This is an interview about AI. My fiancé works at Anthropic. See my full ethics disclosure here.

In the first three episodes of Platformer's miniseries on AI and jobs, we heard from tech leaders who are building and implementing the agents that may one day result in significant job loss. For the most part, though, they have steered away from alarmism. Box CEO Aaron Levie argued that the "last mile" of human labor will resist efforts to automate it. Google's James Manyika explained why it’s much easier to automate a task than a job. Boris Cherny, the creator of Claude Code, did say the end of software engineering as we know it is here — but predicted that would lead to the rise of a new "builder" profession that would see more jobs created than destroyed.

And so as we move into the second half of the series, we're shifting to include some perspectives outside tech. In particular, I wanted to talk with people who study the economy. I want to know what they're seeing in the data now, and to understand what will happen if our first three guests are soft-selling the disruption to come. If AI does start taking jobs in large numbers, what then?

For these and other questions, I turned this week to Kathryn Anne Edwards. Edwards is a labor economist and independent policy consultant who writes a column for Bloomberg Opinion and co-hosts a podcast of her own, Optimist Economy. (She is also a compelling presence on Instagram and other social platforms.)

Like our previous guests, Edwards thinks that much of the AI jobs panic is overblown, and she flatly rejects the Silicon Valley vision that AI will result in (to quote Sam Altman) a permanently unemployed "idle class." She told me that idea feels more than a little classist in how readily it writes off the ingenuity and resilience of the American workforce.

"I am very firmly of the belief that the US worker is an incredible being, and they don't deserve to be written off by their former employer as never being able to work again," Edwards told me. "You don't talk about American workers that way in front of me, and you don't really have a place to."

At the same time, Edwards argues that the United States is woefully unprepared for mass unemployment — and that should be reason enough to start planning for AI job loss scenarios. We already have plenty of people out of work whom we've simply chosen not to help, Edwards reminded me.

We already know what tools would work, she argues: overhauling our unemployment insurance and healthcare systems; government subsidies to help out-of-work people move in search of new opportunities; and raising the estate tax, among other things. Whether any of it happens is a question of political will.

Which is why, despite her worries, Edwards insists she's an optimist — albeit a cynical one. "The lowest bar is the greatest source of optimism," she told me, "because I'd feel very differently if we had tried anything."

Highlights of our conversation are below, edited for clarity and length. Listen to the entire conversation wherever you get your podcasts — just search for Platformer — or watch it on YouTube at youtube.com/caseynewton.

And let us know what you think — we're new to podcast production, and welcome your feedback at casey@platformer.news.

Casey Newton: Your podcast is called Optimist Economy, and its central question is what it would take for the US to have nice things. But your recent columns have been pretty candid about how broken our safety net is. So what is making you an optimist these days?

Kathryn Anne Edwards: It's a very practical optimism — not being a naive Pollyanna, but having such a good handle on the economic evidence of how good policies can be when we do in fact try to make them good, as opposed to leaving them to rot for decades.

Newton: On our first episode, Box CEO Aaron Levie told me jobs aren't going anywhere — that AI has the net effect of giving engineers more to do. Last week, Boris Cherny, who built Claude Code, told me jobs might transform so they become unrecognizable, but there will still be a good job in there somewhere. Some call this the "AI as normal technology" view — that there are laws of gravity that slow the diffusion of new tech. Is that your view, that this is not a case of "it's different this time"?

Edwards: I'm a labor economist, so we all have the prison of our own point of view. Technology in the workplace changes the nature of labor, but throughout the history of technological adoption, it is very hard to see in real time either the addition or deletion of workers, and even after the fact it's very hard to attribute job loss to technology. AI comes out with a splash, but it takes a while for firms to really adjust. It will cause some people to get a new job, some a job change, and some to lose their job — and knowing any of that in detail will be very hard, if not impossible.

And of course it's different. Every technology is different. But I think some of the conversation around AI is deeply, deeply classist ….

It feels like some of that comes from being a little high on your own supply of how amazing your technology is when you hawk it. And it comes from this idea that some workers are better or more special than others, which makes your technology better or more special. That's also a way of saying the value of other workers is a lot lower — so it doesn't matter if they're displaced, it only matters if these workers are displaced. I push back on that, because I don't like being quite so derogatory to so many people in our economy just because they didn't have a knowledge-worker office job.

And two: I'd be willing to accept the exception that it's really, really special when it does something really, really special. I don't mean that in an insulting way. But what have we tangibly seen in our economy as a result of these LLMs being accessible through a subscription-based chatbot?

