market and wage growth?

Arin Dube: This is interesting, because of course in the last year there’s some pretty objective reason for people to have that vibe. A lot of us have said this on Twitter, that inflation has eroded real wage gains for the last year. That’s an important thing to keep in mind.

But this has been going on for some time. This is tricky because there’s some amount of negativity bias that might arise. Some people have argued and shown evidence that economic news coverage has at times been more negative than fundamentals warranted. But at a certain level there was some legitimacy to the bad vibes. What I try to do in the book is to acknowledge that you can have legitimately bad vibes, but it’s also important to tease out the positive parts, because otherwise we get stuck in a self-defeating loop.

What’s the reason for the bad vibes? Part of it is something we’ve known for a while: people don’t like a jump in the price level. That’s really clear from a bunch of pieces of evidence. People don’t like higher prices, even for the same real wage. But it wasn’t even that real wages always rose. In fact, real wages first fell before they caught up and grew, and it was messy.

For low-wage workers, the bottom third of the workforce, real wages did grow very strongly. For example, the 10th percentile grew by a total of 15% real wage growth after COVID, between 2019 and 2025. That was about the same as it had grown in the 40 years prior. So five years versus 40 years. But if you’re in the middle, that wasn’t really the case. So there was a lot of dissatisfaction.

What I do think is important, again, this is me talking about extracting the signal from the noise, is that if we conclude that nothing happened, that no gains were made, that things were just really terrible, then we lose sight of the fact that we actually have had better real wage growth, more broad-based real wage growth, in the last 10 years than we did in the 40 years prior. If we don’t get that, then we get all the doom and gloom, and we lose sight of what works. And that’s my real plea: we have to understand what works. You can be unhappy. There are reasons. The book documents the long-run reasons for legitimate dissatisfaction with the way the American economy has worked and not worked. But if we don’t get what works, then we’re going to repeat our failures of the past, and that would be bad.

Mike Konczal: Yeah, I agree. And it’s tough because you think of the trade-offs. There was a lot of fiscal stimulus in 2021. Maybe that added one percentage point to inflation. The estimates are very complicated. But if that also gave you a real wage recovery and a really good labor market recovery that itself wasn’t particularly inflationary. If you had the same inflation, but the real wage growth would have been much lower. And unemployment didn’t really fall below 5% or 4.5%. How much worse would that situation have been? We see it in a lot of peer countries.

Arin Dube: Exactly. We see the economic consequences, and we see very similar political outcomes of anti-incumbent sentiment. So the idea that if we’d simply done no fiscal policy, or monetary tightening had happened a lot earlier, yeah, I believe that could have led to some reduction in inflation. But look around: not every country, or even most countries, did the fiscal and monetary stimulus we did, and it didn’t lead to better outcomes overall for the typical worker in those countries. So be careful what you wish for, because you could get stuck with much of what you don’t like, plus some new bad things. That probably wouldn’t be good.

AI and the Labor Market

Mike Konczal: To wrap up: artificial intelligence. Obviously everyone’s very concerned and focused on the labor market. Obviously the people creating it are making very big predictions about what’s going to happen to the labor market. I talk to random people at daycare pickup, and someone I know is retraining to be a nurse because she was doing work that’s very AI-exposed. There’s a really wide anxiety about it. As an empirical labor economist, what are you watching for? How are you approaching this as a question about what could happen?

Arin Dube: It is just really weird that I can’t think of too many cases where we have a major technological change and the people really advocating for that technology are mostly out there saying how it’s going to destroy the economy and livelihoods. It’s a very peculiar situation. I don’t know what’s making it that way.

Mike Konczal: It was probably good for fundraising. It’s probably bad for building data centers, and we’re seeing a lot of that.

Arin Dube: But frankly, we do not know. I do not believe anyone really knows. There are layers of uncertainty. Part of it is in the technical space: what happens in terms of actual large language model development, whether it’s going to continue at the same rate or not. Then there’s the economic part, and this is where I think people are sometimes very naive about how adoption actually works. There are lots of examples of new technologies that didn’t lead to the miracle productivity growth you might expect.

One way economists think about this is the O-ring theory: when you have a bunch of tasks, you need all of them to work. Maybe you can improve productivity on nine out of 10 tasks using AI, just to be really generous. But then you’re stuck with that one task where there’s no productivity increase, and that prevents you from getting the boost you might otherwise expect. So how those tasks fit together, and how workplaces rearrange them, are the big questions. They can limit both the gains and the losses from AI.

In general, I’m very agnostic about where I see the impact, because I can see arguments that it would substitute for tasks of a certain sort, but it can also augment the capabilities of some people in ways that could boost productivity, including for workers who haven’t really gained much from technological change in the last 40 years, including blue-collar workers. So I can see pretty different types of impacts on wages and inequality. There’s just a lot of uncertainty.

What I’m trying to pay attention to, frankly, is my students. I see this in my classroom: students who are graduating, there’s a lot of nervousness and dissatisfaction and uncertainty about what’s going to happen. That’s really painful. And it’s not aided by some wild speculations we sometimes hear from people.

One thing I think is clear is that there’s always a tendency to over-interpret the data. For example, there was a lot of discussion that young college graduates more exposed to AI were the ones seeing unemployment rise. But there’s a strong counter-argument that the unemployment rate rise has actually been broader, including for non-college workers, which would be hard to explain using AI as the main thing. So be careful about interpreting secular trends from cyclical patterns. That’s probably always a good thing to keep in the back of our minds.

Mike Konczal: That’s a great way to conclude, but I want to ask one more technical question. I’m just curious on your quick opinion. You’ve got two more minutes. I’m noticing, obviously trying to read a little bit more into the sophisticated modeling of what AI might do. In Acemoglu’s papers from the late 2010s there’s this real focus on approaching the economy as a series of tasks, where previously people maybe were thinking of human capital (individual workers) or aggregates of firms. Now it’s an aggregate of tasks. Does monopsony still provide a useful framework if we’re all going to start wrapping our heads around the economy as a series of tasks over the next few years?

Arin Dube: Absolutely. In fact, one of the things in the revision of the paper with Autor and McGrew is a model of how technological change interacts with tightness and labor market power, monopsony power. We argue that in fact a slack labor market makes the impact of technological change more unequal. This is very relevant when we think about AI. Imagine you have some technological change that displaces workers. If you have other forces tightening the labor market, you’re not going to have the same kind of wage differentiation and wage inequality growth that might otherwise happen. In contrast, if you have a very slack labor market, there’s nothing pushing back against the monopsony-increasing aspect of technological change.

So this is where thinking about how we govern the technology is also a really important point. In the book I talk about the Writers Guild example. The Writers Guild union struck and were able to get not just better streaming residuals and better pay, but also the ability to have protection against the use of AI to replace the work they’re doing. That’s a good example of why we need better institutions, including the sectoral approach I put forward in the book. In a world of AI, in a world of technological change, you need better governance to decide and to shape how the technology affects the labor market. We don’t have to think of the technology as simply something that happens to us. We can also shape it. And I strongly believe that.

Mike Konczal: That’s a great way to conclude it. Arin Dube, the book again is The Wage Standard. I highly recommend it. Thank you again for taking the time.

Arin Dube: Thanks for having me.

But let’s start with a video of my favorite Powell moment and then walk through a little bit about why his legacy will be a bit weird in the short term.

@mtkonczalGoodbye Jay. You were awesome. #jaypowelllastday #jaypowell #fomc

Tiktok failed to load.

Enable 3rd party cookies or use another browser

July 10th, 2019, Jay Powell is testifying before the House.

Representative Ocasio-Cortez: In early 2014, the Federal Reserve believed that the long-run unemployment rate was around 5.4%. […] Unemployment has fallen about 3 full points since 2014 [to 3.7%], but inflation is no higher today than it was 5 years ago. Given these facts, do you think it’s possible that the Fed’s estimates of the lowest sustainable unemployment rate may have been too high?

Powell: Absolutely. <laughs>

Having spent some time thinking about this exchange both before and after it happened, I still remain impressed with Powell’s answers and follow-up. (You can see it all on CSPAN here.) It helps that Ocasio-Cortez is a serious politician who took the novel step of asking a relevant and interesting monetary policy question to the Federal Reserve chair. (The Republican representative right before AOC asked Powell whether or not he believes in the market system and whether or not the United States was the greatest economy in history.)

Powell, in his answer and the full clip, demonstrates both command of the material and humility. When AOC follows up on the flattening Phillips Curve, Powell is honest and open about what the Fed knows and what it doesn’t. I think this temperament served him well navigating the great labor market of 2019, the inflation wave that hit, and the soft landing. A more dogmatic and less open Federal Reserve chair would have cut off that 2019 economy (AOC was right that the general consensus was that unemployment couldn’t go below 5%). They also could have forced a recession in 2022-2023 for inflation to come down. But he understood his role was to maintain inflation expectations while staying open to the idea that they could solve it without causing a recession, which they did.

These are admirable qualities in a central banker, and I hope they outlast him.

Now you might read this and watch that clip and say, “Aha, Powell’s dovishness caused the great inflation!” But in the short term, this legacy is weird. People want to hate him for the inflation wave we’ve seen. Defenders point out that the inflation was global and credit him for the soft landing. I just did. But the soft landing was global too.

Let’s grab some quick OECD data. Here’s headline CPI inflation for the United States and several peer countries. This is year over year, over the past several years.

As you can see, it really was global. We could redo this with different measures of inflation or bring in more countries, but, as someone who has sweated these details, it’s all the same thing. It’s hard to either credit or blame Powell for this specifically. In retrospect I’m sure he wished he raised rates earlier, but I have a hard time thinking it would have changed this dynamic, and most realistic estimates (e.g. Reifschneider, 2024) don’t have much of a difference.

But how unique was the soft landing? Let’s also pull unemployment rates and GDP rates from OECD for these countries. Let’s plot the amount of disinflation for each country, from their highest to lowest values, against the annualized cumulative change in the unemployment rate during this period. Not the most sophisticated, but we’ll get the same results if we do the bells and whistles.

The general thought was that you’d need two percentage points of unemployment to bring down inflation one percentage point. This is called the sacrifice ratio with a value of 2, and it is reflected in the solid line above. As a matter of geometry, the soft landing is that you can have high disinflation, a value far to the right, without being near the line. But the actual measured sacrifice ratio is essentially zero. There are many estimates of the sacrifice ratio in the economics literature, generally in the 1-3 range. If the value is near zero, then something else is going on.

Now unemployment is doing a lot of crazy things during this time period, and we’re not trying to adjust it to compare against more normal rates. So let’s look at GDP growth also.

Following Ball (1994) (more methodology details here), we draw a GDP trend line from when inflation peaked to its lowest point and count up the deviations to get the cost to GDP of reducing inflation. As the graphic above shows, that number is also essentially zero. For the United States, it straight up has the wrong sign. The sacrifice ratio is negative.

Here’s a table with all this information.

The global soft landing looks easy in retrospect. But to give a sense of the outside chance of this, back when it looked impossible I said I would get a tattoo of the Phillips Curve if we pulled off the soft landing. Which we, and I, did.1

If a variable has the wrong sign, as βy does, then we probably need to look at other variables. And there is one you can see on my left forearm: the cost-push shock v term. If that term did the work, Powell’s job wasn’t engineering a recession to break inflation. It was holding expectations steady while supply healed, and that’s what he did.

If President Trump wanted Powell to be remembered poorly, he single-handedly screwed that up. By putting so much political pressure on him in such a comical yet terrifying way, Trump all but ensured Powell will be remembered quite well for standing his ground. I think, with time, his economic record will be remembered as well too.

There’s been a pickup in the labor market in the six months of employment data we’ve gotten since November 2025. The unemployment rate has ticked down slightly, by around 0.2%, and has since held steady. More interesting, job gains have increased and broadened, though only to a point.

(We’re going to stick with private job growth because DOGE's deferred resignation program packed a lot of federal job losses into October.) Since November, when the unemployment rate peaked, private job growth has been 68.3 thousand a month, compared to 23.1 thousand a month during the first 10 months of the Trump administration. (Though it remains lower than the 85.1 thousand private sector jobs a month in 2024.)

There’s a really funny irony in these new numbers, which the Trump administration is promoting as a win. When they took office, the Trump team said the economy was in a secret recession because so much of the economy was dependent on health care jobs. Stephen Miran of the CEA and Secretary Bessent argued at the beginning of 2025 that “73% of all jobs created [at the end of the Biden administration] were due to government and government-adjacent sectors [...] like education, sectors like health care [and] the short-term pain is coming from the reorientation of the economy, from the government to the private sector.”

Read that italicized quote again. 73% is too much of the job growth to come from government-adjacent sectors like private education and health services, and we'll have to take some pain to change the economy and bring that share down. Got it.

They did administer that short-term pain, and, during the first 10 months of their administration, at 58.8 thousand jobs a month out of 23.1 thousand in monthly overall private job growth, private education and health services were responsible for a comical 255% (!) of all job growth. But here we are past that and, after blowing up global trade, destroying historical alliances, raising tariffs to century-high levels, waging a terror campaign against non-citizens, shredding the federal administrative state, and attempting to imprison members of the central bank, they’ve done their “reorientation.”

And, as you can see in the figure above, at 50.2 out of 68.3 thousand jobs averaged over the past six months, private education and health services are responsible for … 73% of job growth. They warned about 73% of job growth coming from government and government-adjacent sectors, and here’s one of the sectors they flagged, doing exactly that on its own. They did all of this to end up in the exact same place, the exact same percentage point they started at, except inflation is now hotter. Ah! Well. Nevertheless.

So what’s driving this less-bad labor market over the past six months? Maybe it has something to do with tariffs.

Again, looking at the chart above, much of this turnaround is that blue-collar jobs aren’t bleeding out anymore. We’re losing fewer manufacturing jobs. Construction turned positive and so did employment services, which is likely a proxy for manufacturing employment via temporary workers (Bowdle and Tito, Federal Reserve, 2024). Together that shifts about 50,000 private sector jobs.

Maybe the worst of the tariff uncertainty is over? That doesn’t seem right; the courts made the endgame much more complicated during this time. Oral arguments for the case that blocked the initial tariffs were from November, with the consensus being that it went poorly for the Trump team. So maybe industry believed the worst was behind them?

But maybe it’s something to do with immigration?

Construction was a big turnaround, and construction tends to have a lot of foreign-born and non-citizen workers. Let’s go back one extra month, so data through March, so we can get more granular job categories. We take the 250 sub-industries that cover most of all private employment (the sub-industries that feed into the diffusion index), and break them into quantiles of their share of non-citizen employment in 2024. (See this post for the methodology.) So Q5 is the ~20% of sub-industries with the most non-citizens.

So for the highest quantile of non-citizen share, the first 10 months of 2025 saw a major drop, going from 29 thousand a month in 2024 to -19.8 thousand a month through October. But since November it’s at 33.3 thousand a month, a major rebound. A similar trend happened with the 4th quantile.

Are native-born workers taking these jobs? Probably not. Below is the prime-age (25-54) employment-to-population ratio by month (as it is not seasonally adjusted), and it’s obviously lower during this time than in prior years. We see the same thing if we use the native-born unemployment rate. I don’t think that’s the channel.

So what’s going on? Whatever it is, the labor market picture looks better, and the inflation rate worse, than in November of last year, when there was a case for potential rate cuts. That case is no more.

Let’s play a game. Below is a list of 13 inflation measures. I want you to guess two of them. Specifically I want you to guess the two inflation metrics Kevin Warsh said were his preferred measures during his Senate nomination hearing last week.

If you’ve followed inflation over the past 5 years you may be familiar with some of these. They have come up at some point to prove or disprove your favorite theory of why inflation has behaved the way it has. They all have links to their actual sources.

Now try to guess which two Warsh pointed to as being the best ones for right now. It’s ok if you don’t know much about inflation or economics, I think you can make an educated guess which are the two.

If you guessed the lowest two, the two “trimmed” ones, the ones that said inflation was the lowest under Donald Trump, congratulations. You got it right. Specifically, Warsh said the following during his testimony (my bold):

Senator Catherine Cortez Masto: So some Federal Reserve officials have said that this excess is due to tariffs. Do you agree with that?

Kevin Warsh: Senator, I don’t. […] the Fed needs to do, is to try to use our new understanding and new data sources to see what’s really the inflation rate in the economy. We used to use Core PCE, Core measures, so we’d exclude food and energy, because it was sort of a rough swag as to what was going on. We don’t have to do a rough swag anymore. What I’m most interested in is, what’s the underlying inflation rate? Not what’s the one time change in prices because of a change in geopolitics or a change in beef, but what’s the underlying generalized change in prices in the economy?

My broad sense is that these inflation risks and the inflation damage the last several years is improving somewhat. It has improved somewhat in the last year. The measures I prefer are looking at things that are called trimmed averages, where we take out all of the tail risks, all of the one-off items, and we ask ourselves whether the generalized change in prices is having second order effects on the economy. Again, they’re not where they should be, but I think that the trend is quite favorable.

I’ve tried to approach the Warsh nomination with an open mind. After all, a wide bipartisan range of people affiliated with Harvard University said he was a good choice.

I understand the liberal worry that Warsh is a sock puppet for President Trump’s ambitions to control the monetary and financial system. I worry about that too. There is something bleak yet familiar about watching someone so rich they can’t disclose all their wealth or plans for divesting thinking they’ll win out against Trump, that they’ll be the one who doesn’t get bitten by the snake they are supporting.

I also don’t take any comfort that the Federal Reserve board would contain him or Trump’s influence. As Bharat Ramamurti discussed, Fed chairs generally get their way. They control the staff, set the agenda, put the first formal policy proposal on the table, and set the tone at the press conferences.

But, and there’s no polite way to say this, I’m more worried that he’s just not up for this moment. This is going to be a very difficult time for the central bank, and nothing Warsh has said has given me much confidence in his abilities. He says a lot of stuff that no doubt works well in the circles he rolls in, but I don’t think it stands up well to anyone thinking seriously about the topics.

The way he discusses the Fed’s balance sheet, for instance, doesn’t convey that he’s thought hard about the tradeoffs. It’s even more apparent when he argues AI productivity would lower interest rates, seemingly unaware of the more obvious counterargument, like Krugman notes, that it would raise them. It all seems shallow.

Trimmed Mean Was a Disaster Measure

Warsh’s quote above continues this. I genuinely mean this in a non-partisan way, the worst lesson you could have taken from the last six years is that trimmed mean measures have some special signal other measures lack. During the inflation wave they consistently performed worse in both directions, leaving many smart researchers positioned the wrong way at the wrong time throughout the whole inflation saga.

Looking at the graphic above, you can see how much trimmed mean inflation trailed overall inflation during the initial inflation wave. Various trimmed and related measures made people underestimate in 2021 how high and widespread inflation would become in 2022, especially expanding to services. Trimmed measures moving up in 2022 also caused smart people to think a recession was necessary to bring down inflation. Instead, we had the soft landing, where GDP went up ~2.8% a year during the two years where inflation fell four percentage points, making traditional measures like the sacrifice ratio have the wrong sign.

I’m going to mention some names here, but I want to be clear I screwed this up more than most. But take Federal Reserve Chair Jay Powell himself. At Jackson Hole in August 2021, when inflation picked up but was not yet widespread, he pointed to “trimmed mean measures […] generally show inflation at or close to our 2 percent longer-run objective.” Yet six months later, the February 2022 Monetary Policy Report used the very same Dallas trimmed mean to argue that inflation was skyrocketing.

You can see an exaggerated version of this among academics. In a December 2021 IMF paper, Laurence Ball and co-authors held up the Cleveland weighted median and Dallas trimmed mean as the measures that had stayed smooth through the pandemic to imply that underlying inflation was still contained. But by the time of Ball’s September 2022 Brookings paper, built on the Cleveland weighted median (an extreme version of trimmed mean), Ball had flipped to the opposite side, arguing that the trimmed measures showed the Fed would need to push unemployment “far higher” to get inflation back to target, and either inflation would stay high or the country was headed for “a substantial economic slowdown.”

As Preston Mui of Employ America later noted, the Brookings event discussing the 2022 Ball paper was a venue for many to conclude that a higher trimmed mean inflation necessarily meant a recession was required to bring down overall inflation rates.1 Jason Furman wrote a Wall Street Journal piece on the paper, calling it “the scariest economics paper of 2022” and arguing an elevated trimmed measure meant “we may need to tolerate unemployment of 6.5% for two years” to bring inflation down. Thankfully the soft landing didn’t require any of it.

My own theory of the inflation wave is that there were several overlapping shocks, especially to the composition of spending, to working remotely, and key inputs like semiconductors as well as commodities following the war in Ukraine. That pulled up many prices, and because we live in a Keynesian world full of rigidities, other prices did not fall. Then relative prices were out of sync, so that pulled up other lagging prices, moving the distribution of items to line up. You can follow this logic with microfounded models as in Lorenzoni and Werning (2023), ironically the BPEA paper of the following year.

This is even more important as the divergence picks up again, as more is being excluded in these tails. The Dallas Fed did a deep dive into what their methodology finds in this moment. After replicating for a future post, I’m reading that in recent months it looks like service inflation is picking up, something we’ve been covering, but is being hidden by the trimming method. So we’re back to where we started: what exactly is going on at this moment?

These kinds of stories won't be understood through just looking at the central distribution of inflation. These overlapping shocks generate the misleading trimmed-mean signal that fooled analysts during the actual inflation wave. Figuring this out requires wisdom and judgment.

1

Rewatching the video of that 2022 Brookings Papers event, with all the economists calling for a major recession as soon as possible, there’s a funny moment where, as a result of multiple cameras, you can watch Claudia Sahm’s reaction in real time from her back row to Frederic Mishkin, of financial (in)stability in Iceland fame, saying that “bottom line is that the recession is probably going to be a serious recession” to bring inflation down.

Is it an epic eye-roll? I’ll leave you to judge. Claudia was one of the smartest people on this inflation wave, so subscribe to her Substack. Hi Claudia! 👋

I’ve written before about why I think people are concerned about affordability given high price growth in key items. I want to take a different angle. It’s a thought experiment, and it probably needs more analytical firepower than I can give it here, but I want to float it.

Making Comms People Sweat, Racial Wealth Gap Edition

First, let’s start with something you probably haven’t spent a lot of time fixated on: how would you describe the black-white racial wealth gap during the Biden administration?

Let’s take data from a Federal Reserve study, derived from the Survey of Consumer Finances, a high-quality survey of wealth administered every three years, last administered in 2022 (we don’t have 2025 yet). Here is black and white median wealth, inflation adjusted, for several years.1

One might say that “the racial wealth gap is the lowest it's been in 20 years.” This is the standard measurement of the ratio between the two, where 6.35 is the lowest since 2001. But reporters and others could ask, isn’t it the highest in decades? The absolute difference is the largest among all these measured here.

I would think a good response is: no, our measures of inequality are scale invariant. If you double incomes, the inequality measure should remain the same. This isn’t partisan. Check out the link for inequality measures on Wikipedia and there’s a whole “properties of inequality metrics” section that discusses scale independence. We talk about the 1% and the 99%, and if you double everyone’s income the 1% and the 99% still get the same shares. This is even more important since we compare inequality across decades, even a century, from the 1980s, 1950s, and 1920s. If we brought in an 18th century farmer to discuss absolute inequality measures, it wouldn’t answer any question that’s relevant.

(You can also imagine the look of shock and horror on the faces of a team of communication professionals when they realize you would really say this to a reporter, in fact you think this is not just a great and sufficient answer, but you can’t even wrap your head around an alternative argument.)

(And in case you are wondering on the ‘lowest in 20 years’ verdict, alas, Politifact: “We rate this claim Half True.”)

Economists are trained to think in ratios and how results change if they are multiplied by constants. Our inequality measures work this way. It’s also a stronger, key assumption built into macroeconomic theory, that if you double every price and every income then real variables shouldn’t budge. That is homogeneity of degree zero.

But everyday people have concerns, in the short term, over absolute differences. I want to be clear that, by talking about absolute levels, I’m not invoking money illusion. I want to at least approximate stories that could survive rational expectations. And I think absolutes can matter because people believe their room to absorb an unexpected cost scales with absolute income. I think there are three such vibecession stories that survive this.

Three Reasons In Play

It’ll be easier with some examples, and those are best based on actual numbers. Here, from the Economic Policy Institute, are real wages by percentile, from 2019 to 2024.

Wages are up following inflation. These are big wage increases for those at the bottom. But it still nets to a real absolute increase of under $2 an hour, indeed most are under that threshold though the 80th percentile has higher absolute increases than the 20th. If the new price environment is more volatile and complicated then absolute differences might matter.

Let’s discuss these three.

Essentials

Imagine a person near the bottom of the income distribution, who is spending as much or more than they make as a result of borrowing. This makes them very sensitive to specific costs, like food, electricity, and auto insurance, all of which are hard to cut.2 Even if essentials rose in proportion, the risks of having little should make people concerned about absolute prices. But these items rose much faster than overall prices, as shown in Figure 1 above. For those with low incomes, the squeeze here is particularly difficult, and if you view utility as spending above a certain baseline, that baseline increasing faster causes distress.

Housing

I tend to be a housing-theory-of-everything person, so much so that I’m excited to say I’ll have a big housing report out in the next month or so. And the housing situation is bad.

The median sale price of a newly built home bought with an FHA mortgage, a common financing path for first-time buyers, went from about $240,000 in 2019 to around $356,000 in late 2025, up roughly 48%. The 30-year mortgage rate went from around 4% in 2019 to over 6% now, which means the same house costs more every month even before the sticker price moves. Homeowners benefit from the wealth, but they also realize any new home they are going to buy has gone up in price too, and interest rates are higher. Fully-anticipated inflation can make home-buying more inaccessible (Modigliani and Lessard, 1975). But given the even higher costs here, it is easy to see how this causes angst.

Risks

I think the absolute level also matters for tail risk situations, where sudden, unexpected costs could impose a much higher penalty even adjusted for inflation. Let’s take health care. Kaiser Family Foundation’s Employer Health Benefits Survey reports that family coverage premiums in 2024 are up 24% since 2019, comparable to overall inflation. But the average isn’t what you worry about, it’s the tail. In any given year you might face a plan that dropped your doctor this cycle, a partial denial on a procedure, or a specialty referral that sends you outside the network. And now the premium tax credits have expired, while there’s also a trillion dollars of cuts to Medicaid on the horizon.

Essentials, housing, and tail risks all survive two key tests. First, they follow Binetti, Nuzzi, and Stantcheva (2024) in that “increased complexity and difficulty in household decision-making” are a central problem people cite for inflation in surveys. All of these fit that characterization.

Second, these were made worse under Trump, so they can justify the further decrease in sentiment and economic conditions in 2025. Tariffs made food prices worse, housing was hit by tariffs, interest rates, and likely deportations, and insurance is being hit by price increases from the end of the premium tax credits and future Medicaid cuts which will devastate rural and regional health care providers. This is a pretty bad situation for people. But these are solvable problems, in fact we can solve them.

The argument of the piece is that women’s employment rates are currently being “propped up” by government policies. Which policies? She names several that trace to the early 1990s. She explicitly names and dates the Civil Rights Act of 1991. She references student loans the government backs, which come out of a 1993 law, as well as immigration, which particularly takes off in the 1990s. She also names other events from the past decade, like DoorDash and the growth of HR departments, which she believes accelerated women’s inflated employment rates.

This is part of a broader argument on the right, drawing on Helen Andrews on the “feminization” of corporate culture and institutions, and the broader “longhouse” discourse. (As a response, I’d read Darby Saxbe at The Argument.) Stepman has a reference to the viral social media skits of the Australian skincare company The Breakout Hack, as is standard in these arguments. Andrews times her argument to the idea women took over fields in the late 2010s, and makes two empirical points: women have reached half of workers by taking over new institutions, and once “institutions reach a 50–50 split, they tend to blow past gender parity and become more and more female.”

The problem with this argument is that it doesn’t fit the key labor market economic data points. Women’s labor force participation plateaued in the mid-1990s and has been flat since. As a graphic from a Treasury report shows above, it is below that of peer countries. Before that there was no particular break where it accelerated; it followed a steady trend upward from the beginning of the 20th century, a familiar U-curve tied to capitalist development. There’s been no recent takeover of industries; the fields that already employed women simply grew faster. Specific industries tend to mean-revert their gender mix.

Let’s dig in.

Women’s Labor Force Participation Plateaued in the 1990s

The most basic problem with the “women are increasingly dominating the workplace” narrative is that they aren’t. As Figure 1 shows, prime-age women’s labor force participation rate, the share of women aged 25-54 active in the labor market by working or actively looking for work, rose at a consistent pace from the 1940s through the late 1990s and then stopped.1 It has been nearly flat for a quarter century, rising only about 1 percentage point in the very strong labor market of the past few years.

We can also see this in the employer data. The Current Employment Statistics, the employer survey behind the monthly jobs numbers, tracks total payroll employment and women’s employment at the industry level. Women’s share of total nonfarm jobs rose steadily from 32% in 1964 to 48% by 1995. Since then it has barely moved, reaching 50% in 2025. (We’ll talk about that 2 percentage point growth in a minute.) If the policies since the 1990s were supposed to turbocharge women’s entry into the labor force, there is no sign of it in any of the key data sources.

Maybe the aggregate numbers are flat because older workers mask what’s happening among younger cohorts, especially college-educated ones. We can check this directly with the CPS microdata.

I don’t see it. As Table 1 shows2, among college-educated 20-somethings, women’s labor force participation rate has been about 4 percentage points lower than men’s in 1992, 2019, and 2024 alike. Among non-college 20-somethings the gap narrowed, but only because men's participation fell.

What About the Kinds of Jobs Women Have?

Has the composition of jobs changed? Perhaps women aren’t entering at higher rates overall but are concentrating in specific sectors and pushing men out. The data argues otherwise.

To start, consider three major sectors:

For manufacturing, men lost more jobs in absolute terms as the sector collapsed. But women lost a larger share of their jobs, so women’s representation actually fell. Women’s share went from 32% to 29%.

Over 16 million jobs were created in private education and health services since 1990, more than any other supersector. Women gained 12.5 million of those jobs and men gained 3.8 million. But women’s share was 77% in 1990 and 77% in 2025. The sector grew enormously, but its gender composition didn’t change at all.

When it comes to professional and business services, presumably the sector where the elusive girlboss lives, jobs grew enormously, adding millions of jobs for both men and women. But women’s share of the sector was 47% in 1990 and 45% in 2025. It was essentially flat, even as both men and women gained a lot of jobs.

We can test this more granularly. Economists have a set of 250 private-sector sub-industries that they use to gauge the breadth of the labor market.3 These cover 100% of the private-sector economy. Of them, 209 give us the number of female employees. These cover 93% of the private-sector economy; the missing 41 are largely small and in manufacturing, which would likely amplify the results we’re about to see.

For each of these 209 industries, we can compare women’s share in 1990 to their share in 2025. The gender composition of industries is overwhelmingly determined by where they started. The correlation between an industry’s women’s share in 1990 and its share in 2025 is 0.955. The regression line sits just below the 45-degree line, which shows mean reversion, not acceleration. Industries that started most female-dominated saw their share decline slightly on average, while the most male-dominated saw slight increases. This is the opposite of a takeover.

Women’s share of total employment did tick up about 2 percentage points since the 1990s, reaching the 50 percent threshold Andrews flags. How? It could be because women are taking over more jobs across all industries, an increase within industries. Or it could be because sub-industries with more women are just growing faster than the rest, like health care. That would be an increase between industries, the economy shifting toward industries that already employed more women.

Economists call this a shift-share decomposition.4 We can use these sub-industries to ask why the aggregate women’s share of payroll employment rose over this period.

The within-industry component is -0.5 percentage points. That is, industries became slightly more male over this period. But the between-industry component is +2.3 percentage points. The entire aggregate increase, and then some, comes from the economy shifting toward industries that were already disproportionately female. Women did not move into new industries. Industries that already employed women simply grew faster. That's the story.

Ceilings and Floors

But aren’t men in crisis? As you can see from Table 1, non-college-educated men’s labor force participation has been declining for several decades. Men leaving the labor force have ended up in dark places. This is where many argue the resurgence of deaths from suicide, firearms, and overdoses shows up, dragging men’s life expectancy down.

Here’s how I think about this: we want to address the ceilings for women and the floors for men. What are the institutions that keep women’s labor force participation below that of women in other countries? And what is the floor for men who fall through in the wake of changes to capitalist development and technological change? How do we keep them from falling all the way through? We can do better than invoking girlbosses.

4

For those interested:

The main explanation is that net immigration collapsed. Job growth fell dramatically even as the unemployment rate remained roughly steady. Two new major papers from both the Dallas Fed and the Federal Reserve argue that break-even job growth may even be around zero now because of this decline in immigration.

These analyses generally work by projecting population growth, participation rates, and unemployment, then multiplying them together. They’re important studies and they’re difficult to get right because they depend on the unknowable future of population growth and the unknowable present of steady levels of unemployment and participation. But one thing I’ve seen less of in real-time is the cross-section of immigration and job growth. Immigrants aren’t scattered randomly across job categories. What can we see of their evolution in the monthly data?

This is a quick pass, designed to put some methods into play and spark debate. On this first glance, we can see the expected collapse in job growth for industries with high levels of non-citizen workers. But additional things jump out. The recent rebound in employment numbers over the past several months looks to be driven by this group of industries. Yet the wages for workers in these industries slowed among the fastest, the opposite of what conservatives were hoping would happen.

Methods

We bring in three things here1:

First, the diffusion index from the Current Employment Statistics (CES). This is a set of 250 sub-industries that together comprise nearly 100% of private-sector employment, giving us a broad cross-section of job growth. Since CES is a survey of businesses, we get good numbers on employment level and changes, as well as their own detailed industry code (NAICS). But as a result, we don’t know anything about the workers (except their gender).2

Second, citizenship status from the Census’s American Community Survey (ACS). The ACS collects citizenship status and provides its own NAICS code. We pull the 2024 1-Year ACS Public Use Microdata straight from the Census website. From that, for every available NAICS code, we calculate the percent of workers who are not citizens.

Third, we then match those NAICS codes with non-citizen percentages to the CES NAICS industries in a crosswalk algorithm that tries to match the industries with as much specificity as possible.3 If there’s an error in this analysis or if it is fundamentally misguided, it happens in this step. The match isn’t perfect but it works reasonably well. Of the 250 industries, 136 match at the 4-digit NAICS level, 98 at 3-digit, and 15 at 2-digit, for a total of 249 matches out of 250. The one notable issue is construction, which only shows up once in the ACS at the highest aggregation level. So the 10 construction sub-categories in the diffusion index for jobs data all get the same high non-citizen share of 19%.

The Census definition of non-citizen here broadly includes “temporary migrants (such as foreign students), humanitarian migrants (such as refugees and asylees), and unauthorized migrants.” A last methodological choice here is to keep all non-citizens, so it reflects those here legally and illegally. There is excellent work from the people at e.g. Pew that tries to disentangle legal versus unauthorized immigration levels. However, the Trump administration has revoked student visas and has targeted green card holders for deportation proceedings. For our exercise, the pressure campaign the administration has pushed on non-citizens here legally causes me to include the entire group.

The Slowdown

With that in hand, we can plot each industry’s job growth slowdown against its non-citizen share. For each of the 249 industries, I compute average monthly job growth in 2024 and 2025, then take 2025 subtracting 2024, and put it on the y-axis. Negative here means the industry slowed down. X-axis is the percent of noncitizen workers.

There’s a clear downward slope, and in particular at the right side of the distribution it’s basically all negative. The industries with the most non-citizen workers nearly all slowed down. A regression is significant but has little R-squared explanatory power.

You can also see the industry we teased in the first paragraph, the one with the largest negative value on the y-axis. That’s home health care services, which went from 13,000 a month in 2024 to 1,000 a month in 2025.

Home health aides do work we know needs to be done. It won’t be automated or done with AI. It is labor-intensive. There’s no sense that immigrants contributing to it detract from national security or keep us from upstream innovation. As this excellent overview piece last December by Shannon Najmabadi at the Washington Post showed, home health care is getting squeezed by immigration restrictions, cuts to Medicaid and Medicare, and the growth of Medicare Advantage. What are we even doing here?

Slowdown, Then Speedup

I take the 249 matched industries and divide them into five quantile groups by non-citizen share.

There’s a little lumpiness here at the top end, but I think we can proceed. Let’s make this clearer. What was the job growth in 2024 versus 2025 for those five categories?

It’s the fifth quintile, with the highest percent of non-citizen workers, that has the biggest drop, from contributing 29,000 jobs a month in 2024 to -6,000 in 2025. There is a slowdown across the board, but Q5 stands out. What everyone assumes is happening actually shows up in the data.

How closely can we see a break early in 2025? And since last November, there’s been some pickup in job numbers; what’s driving that? To look into this, here’s 3-month average job growth for the five quintiles from 2023 through February 2026.

You can see this Q5 drop happening right as the Trump administration takes office. Yet the recent pickup is also mostly driven by the fifth quintile, the industries with the most non-citizen workers. Let’s redo the comparison above but add a third bar for average monthly growth since last October.

There’s a big shift. Q5 jobs have surged since October while jobs in Q1 are actually falling. What do we think is happening here?

When you’re looking at data this granular, you have to be careful about over-interpreting what could be random noise in small categories. So I create a breadth measure for Q5: the percent of Q5 industries with positive 3-month average job growth.

It’s actually picked back up to old, broader, levels. So the recovery is fairly broad-based across Q5 industries, not just a few big movers.

What Isn’t Happening: Wages

There are several possible explanations, but let’s get one out of the way. I don’t think the Q5 pickup is explained by workers being pulled over from Q1 industries, which are declining. Q5 industries pay about $32/hour on average versus $43/hour in Q1. I also don’t think it’s workers being pulled into the labor force because we don’t see it in the native-born employment-to-population data, which is shrinking.

If employers in high-non-citizen industries were bidding up wages to attract replacement workers from other sectors, we should see wage acceleration in Q5. Here’s what we see, with both nonsupervisory and total wage growth in Graphic 6-7 and 8 respectively:

Q5 nonsupervisory wage growth decelerated by 0.69 percentage points. This is even more driven by Q5 for all-employee wage growth. This is true if we use fixed-weights to deal with compositional issues as well. Friend of the Substack Matthew Klein has also noticed a weird wage divergence among industries; perhaps this is part of the explanation.

Jobs decreasing and wages slowing at the same time is something we’d normally associate with a negative demand shock. But that doesn’t obviously characterize what’s happened with the step-up in immigration enforcement. To check if a negative demand shock is in play, we look at average weekly hours for nonsupervisory workers:

We see that hours are increasing for Q5. (They are flat or recently increasing if we look at all hours.) So not an obvious demand shock.

One story is that this is tariffs. Q4 and Q5 have a lot of workers in construction, durable goods, and transportation, all of which have had a major slowdown in job growth. Trying to estimate this versus cyclical measurements is difficult given the lack of actual aggregate slowdown in unemployment rates. But all of these workers are being impacted by the administration’s trade war.

Another more worrying possibility is that employers are using the threat of ICE to slow wages, even as hours increase, which would mean the enforcement regime is bad even for the remaining workers. We’ll get more granular data in the years ahead, but consider this post a signal flare for all students of monopsony power looking for research topics.

The industries with the most non-citizen workers had the biggest job growth collapse, they’re now driving the rebound, and their wages are decelerating. The first fact is what the break-even papers predict. The second and third are harder to square with any simple story about labor supply. Something is happening to employer behavior in these industries, and it’s not the high-wage workforce the Trump administration promised.

The Van Hollen and Cory Booker tax bills have set off a debate in Democratic politics. Both bills raise taxes on rich people roughly $1.5 trillion. They then, in turn, use those revenues to do a middle-class tax cut. According to the Tax Foundation, Van Hollen targets the $40,000 to $140,000 range for a tax cut; Booker reaches up to $280,000. There’s been significant pushback on the agenda among economic wonks. But another question remains: how did we get here, where some consider middle-class tax cuts an affirmative Democratic agenda?

There’s a clear short-term motivation for the policymakers who’ve put forth these proposals. As Groundwork’s Alex Jacquez helpfully explains in a recent American Prospect piece by David Dayen and Ryan Cooper, this push is largely driven by consultants chasing the perceived political benefit of Trump’s No Tax on Tips. I’m skeptical that those pledges drove Trump’s 2024 win, which has plenty of other sufficient explanations. And even if they did, the right response, in my personal opinion, is to fight on better and more consequential terrain like healthcare.

As for the longer story of how we got here, I don’t have the full answer, but I want to explore it through two moments, one in 2016 and one around 2008. Each is usefully anchored by a contemporaneous editorial that identified the problem as it was happening. In 2016, the Hillary Clinton campaign proposed to fund paid leave not with a payroll tax, not as an extension of Social Security, but with taxes on the wealthy. That decision foreclosed options for building universal social insurance. But it built off the earlier one, around 2008, when then-presidential candidate Barack Obama pledged not to raise taxes for those making below $250,000.

The Bush Trap

By cutting taxes for everyone in the early 2000s with an expiration date, George W. Bush effectively created a ratchet. When the cuts were set to expire at the end of 2010, Obama was boxed in twice over. He had campaigned on a pledge not to raise taxes on families under $250,000, so he could only let the top rates snap back. But the economy was fragile, with unemployment near 9.5 percent, and letting any taxes rise risked pulling demand out of an economy that couldn’t afford it. The result was the December 2010 deal: all the Bush cuts extended, top rates and all, in exchange for unemployment insurance and a payroll tax cut.

When the cuts came up again at the end of 2012, Obama framed the fight around middle-class tax cuts. Then-White House comms director Dan Pfeiffer argued “President Obama today will push for extension of middle class tax cuts. Will the GOP join him to provide certainty for 98% of Americans?”

He ended up striking a different deal, making the Bush rates permanent under $400,000 and raising taxes above that. That decision locked in a tax architecture. The vast majority of the Bush tax cuts were now bipartisan law.

To see criticism of this from when it happened, read Matt Yglesias’s 2009 American Prospect piece, “The Next Tax Revolt.” In it, Yglesias argues that Obama’s pledge didn’t change the anti-tax framework so much as find a way to survive within it, reinforcing the right’s claim that public services aren’t worth paying for. He concludes (italics in original):

Progressive taxation is an important principle. But the idea that further changes to the tax code should exclusively target the wealthy is ultimately counterproductive. Making the case may be difficult, but refusing to try to make it amounts to conceding defeat. At the end of the day, persuading people to support a more active role for government means persuading all of them that such a government is worth paying for.

Those concerns about broader taxation were kept at a distance, in part, because an era of falling interest rates, often called ‘secular stagnation,’ made debt burdens significantly less onerous. The long Great Recession, where the Federal Reserve’s interest rates remained at zero percent for seven years, from the end of 2008 to the end of 2015, made the immediate concern of getting more fiscal spending into the economy central. Meanwhile, growing awareness of inequality pushed policy attention toward the top of the distribution. Ideas like the Buffett Rule, a proposed minimum tax on incomes over $1 million, inspired by Warren Buffett’s observation that he paid a lower rate than his secretary, captured the mood.

Each campaign cycle deepened the tax threshold promise and boxed in policymakers more. Secretary Hillary Clinton championed the Buffett Rule in 2016, and Obama’s pledge was increased to $400,000 by Vice-President Biden while running in 2020. But let’s stick with 2016, and a specific decision flagged by the writer Bryce Covert at the time.

Paid Leave and Payroll Taxes

By 2016 the worst of the recession was behind us (unemployment was just under 5 percent), and there could have been space to discuss broader taxation without wrecking the economy. But the Hillary Clinton campaign kept the pledge and added an important choice: moving paid family leave financing from payroll taxes to taxes on the wealthy.

The writer Bryce Covert laid out the case against this in an important 2016 Slate piece, “Hillary Clinton’s Plan for Paid Family Leave Is Bad Policy Design“:

The proposal in Congress would [fund paid leave] by creating a social insurance program akin to Social Security. The bill would levy a new payroll tax—0.2 percent, or about $1.50 a week—on both employees and their employers to fill up the fund, and then the fund would pay out benefits to everyone eligible to receive them when they wanted to take family leave.

Clinton’s problem with this model is that, technically, backing it would bar her from making a promise she’s gotten very keen on making: that she won’t increase taxes, not by one cent, on people who make $250,000 a year or less. “Hillary strongly believes that middle-class families deserve a raise, not a tax increase,” her paid leave proposal states.

The deeper problem, as Covert emphasized, was political durability. Social Security and Medicare have become politically untouchable precisely because people pay in and feel ownership over the benefits.1 Payroll taxes had also been a highly effective model for state paid leave programs, so the Clinton approach meant abandoning a design that was already working.

Social insurance is insurance against lost wages. Payroll taxes are part of the mechanism through which workers contribute to insure themselves against unemployment, disability, sickness, or old age across generations. From the beginning of the study of social insurance, scholars knew that giving birth to children also disrupted earned wages, and thus paid leave belonged in the system.2 It was part of the unequal gendered nature of the New Deal that, as political scientist Suzanne Mettler has argued, the system divided citizens so that men got full benefits and women didn’t. Bringing paid leave into Social Security thus had long-term political benefits.

But none of this was enough to break the no-tax pledge. And during this period, the wonk class had largely stopped making the civic case for payroll taxes. The New Deal understanding that social insurance carries a deeper meaning, that “keep your government hands off my Medicare” reflects something real about how people relate to programs they pay into, that universal contributory programs can be the foundation of democratic solidarity, had faded from the conversation, left largely to heterodox thinkers like Michael Lind. Covert is a friend and I remember discussing this with her at the time. Though I then saw the case for moving the tax burden to the rich, I increasingly think she got this argument right as it occurred.

That is not to say there isn’t a straightforward case for higher taxes on the wealthy funding a broad set of programs and social insurance. You hear from pundits that Americans want Scandinavian social insurance without Scandinavian taxes. I actually don’t buy that’s the right point. The comparison is harder than it looks, and healthcare is the main reason. As both the economists Saez and Zucman in The Triumph of Injustice and Matt Bruenig at People's Policy Project have shown, adding in employers’ U.S. health spending as a tax dramatically increases the regressivity of taxes and largely closes the international gap. That international difference is fundamentally about the way we fund health care.

But as Democrats shape their policy agenda going forward, there is something worth exploring here: whether social insurance can be expanded through two tracks. The rich pay their fair share in our deeply unequal society. But also we affirm a broader commitment that everybody pays in and everybody benefits. That is what was underpinned Social Security in the first place, and it may be what’s needed to get out of the trap we’re in now.

Yet there remain two central ironies. For all the Democratic reluctance to ask anything of the broad middle class, it was Trump who imposed broad-based taxation on working-class individuals through tariffs. He just did it in the worst possible way, for the worst possible reasons, based on lies, and with the revenue going nowhere useful. More, for all that George W. Bush was thrown into the trash heap of history in 2009, here we are in 2026, unable to increase middle-class taxes above the path he set in 2001. And, after the Iraq War we still live with3, we are now waging another full-on war against Iran, another country he named to the Axis of Evil in his 2002 State of the Union. His strategery of playing the long game won, and we all lost.

Now subscribe and watch this drive:

Subscribe now

2

In his classic Social Insurance (1913), the first textbook on the topic published in the United States, the actuary I.M. Rubinow translated a 1912 German taxonomy of the types of social insurance, with my red text box flagging what we’d now call paid leave among all the important types:

It’s both refreshing to know the central risks of the capitalist modes of production were well articulated in this early 20C book I recommend you check out, and depressing to think the insufficient progress made on addressing them since then.

I don’t think the severity of this has been conveyed well in the media. Coverage tends to focus on CPI inflation, which is the majority of inputs into PCE, but it is not the Fed’s target and is being pulled lower by its heavier weight on housing. Important methodological debates have taken up airtime. And when PCE is discussed, it’s generally in year-over-year terms, which smooths over how much things have heated up over the past three to six months.1

All of this is before the war with Iran, which started on February 28th. The problem existed before the energy shock that is now ongoing. WTI crude is running roughly $30 above its pre-war level, and the hit to production and consumer spending would on its own put the central bank on pause. As Bob Elliott notes, an oil shock is “like the opposite of a productivity boom.” He finds that historically “the best case scenario is a pause to easing and a worst case scenario is hikes into weakness” in order to maintain expectations.2 The fact that inflation was already a problem before this shock only makes the Fed's position that much harder.

Overall

Figure 1 is the chart I follow most closely, the breakdown of PCE inflation by type across several measures. The top line is total PCE inflation over several dates and intervals, and the subsequent lines are how much each part contributes to that total number.

PCE inflation has increased. In 2024, it ran at 2.73%, and there was a clear runway for roughly 0.3 percentage points of decline in housing, which did materialize. That would have gotten inflation closer to target at around 2.4%, with the central bank doing a careful dance through the last mile of disinflation.

Instead, inflation picked up. It ran 2.91% across 2025. Over the past three and six months it ran 3.47% and 3.19%, respectively. It has been accelerating, getting hotter the more recently you look. And that includes banking the housing disinflation that came through on schedule.

What can make this situation better?

Tariffs and Goods

Core goods is a new and significant contributor to inflation, and some of that is tariffs. As a kind of tax on consumption, tariff-driven price increases should be looked through as long as inflation expectations aren’t breaking, and they haven’t. There’s significant debate about how much of the pickup in goods inflation is tariff-driven. The Trump administration is in the awkward position of arguing that core goods inflation is global and structural rather than temporary and driven by their tariff actions, which, if true, means the Fed should probably be tightening.

If we assume half of goods inflation is tariffs, total PCE inflation still sits at or above 3% on the three- and six-month horizon. If we assume all of goods inflation is tariffs, and it would be 0% otherwise, overall inflation is still higher than 2024. We’ve lost ground going into 2026.

We can also just look at non-housing services inflation. This measure was originally emphasized to deal with goods prices spiking on shortages, and here too we see a spike in recent months. It’s contributing a full 2.04 percentage points, the entire target, over the past six months. There’s a problem here.

Market-Based Measures

Another move has been to emphasize the imputed elements of core non-housing services, particularly portfolio management services. Stephen Miran of the FOMC emphasized looking at market-based core inflation in a December speech, arguing it showed genuine cooling. Recent months have gone against that analysis.

As Figures 2 and 3 show, market-based core inflation and market-based core excluding housing have accelerated in recent months:

Footnote on creation below.3 Here in Figure 4 are those two series, along with market-based core services ex housing, showing the annualized percent increase in the respective index. So market-based core inflation increased 2.86% over the past six months, compared to 2.5% in 2024.

Every one of these measures has accelerated in recent months. Over the past three, six, or twelve months, they all run higher than their 2024 values (annualized). Miran, in his speech, described 2002-2007 as a period in which overall inflation was at target, and thus a good benchmark for these market-based measures. They are all running much higher than that time period. (Hat-tip to Omair Sharif making this point here.) The various supercore measures, the ones specifically designed to strip out noise and find the underlying signal among imputed prices, are all heating up faster than we’d expect.

Labor Market

Another argument is that the labor market is softening enough to bring inflation down on its own. I think this is more complicated than it looks. The unemployment rate has been essentially flat for almost a year, ranging roughly between 4.3 and 4.4 percent since last May. Average hourly earnings have moved sideways. There is weakness in hiring and openings, and I think young college graduates show a surprisingly high level of unemployment given the headline rate. Personally, I tend to think there are two unemployment numbers, one for those under 26, which has increased, and one for those above, which is more normal.

But the case that the labor market is decelerating fast enough to take pressure off nominal spending growth isn’t there in the data right now.

Fiscal Impulse

The oil shock is coming. But before that, there was another shock already in the pipeline: the One Big Beautiful Bill Act (OBBBA). Below in Figure 5 is the fiscal impulse tracked by the Brookings Institution Hutchins Center Fiscal Impact Measure, which I’ve followed closely over the years. There’s a government shutdown which is shifting Q4 2025 spending to Q1 2026. But they also note an increase in GDP from “the stimulative effects of the OBBBA on both purchases and taxes.”

Going into the year, the fiscal impulse from OBBBA will be substantial in Q1, and the tariffs do not offset it. Ironically, a massive gas price spike may have arrived just in time to choke off the stimulative effects of OBBBA which was already pouring into a higher-inflation environment. But fiscal impulse is still in play as the Fed looks ahead through 2026, and this is before any additional war spending for Iran, which could get very expensive and become a large additional fiscal impulse if the conflict continues.

How Far Off Is the Fed?

Figure 6 gives us a sense on where the Federal Reserve’s own Taylor Rule would land with a couple of assumptions. Assuming, as the Fed does, that r* is 1% and NAIRU is 4.2%, then the year-over-year value of PCE would have the Fed raising rates. Assuming all core goods inflation would be zero without the tariffs in a year-over-year number has rates basically right where they currently are (yellow line).

This graphic gives a sense of what the Fed was thinking; it understood itself as too tight (the blue Fed funds rate is above measured inflation). However there hasn’t been continued disinflation, and in recent months inflation has picked back up. This graphic reflects year-over-year inflation, which is lower than the inflation rate over the past six or three months, which would have rates from a Taylor Rule even higher. The more the Fed looks at recent information, the more they should be worried they got ahead of things.

There are things the Trump administration could do to try and ease this. Pardon the entire FOMC and back off independence, pick selective tariffs to execute, and stop deportation quotas might be a baseline if they want to wage an ongoing war against Iran. But they seem unlikely to try and prioritize anything, much less their crises putting pressure on aggregate supply while pushing demand.

Maybe this all turns over. Though note that in recent years the second half has been the cooler part of the inflation calendar, even with seasonal adjustments, which makes it running hot especially worrying. The decisions the central bank made made sense as they were happening. But at this point the Fed should not be in any hurry to cut further. That was true even before a major energy war landed on top of everything else.

If you're at the Powell press conference today, all your colleagues are going to have questions about Iran and criminal charges against Powell, so you want something unique. But you likely are stressed hunting down oil and energy sources for stories and having had to watch your kids home from school for the D.C. storm that never came. So steal this blog post for your question! Some questions that'll get attention:

• The three- and six-month PCE numbers, both overall and supercore, are running well above target and have been accelerating even before the energy shock.

  • At what point do shorter-horizon measures like these change the Committee's assessment of where inflation is headed?

  • How much would this acceleration have to continue in order for the Committee to consider rate hikes?

  • Looking back, does the Committee believe the December cut was premature given how inflation has evolved since?

• Your own Taylor Rule, using the Committee’s r* and NAIRU estimates, suggests rates should be at or above where they are now, whether or not you look through tariffs, and especially above with more recent PCE inflation data. How do you make the case that the current stance isn’t accommodative, especially this far from the inflation target?

• Even zeroing out all goods inflation, non-housing services is contributing 2 percentage points on a six-month basis. Is the tariff look-through framework relevant when the non-tariff components are the ones accelerating?

• According to estimates, the OBBBA is delivering a meaningful fiscal impulse into an economy where inflation was already reaccelerating. How does the Committee factor that into its outlook?

It’s gotten pushback from economists and Fed watchers, especially those who question whether policymakers would really sit on their hands in the face of a disinflationary spiral. You can read a lot about it online. I’m going to use this moment instead as an opportunity to start digging into the macroeconomics of AI in a more disciplined way. This is material I’m learning in public here, please feel free to leave critical and technical comments.

My hypothesis is that AI lowering wages will make it harder for the Federal Reserve to stabilize the economy, and will create an ugly situation in terms of labor share and consumption (both central in Citrini’s story). When I embedded that argument in a leading DSGE framework, one of those predictions held up and the other didn’t.

The simplest, representative-agent New Keynesian (RANK) model doesn’t get us very far. We need income distribution to matter for Citrini’s story, where workers have high marginal propensities to consume (MPCs), capital owners have lower MPCs, and shifting income from one to the other affects demand. At the same time, I don’t want us to get lost in the pyrotechnics of a full heterogeneous-agent New Keynesian (HANK) model, where a continuum of agents generates their own wealth distributions.

A two-agent New Keynesian (TANK) model splits the difference. The model I’m going to use is from “Workers, capitalists, and the government: fiscal policy and income (re)distribution” (2021) by Cristiano Cantore and Lukas B. Freund. They are kind enough to include their Dynare .mod files, which makes replication and modification straightforward, a huge plus. Cantore and Freund describe their model as “tractable laborator[y] for understanding various macroeconomic experiments,” which is exactly what we need.

In this model there are two agents, a worker who works to earn labor market income and a capitalist who doesn’t work and receives profits. (I’m already sold.) Both are forward-looking rational agents. Workers can save in bonds but face portfolio adjustment costs. That friction makes them partially constrained and more responsive to current income than a permanent-income consumer, so they have a higher MPC. The rich capitalist class smooths their entire life’s income independently based on their permanent income.1

In order to explore this, I take their medium-scale model and do my best to insert a task-model loosely based on Acemoglu-Restrepo (2019) (especially read through these Jon Steinsson notes). We take a permanent innovation shock to the automation of tasks with three scenarios: one where the shock is fully labor-augmenting, benefitting workers, a second where it is fully labor replacing, and one in-between.

We end up with 43 endogenous equations that we send into MATLAB to be solved.2 (It’s ironic to model the collapse of software while relying on proprietary MATLAB.)

Here’s what we get:

In everyday language, red is where workers get screwed and blue is where workers fully benefit from AI. 0 is the steady-state, or what things were settled at prior to the shock. Don’t worry too much about units, we’re here to see the signs.

What do we see?

Interest rates: All three lines dip below zero, meaning the Fed cuts rates after the AI shock in each scenario. But the cut is dramatically larger under the blue scenario than the red one. This seems backwards. Shouldn’t we expect the scenario where the labor share falls to require deeper cuts?

What I think is happening is mechanical but unintuitive. In the labor-augmenting case, productivity rises faster than wages because of nominal rigidities. Unit labor costs fall sharply, generating deflationary pressure. The Taylor rule responds with aggressive cuts. In the displacement case, human marginal product falls along with wages. Deviations in marginal cost, which is what drives inflation in New Keynesian models, barely moves. So despite the collapse in labor share, there is little disinflation and therefore little pressure for rate cuts.

These kind of New Keynesian results, where inflation is largely about deviations of marginal costs, might feel like its own kind of science fiction. But it’s worth considering. It’s also funny to imagine the Federal Reserve staff who will have to explain to potential new bosses that AI doesn’t magically mean you get to do whatever you want to interest rates.

Workers: It’s not great. In the scenario where workers are augmented by AI they get more consumption and work fewer hours and the labor share fall isn’t that bad. In the displacement case, real wages barely rise, hours initially increase and then drift down, and labor share collapses. The modest rise in worker consumption comes primarily from higher labor supply rather than higher wages. They even, for some periods of the shock, work more hours than the baseline. Their labor share plummets, and their increased consumption is mostly from just working more hours.

I am still not sure if I’m doing justice to the AI shocks. I get a similar result when I instead just use a productivity shock to the model and make it so wages are fully inflexible, they are basically impossible to update. This extreme example gives us the same kinds of movements.3

From a standard business-cycle perspective, this remains within the reach of conventional macro stabilization. In the displacement scenario, inflation barely moves because marginal cost barely moves, so the Taylor rule doesn’t face dramatic tradeoffs.

That said, I’m still uneasy. If AI were sufficiently disinflationary, or if nominal rigidities behaved differently than in this calibration, we could find ourselves closer to the zero lower bound and facing a more prolonged demand shortfall than this baseline suggests.

I don’t like AI art. But one thing I always do to test the picture and video models is prompt it with “the world as if John Maynard Keynes’s ‘Economic Possibilities for Our Grandchildren’ came true.” In the two videos above you can see Sora, OpenAI’s video producer, from its public launch in October 2025 (top) to a video yesterday (second), each with that prompt, and chart the progress.

So I feel a visceral pain seeing a DSGE model that gives us periods where workers are compelled to work more hours as a result of AI-based skyrocketing productivity. My worry now is less that AI causes a recession. It’s that AI can raise output while worsening workers’ position, a distributional transformation without a traditional macro crisis.

That distributional question of who captures the AI dividend and how it feeds back into demand and political economy is where the real work lies. I’ll be digging into that more in the months ahead.

2

Code here. For the sickos, here’s the whole thing. 9-11, 13, 15, 33-34 are the lines where our additional task-based elements show up:

3

Funny enough here the Fed cuts less because output is so high because workers are working a lot more hours with their wages frozen. But as a quick check with a completely different methodology it gives us similar results.

Subscribe now

But much of the discussion has been caught between people building applications as fun personal hobbies on one end, and massive enterprise software on the other. Most of us who use these professionally will live somewhere in the middle. (The discourse is also wedged against whether AI will cause large-scale unemployment or otherwise destabilize society.)

I’ve integrated these terminal tools into my workflow over the past two months. Below are three specific ways I’m using them that are genuinely new, and where I’m not going back to how I worked before. Terminal AI compresses the setup and robustness-checking phase of knowledge work. I’ll also explain what makes them different from the browser-based chat tools, and whether I, Mike Konczal, am about to be automated out of a job. My relevant background is in a footnote here.1 This is not a setup guide, though I’ll try to explain ways these examples might be applicable to you. That will be helpful, as these tools are not going away.

Use 1: Real-Time Analysis, Without the Prep Work

I maintain a set of R files, the statistical programming language, for instant analysis of the economic data releases. Those numbers generally come out at 8:30 a.m. several times a month, and by 8:35 a.m. I need to know the big takeaways.

I’m not trading on this information, so I don’t need to know within milliseconds. But I do make statements that other political actors use, and accuracy is vital. For instance, I was interviewed by the New York Times shortly after the jobs numbers last week, and I want to be able to say something useful, interesting, and, most important, true.

My old workflow was that I’d take one or two hours the day before the major monthly data releases and manually code a few new graphics based on what I expected would be important. Maybe I post them or maybe I don’t, but they help me keep track of how the economy is moving, and they force me to think of new pieces I may have been missing.

Now I just ask Code/Codex to write them. It’s usually 95% of the way there on the first go. I check the math and get it fully there with one or two more exchanges.

What used to be 4–5 hours of work in an average month is now 15 minutes. The reduction in time allows me to ask even more questions and prepare more materials. I have a lot of experience in this, so I’m able to proofread the code and I double-check that it is doing what I want it to do.

Why not just use the Claude or ChatGPT webpage for this? I used to do this. But there is significant copy-and-paste overhead. The old way meant copying code out of the browser, pasting it, running it, copying any error messages or changes that needed to be made back into the chat, and repeating. With the terminal tool, the LLM writes the code and runs it in the same folder. It can see whether what it did worked, investigate failures on its own, and I see the result after every pass without touching a clipboard.

Why not just let the AI find the interesting result? The LLM actually does a poor job of this, as it lacks the context for what people are searching for or what stands out. It can summarize news reporting on those numbers after they circulate, but influencing that coverage is the point of the exercise.

So if your job involves monitoring something that changes on a schedule, such as earnings releases, polling averages, clinical trial updates, or monthly sales figures, and you currently spend the hours before and after each release manually refreshing Excel models or rerunning Tableau dashboards, this is the use case that will feel most immediately different. The tool doesn't just help you go faster. It makes it cheap to add new angles of analysis instead of being locked into what you could cover in limited prep time.

Use 2: Building Out a Report

Last Monday, I published a blog post about the affordability crisis. Check it out! Near the beginning are three graphics I made on the price, spending, and indexing of essential items. To do this I created a folder on my desktop and dropped in the relevant data files. Then I gave this opening prompt:

Note I ask it to create a qmd file (Quarto Markdown), which can render to html, pdf, or slides with minimal changes. You’ll probably be using a similar method in the future.

It got very close on the first pass. I scanned the code to check the methodology, spot-checked one item from each by manually calculating it (knowing the algorithm means that if one is correct, all of them are).

From there the prompts are mostly about cleaning up:

I’m just talking into the microphone at this point, deciding which essentials to include (some categories overlap in ways the LLM wouldn’t know) while simultaneously tweaking colors, legends, titles, and sizing, while it re-renders the graphic.

One thing I worry about is that if I’m not mucking around with the data myself, I’ll miss important side observations. But even in this pass I notice things: cigarettes show a huge price increase (it’s very funny that the LLM thought cigarettes are an essential), and I wonder if it’s driven by excise taxes, the same way tariffs are mechanically increasing the price of goods in 2025 and 2026. Something to investigate.

A benefit here is that stress-testing became trivial. Want to check whether your finding holds under a different index, or a different date range? If you are using Excel this is a nightmare, and even with statistical programming it’s a slog. But I get this done in a minute and can be more confident my results will hold under scrutiny.

Doing this part of my blog post last week would previously have taken 1-2 hours, with an additional 1-2 hours if I wanted to stress-test the indexes to make sure the results are robust. It took me 15 minutes to do this here.

So any future project that involves empirical work will be built out of a folder where the terminal LLM is executing work that I check. The terminal allows it to do much more complicated work, faster.

So if your job requires you to produce regular analytical reports, for clients, for leadership, or for publication, you will use this to cut out difficult first steps. Often the bottleneck is not the writing, it’s getting from raw, unorganized data to a defensible visual that you trust and can defend. The terminal tools compress that step dramatically, and more importantly, they make it cheap to check whether your finding survives different methodological choices. That robustness-checking step is the one most people skip under deadline pressure. Now you don’t have to.

You Bought a New Car and Auto Inflation’s Really Taken Off

I really want to emphasize how much of the work for any report is this basic setup, the data-wrangling and first-pass results to see if the results are worth exploring in full. This can be time-consuming, especially if it doesn’t pan out. Unless you try it, I can’t describe to you how fast the terminal LLM can get through this stage, and keeps the building blocks in place to iterate on.

Two examples I explored recently. There’s an interesting new paper from Louie, Mondragon, and Wieland on whether supply constraints actually explain house price and quantity growth across U.S. cities. I wanted to get into the guts of the paper before I formed an opinion and wrote about it in the future here. But the actual data, merged across multiple datasets of prices, quantities, and restrictions, all from different sources, each with its own formatting and naming conventions for local areas, is a nightmare.

I put the data in a folder with the paper and had the tool merge it all in a single 10-minute session. It closely reproduced their key graphic, giving me a foundation to explore their work more fully. (I was told by one of the authors that the research assistant who led that data merging had a lot of big feelings about this.) Data merging in the browser tool, by contrast, had never worked reliably for me; but being able to iterate in place, with the LLM reacting to exactly what went wrong and adjusting, got it done.

Or take the standard “Phillips curve” style analysis of inflation. From around 2022-2023, a lot of people were running these regressions adding variables that followed the timing of inflation, like global supply chain stress metrics or the vacancies-to-unemployment ratio. I was frustrated because those don’t predict inflation out-of-sample and basically anything with an up-and-down pattern during those years would “predict” inflation.

Back then I had some code that estimated inflation with those variables along with other things that just happened to pop in 2021-2022. First, a dummy variable with a value of 1 in 2021-2022 and 0 every other year. Second YouTube views for Olivia Rodrigo’s “good 4 u” video, which debuted in May 2021, just like the inflation spike. What would be the best predictor of inflation?

I never got around to finishing it all at the time. But then James Stock and Mark Watson, the godfathers of empirical macro, published a paper in late 2025 finding that basically only COVID deaths explain the inflation in a component analysis. It’s a big victory for the cost-push shock side of the inflation debate. We should do a full post on it, as the entire episode traces back to a single supply-side shock. That made me want to dig my idea back out.

Sure enough, Codex one-shotted it.

It got the data and did the analysis with one call.2 From the R-squared above, reflecting a few clean-up calls, Olivia Rodrigo does a better job estimating inflation than vacancies-to-unemployment. (The unit is log(good 4 u views + 1)). If you put them all together, only the dummy variable and the “good 4 u” variable are significant.

You will, no doubt, have more normal and productive projects you want to explore and see if they are worth developing fully. The real work is finding good questions and understanding how to make the results rigorous. Using the terminal gets you to where you can focus on what matters the most.

Use 3: Anticipating Arguments You Might Not Know Ahead of Time

Keeping with that blog post on essentials: I do not know the complicated microeconomic derivations as well as I should. What differentiates Hicksian from Marshallian demand does not come naturally to me. It’s been almost twenty years, and I still occasionally have nightmares about that one chart from Chapter 3 of Mas-Colell.

But my recent post on essentials (and a previous one on DoorDash) touches on how groceries and shelter behave differently from restaurants and delivery services as income changes. I wanted to understand how rigorous demand theory would assess what I’d written. So I asked the terminal to take a look:

AIDS here is the Almost Ideal Demand System, a standard model in microeconomics for estimating how consumer spending shifts across goods when prices and incomes change. (Angus Deaton and John Muellbauer published their groundbreaking work in 1980, roughly two years before they would have named it literally anything else.) It generated this analysis:

I have no idea if this analysis was executed correctly. But I do know that on a first pass, an approximation of the analysis doesn’t disprove my instinct that food and shelter behave like my post argues they do. My argument wasn’t conditional on this, and I don’t advertise it. But I now think I’m not going to take obvious incoming criticism on something I missed. And more importantly, this gives me an incentive to learn more and continue digging on this topic.

A lot of analytical work is anticipating good-faith skepticism. You know the argument you're making, but you're less sure what the specialist counterargument looks like. These tools let you run a first-pass stress test against these other frameworks, cheaply enough that you can do it before every major piece rather than only after you've already taken the incoming.

Is It Game Over?

All of this makes me better at my work. My software background helps, as I can proofread the code, double-check what it’s actually doing, and catch errors in the methodology. But I’m also wondering if I should be switching fields ASAP.

Who knows what the future holds? But I keep running a version of this test: I put the Survey of Consumer Finances data, a triennial survey by the Federal Reserve on household incomes, debts, and assets, in a folder, and ask the AI to find the most interesting results. It gives me simple summary statistics. I ask it to find both unique policy insights and business opportunities, and it gave me very generic ideas. I asked it to use linear regressions, figuring it could just p-hack something, and it gave me very little. I keep running a version of this, asking AI to find what’s interesting about a project independent of any direction from me, and it really can’t do it.

So this still looks like an extreme version of labor-saving technology. It makes people faster at a given set of tasks, and it lets you cover more ground with the same resources. Which effect dominates, fewer people or more output per person, is an open question. My experience, at least for now, is that it can complement people who know how to use it, but risks shortcutting those earlier in their careers before they’ve learned the building blocks.

Last summer I was so conflicted about giving up RStudio for Positron, a VS Code fork that integrates Python and Quarto more naturally into R development. Now I keep Positron open to check while I work in the terminal. Whatever happens with AI bubbles and buildouts, the economics of my own work have changed permanently. I can do in twenty minutes what used to take half a day. This means I ask more questions, check more assumptions, and cover more ground. I still believe the tools can’t identify what’s interesting or draw the right conclusions on their own. But they’ve made exploring what’s right a lot cheaper.

useR!: EconTwitter to the White House

I gave a short 15-minute talk last summer at the useR! conference for the R statistical programming language, titled “From #EconTwitter to the White House: Real Time Economic Data with R.” (A sentence I still can't quite believe is autobiographical.) The pitch is four reasons to use R, convincing analysts not to use Excel and economists not to use Stata.

There are a lot of fun tidbits from my experiences in and out of government working on data releases that you might enjoy.

Slides here. Even as more statistical analysis will be done through LLM terminal prompts rather than working directly in IDEs like RStudio, the reasons to base your code in R are still relevant. The ease of the tidyverse, the grammar of graphics framework, and the variety of data libraries available make it a great foundation to iterate with AI tools.

Yale Budget Lab

I was on a panel this past week at the Yale Budget Lab and the Tobin Center for Economic Policy titled “Tariffs, AI, and Our Economic Futures,” looking at the first year of the Trump second term. Moderated by business journalist Alexandra Scaggs, with Michael Faulkender (a Trump first-term Treasury official and Deputy Secretary of the Treasury, the number two, from January to August 2025) and Natasha Sarin (Budget Lab President, and former Biden Treasury official).

Around 32m55s, I pointedly bring up the IRS giving ICE access to ITIN tax records for undocumented workers, which Faulkender signed off on while at Treasury. There’s a back and forth among the panelists that is interesting if you follow the IRS/ICE story, one that ends with a chill coming over the room. (You’ll know when it happens. This was recorded before the Washington Post broke a story about related improper breaches.) There’s also a debate starting around 10m44s about why exactly Trump is doing tariffs the way he is, where I push back on the administration’s rationale.

Columbia / Groundwork Collaborative Webinar

I did a webinar with the Initiative for Policy Dialogue at Columbia and the Groundwork Collaborative examining the impacts of Trump’s economic agenda.

It was chaired by Joseph Stiglitz, with Groundwork’s Alex Jacquez, University of Michigan economist Sarah Miller, and Bharat Ramamurti. It’s a great overview. Even as someone following this closely, I’m still learning and amazed at the damage the Trump administration is doing to healthcare. Sarah Miller’s comments (starting 29m) on healthcare were informative.

Stiglitz & Konczal in the Boston Globe

Stiglitz and I keep our buddy cop routine going with an opinion piece in the Boston Globe covering much of what we discussed there: “The high cost of Trumponomics.” We take on Trump’s recent Wall Street Journal op-ed claiming his tariffs have brought about an economic miracle, and walk through why those claims are unmoored from the economy around us.

I was also quoted in the New York Times about recent data releases and the Washington Post on why affordability has become a campaign buzzword.

If you want me to appear or write for your next event, feel free to reach out!

Current Mood: Burn it down.

Current Music: St. Vincent’s excellent cover of Big Black’s “Kerosene.”

You’ve probably seen some version of Graphic 1 below, which shows real median household income from Census. After falling during the post-pandemic inflation surge, real incomes recovered. By 2024–2025, they had ended up higher than in 2019. This result holds across multiple ways of looking at this data, including hourly wage data. Yet consumer sentiment is stuck near historic lows, about as pessimistic as during the financial crisis and the depths of the Great Recession.

These two facts coexist. And the politics of “affordability” has rushed into the gap between them, as politicians and advocates try to speak to people’s persistent anxiety about prices and living costs.

But there has also been a subtle, yet real, pushback against this focus. That pushback usually starts from the observation that today’s affordability debate is inseparable from the 2021–2024 inflation episode. From there, two critiques follow.

The first is a money-illusion story: people fail to recognize that their incomes rose alongside prices, so their distress reflects confusion rather than material harm. This argument often emphasizes that incomes at the bottom of the distribution rose faster than those at the top, producing a durable wage compression. And yet polling consistently finds that lower-income households report more dissatisfaction with inflation, not less.

The second critique follows naturally. If what people really want is their old price level back, that is simply not something policymakers can deliver. Broad-based price declines tend to occur only in deep recessions, and even then only modestly. So it is said to be dangerous politics and bad economics to make affordability central. Doing so risks promising something impossible, or worse, flirting with economic collapse as a policy goal. Matthew Yglesias has made a version of this argument, affordability is “just high nominal wages” and “basically just anger at inflation,” as have others.

This conversation echoes what you often hear about the so-called “vibecession.” The term was originally coined by the writer Kyla Scanlon in 2022 in a nuanced way, to describe self-reinforcing pessimism. Today, it’s more often weaponized to imply that consumer sentiment is untethered from material reality, a reflection of the circulating bad vibes.

One could correctly say that these affordability problems existed in 2019 and are independent of the inflation wave. But we should take the bait. Over the past five years I’ve been haunted and utterly consumed by a keen observer of the debates over inflation. And I think it’s worth being explicit about why a wave of inflation can generate real affordability problems.

These mechanisms point toward specific policy interventions, many of which have already bubbled up in political campaigns. And they also help explain why President Trump’s current policy agenda is depressing sentiment, by placing pressure precisely on these channels. There are several, but the first is what matters the most.

1. The Essentials Squeeze

The simplest story is that essentials have been squeezed: their prices have risen faster than overall inflation, even as households are forced to devote more of their budgets to them.

Consumer prices have increased 26.1 percent over the past six years. But not all price increases are alike. Graphic 2 has some prices that have increased faster than overall prices during that time, ones that are pretty important for people. Food, shelter, transportation, hospitals, and veterinarian services all pop out. They’ve increased faster than both overall prices as well as core services (which rose 27.8 percent during this period).

Last week we discussed DoorDash. From Graphic 2 above, we can see that the price of food away from home (which includes delivery services like DoorDash) increased faster than groceries (food at home), and people shifted their spending to groceries. Given that spending on groceries tends to decline with income, we can understand this shift as a penalty people experience. Even if incomes stay the same, vibes (i.e. utility) decline.

Let’s create two definitions of essentials. Core essentials are groceries and shelter. Essentials are groceries, shelter, healthcare and transportation. Taking the Consumer Expenditure Survey (CEX) data, we can see in Graphic 3 that most households are devoting a larger share of their budgets to these essentials than they did before the pandemic, especially at lower incomes.

Is that because these goods are getting cheaper? No. Using PCE data, we can construct price indices for these bundles, weighted by their consumption shares.1 As Graphic 4 shows, both measures of essentials inflation have run well above overall PCE inflation.

The green line is overall PCE inflation, the blue line is an inflation index of groceries and shelter, and the red is the blue plus healthcare and transportation, all indexed by their consumption weight. As you can see, and as hinted from the CPI data in Graphic 2, the essentials are running much faster than overall inflation. This was true before as well. From end of 2013 to 2019, core essentials rose about 14.3%, essentials rose about 10.5%, and overall PCE rose about 7.9%. But this was supercharged in the recent period.

When prices rise and budget shares still increase, standard demand theory tells us these goods are necessities. If utility comes from consumption above some baseline floor, then a rising share devoted to essentials leaves less room for discretionary consumption and lowers welfare, even if total income keeps pace with total prices.

This also helps explain why lower-income workers are angry about inflation even as their wages rose faster than average. Households traded down to cheaper and generic food brands, a trend coined “cheapflation,” which drove up the prices for the basics that lower-income families depend on most. And as Catherine Rampell noted at the peak of inflation, CEX data show that lower-income households typically consume more than they earn and face more volatile work hours, both of which magnify the pain from price spikes.

The essentials squeeze alone is enough to provide microfoundations for the vibecession and affordability crisis. But it’s not the only channel. Let’s look at other ways inflation creates welfare losses that survive rational expectations. I’ll sketch them briefly.

2. The Housing Tilt Problem

This one is big in the older literature but is oddly absent from today’s debate.

As Modigliani and Lessard (1975) showed, standard fixed-payment mortgages interact badly with inflation even when inflation is perfectly anticipated.2 Higher inflation raises nominal mortgage rates, preserving real interest rates. But because mortgage payments are fixed in nominal terms, higher nominal rates mean much higher initial real payments that then fall rapidly over time. Real payments are front-loaded.

MacGee and Yao (2025) show, using modern life-cycle models, that when borrowing constraints bind at origination, this front-loading tightens credit for first-time buyers. People with steep expected income growth but limited current earnings must qualify against today’s higher nominal payment even though that payment will shrink quickly in real terms as wages rise.

If this is too complicated, it’s just another reason the housing market has been a mess.

3. Planning Under Uncertainty

Binetti, Nuzzi, and Stantcheva (2024) find from surveys that the most commonly cited consequence of inflation is not any specific price increase, but the complexity inflation introduces into everyday decision-making. Eighty-five percent of respondents identify this as a major effect, and more than a third rank it as the single most important one.

When prices are stable, budgeting is boring. When prices are volatile, households must constantly plan when to buy, what to delay, how much to save, and how to interpret nominal changes. That cognitive effort is costly.3

4. The Cost of Money Is Part of the Cost of Living

Another issue is that we largely exclude borrowing costs from how we talk about inflation.

As Bolhuis, Cramer, Schulz, and Summers (2024) show, incorporating borrowing costs into measures of consumer sentiment explains a lot of the U.S. sentiment gap. From a household perspective this is obvious. The interest payment on a new mortgage is several times higher. Interest on new car loans is up sharply. People, especially lower-income ones, experience this as part of the cost of living.

Official inflation measures abstract from interest costs for good monetary and economic reasons. But that creates a gap between measured inflation and lived affordability.

There are others, which we might discuss in the future, but this is a solid grounding.4

Conclusion

When the prices of necessities rise faster than everything else, when housing becomes mechanically harder to access, when planning gets more cognitively costly, and when borrowing is more expensive, welfare can fall even if average real incomes recover.

The good news is that these binding constraints are solvable problems. We have many ideas for tackling housing, healthcare, and food costs. Simple steps like not slapping century-high tariffs on foods like bananas on the fantasy that they create leverage, or not cutting a trillion dollars from Medicaid to cut Amazon’s corporate tax bill in half, would be a great start.

But the first step is to believe that what people have been screaming about their lives for the past several years actually exists. Even a representative agent, forward-looking and fully aware of all the parameters surrounding them, can feel the vibecession.

Nothing gets the internet going quite like a round of hand-wringing about whether too many people, especially young ones, are spending too much money on delivery services, distorting their sense of budgeting and of how the economy actually works.

However, DoorDash Discourse is mostly wrong.

People are not, in aggregate, spending more on eating out and delivery. They are spending more on groceries at home and less on food away from home. This is especially true of young people.

The last round was kicked off over the weekend with “Freedom With a Side of Guilt: How Food Delivery Is Reshaping Mealtime” from Priya Krishna at the New York Times:

In 2024, almost three of every four restaurant orders were not eaten in a restaurant, according to data from the National Restaurant Association. The number of households using delivery had roughly doubled from 2019, just before the pandemic, the group said. And in a survey last year, about one-third of American adults told the association that they ordered food for delivery at least once a week. […]

That disconnect extends to restaurants, many of which have accepted a trade-off: Delivery helped keep them afloat during the pandemic and expanded their customer base, but they now have fewer in-house diners.

This predictably produced a wave of takes about personal irresponsibility and moral failure. A representative tweet comes from Matt Yglesias: “Part of the affordability crisis is pretty clearly people just refusing to be thrifty — you should not be spending a quarter of your salary on DoorDash.”

I have no doubt that food delivery has picked up and normalized since the pandemic. The relevant economic question, though, is not whether delivery exists. It’s what it is displacing. Is delivery mostly replacing groceries cooked at home? Or is it mostly replacing eating in restaurants?

Positive Analysis

We have a place to look. The Bureau of Labor Statistics Consumer Expenditure Survey (BLS CEX) is a nationally representative annual survey of consumer units that reports annual spending by detailed category. It is the primary federal source for household expenditure patterns and is widely used to understand cost burdens, to inform CPI weights, and to benchmark how budgets shift over time. The CEX makes it possible to compare budget shares across time and across demographic groups such as income quintiles and age.

The latest data for 2024 came out in December 2025, delayed because of the government shutdown. This data can help explain why people’s experience of the economy has been poor even when aggregate income growth keeps pace with inflation. A rising share of budgets devoted to necessities will make households feel worse off.

Here is “food at home” and “food away from home”1 as a percent of total consumption2, for all families and by income quintile.3

The dotted-line and middle number are the average for 2015-2019. Here’s those numbers in a simplified chart:

What do we have here? Start with the post-2019 period. Overall, Americans are spending about the same share of their total budgets on food. For the bottom 40 percent of households, the food share is modestly higher.

But compositionally, there has been a clear shift. Less of the food budget is going to food away from home (which includes delivery and takeout). And much more of the food budget is going to groceries.

In other words, society as a whole is reallocating food spending toward cooking at home, not away from it. DoorDash Discourse gets the direction wrong.

Normative Analysis

Is this shift necessarily bad? And does it support affordability as a real political and economic concern?

I think yes, for two reasons.

First, the shift is bigger for the bottom 40 percent of incomes. Lower-income households are doing much more of the adjustment toward food at home than higher-income households. This is not just a uniform cultural pivot toward staring at your phone rather than going to a restaurant. It is a stratified adjustment likely driven by constraint.

Second, the longer-run trend matters. Historically, as incomes rise, the share of spending devoted to food at home falls. You can see that above, which is the percent of the budget spent versus real total expenditures (the same is true for nominal) for each quintiles of each year from 1980 to 2024. There is a very strong decline in food at home as people get wealthier.4

That doesn’t just mean people dislike cooking. It means that, with more resources, people tend to allocate more toward convenience and variety. A lot has changed in preferences and attitudes these past seven years, but the fact that more of the budget has shifted to a good that declines with income should give us a strong presumption of disutility. The affordability crisis is real.

What About the Kids?

Ok, but can we still blame the kids?

The same exact thing happens, with the pattern being most pronounced among people under 25. They experienced the largest shift toward food at home. The 1.7 percentage point shift toward cooking at home for those under 25 is 2.5 times larger than the 0.7 shift for the general population.

The kids are alright. They're just broke and cooking at home to make up for it.

I want to look at the labor market with fresh eyes as we go into 2026. And the first place I want to do this is the debate over weakening prospects for people in their 20s with college degrees. This was a highly debated topic in 2025, with a focus on whether this was consistent with an overall weakening labor market or whether it was a sign of the impact of AI in the workplace.

Subscribe now

My loosely-held hot take a year ago would have been, echoing Joey Politano making this argument at Big Think, that AI would increase jobs in the short-run. My view is AI makes productive workers far more productive, but its unreliability makes it difficult to automate positions that require certainty over how outcomes are determined. But entry-level workers might be vulnerable, both because of the way the career ladder is structured and because bosses, wanting to cut corners, would try to haphazardly replace workers with AI there first.

And we did see unemployment increase for new college graduates. But, as friend of the blog Guy Berger has noted, unemployment is also higher for young people without a college degree, consistent with a weakening labor market and low hiring. Overall unemployment is up, from 3.5% in the summer of 2023 to around 4.4% now, and we expect that impacts younger people more.

Is there a way to distinguish if it’s increasing more than we’d expect?

Regression Relationship Approach

Let’s build up to how we’re going to approach this. There are many unemployment rates, whether for race, education, gender, or age. But they all tend to track the headline unemployment rate you hear about in the news, called U-3 unemployment, even over decades.

Let’s take two groups: (1) men and (2) workers aged 45-54, and see if their rise in unemployment tracks what we’d expect given the overall increase. (Note that the data sourcing will be important throughout this: here the data is from the BLS itself, seasonally-adjusted and monthly.)

Let’s run a regression on their individual (log) unemployment rates versus the overall (log) unemployment rate for the years 1994 to 2019. We get an equation, which we then use to project forward to 2025. We take the actual increase in U-3 unemployment and, given the historical relationship, we ask: what would we expect the specific unemployment rate for those two subgroups to be?1

Figure 1.a shows the full history back to 1994. The dotted lines are projections based on the log-log relationship between each group’s unemployment rate and overall unemployment, trained on data through 2019. This is a pretty good fit across multiple business cycles.

Figure 1.b zooms in on the post-pandemic period, where we are projecting forward based on the 1994-2019 relationship. Both men and 45-54 year-olds track almost perfectly with their expected values. The labor market slowdown since 2023 has increased unemployment across the board, but these groups are rising exactly as much as their historical relationship with U-3 would predict. Nothing unusual here.

Multiple Groups

Let’s extend this across all demographic groups. I take seasonally adjusted monthly unemployment rates from the BLS by gender, race, age, and education, regress each against the overall unemployment rate 1994 through 2019, and calculate the average difference between actual and predicted values over the most recent 2 months:

A positive value means higher than expected from 4.4% unemployment. These results are consistent across many date ranges. Most groups cluster right around zero, with their unemployment rates exactly where you’d expect given the current level of overall unemployment. But young people (16-19 and 20-24) and BA+ stands out, with unemployment notably higher than its historical relationship would predict.

Many, including myself, have noted the statistically significant increase in black unemployment over the past year. Why doesn’t that flag above? Let’s look at this more closely:

Here we can see that black unemployment has sadly returned to its longer-term trend. The tight labor market and full employment environment of 2021-2024 especially benefited black workers, with the black unemployment rate falling much further than historical patterns would predict, and the black-white unemployment gap collapsed to record lows. It’s a terrible fact that this progress has reversed in this past year. As the labor market has cooled, black unemployment has unfortunately returned to its historical, and far too high, relationship with overall unemployment.

Young People

Now to focus on younger people. Let’s take data from the New York Fed’s The Labor Market for Recent College Graduates website. This longer-standing measure was given a lot of attention last year in an article from Derek Thompson. These are 22-27 year olds. Recent grads have a completed BA or higher; young workers are everyone else. The researchers make their own seasonal adjustment. This is their creation from the microdata. Let’s run the same regression from 1990 to 2019, their starting years, and project:

(EDIT: as requested by readers, added a new figure that shows the relationship back to 1990, and extends the recent figure back to 2017. Still same analysis.)

As you can see, people without college degrees who are 22-27 more or less track exactly what we’d expect given the labor market slowdown. But young college is much higher than our trendline we’d expect from the slowdown.

To be clear: college-educated unemployment is still lower than non-college unemployment, and everyone's unemployment is up. Let’s not get tripped up on this on social media. What's unusual is the gap between where college unemployment should be historically and where it actually is today.

Now, as John Burn-Murdoch and Guy Berger have each noted in different ways with excellent analysis, college-educated workers are likely to be new entrants to the labor force in their 20s compared to those entering the labor force earlier.2 So in a slowing labor market it might not be correct to compare these two age groups.

To dig into this, I take the BLS microdata from IPUMS and create my own categories. I use 12-month rolling values instead of seasonally adjusting the data. I take 3-year age bands (so, e.g., 22-24 for age 23) for college-plus and high-school-plus (HS+) without a BA. The College+ series starts at ages 21-23, and the HS+ series starts at ages 18-20. I then run a log-log regression on each from 1989 to 2019 and take the average difference between the projected and the actual over the past 3 months (ending in September, to avoid the month of missing data). I then chart it across the age spectrum:

To read this graphic, for someone who is 23 (the age-range 22-24) with a BA+, they have unemployment that is over 2 percentage points more than we’d expect based on their historical relationship with overall unemployment. That excess fades by their late 20s. This gap is also high for non-college workers in their early 20s, but their peak is lower than the college one and it collapses faster.3

So my current read is that young people have higher unemployment than we’d expect at 4.4% overall unemployment. It’s especially higher at its peak and throughout their 20s for people with a college degree. Their recent unemployment rate is historically a surprise. The bad kind of surprise.