Newton: That's the Solow paradox — the economist Robert Solow said in the '80s that you can see the computer age “everywhere but in the productivity statistics.” A lot of economists feel the same about AI: it's everywhere, and yet it's not obvious it's having a massive, measurable effect.

Edwards: You went highbrow; I'll go lowbrow. It's like teenage sex: everyone's talking about it, no one's doing it.

I find this question itself a little distracting. What do you need to know, and why? If your concern is that there's going to be job loss, and therefore we need to have a better system for handling what could be cascading layoffs — do you need to know the exact number of people who lost their job, or is the problem the same regardless? Will you act differently if it's 7 million people versus 7.5 million? 

If yes, then I really question whether you actually care about helping people affected by job loss. And do you care about all the people who currently don't have a job, who in many respects outnumber most of the projections of how many jobs will be lost to AI?

Newton: We've heard companies from Amazon to Salesforce cite AI as a reason behind layoffs. Are there elements of those layoffs you think can properly be attributed to AI? Or are people AI-washing layoffs, and papering over the over-hiring they did during the pandemic?

Edwards: We don't know, we might never know, and honestly the company itself might not know. Rarely do actions like that have a single determinant. It could be that you're optimistic about the effect AI will have and could go forward with a leaner staff — and also that you over-hired in the pandemic, and also that the economy has been plagued with uncertainty, and it would just be nice to be leaner anyway, so you have a little less risk through a volatile economic period. All of those can be true. And then in the press release you say, "We've adopted AI," because that has the largest [reward] in the stock market for your shareholders.

What economists will do is look at specific occupations and industries and look for changes apart from the overall trend, or dramatically different from their past trend. Saying "lots of young people are having a hard time finding a job right now, and therefore AI" — I don't think we'd make that claim. “At the same time” is not the same as cause and effect.

And tech, even 15 years ago, is an industry that eats its young. A new technology comes along and they prefer to lay people off and hire young people who know the new programming language, rather than retraining older people. They tend not to employ people as much over the age of 50. So looking for job loss within industries that have long-term high turnover anyway — it's tough.

Newton: Let's zoom out a little bit. In May, the US Bureau of Labor Statistics reported that employment in 18 occupations that are considered exposed to AI declined by 0.2% while employment rose 0.8% everywhere else. Is that the kind of statistic that makes you think, okay, now we're starting to understand something that is really happening? Or is it just another confounding variable?

Edwards: It's illustrative, but not definitive. You have to ask yourself: is there another reason why an occupation exposed to AI could be experiencing slower-than-average job growth? Well, if the funding model of those companies is reliant on investment, and interest rates might be poised to rise again, it could be that they're wary about taking on more headcount in an area where they're not profitable to begin with, right? So it's not necessarily ruling out that explanation that AI is causing job loss, so much as leaving in everything that can't be eliminated.

But I genuinely believe AI is already causing job loss, just like I genuinely believe AI is already causing job growth. People want economists to say, "What's the right number? Can we pinpoint it?" And we're kind of pushing back: what do you need to know in order to make a conclusion, and what do you want that conclusion to serve?

Newton: One conclusion I think people are looking for is basically just: is the rate of AI job creation higher than the rate of AI job loss? Or what is the relationship between those two things? 

Edwards: It's going to be hard to know, ever. 

Go back and look at a big technology — the internet, Microsoft Office, computers. You're not going to find many reliable estimates of the total job loss and job creation from computers entering the US workforce. Someone has come up with a number, but the number would be one paragraph and the caveats would be 10 pages. 

The people who have been loudest about AI creating some permanent idle class — in my assessment, that's all coming from AI founders themselves.

Newton: There's a view in Silicon Valley, most loudly espoused by the AI CEOs, that eventually the models will get good enough that people will just hire less. My sense is you're skeptical that we reach that point.

Edwards: So you are able to run a shop with fewer people, because those people can use AI. That will happen, absolutely. It's the next step that makes their claim look ridiculous: they say technology shops will need two workers instead of 10, or 20 instead of 200, and therefore we'll have tens of millions of permanently unemployed people. It's the conclusion they draw that you have to separate out.

The first thing will happen — that's what happened to manufacturing workers. Think of people making shoes in 1905, 2005, 2105. How many are going to be in the factory in 80 years? Not many. How many were there in 1905? A lot. We know that happens with technology. It's the conclusion — "and therefore we have a new idle class" — that I take issue with. That one is borderline absurd. I could see how it could happen, but it's nothing like anything that has happened before, and a lot has to go wrong for it to be the case.

Newton: Let's sketch it out. Say we hit the point where a firm that used to hire 200 people can now get away with 20. We now have 180 people who would have been employed there who aren't. What do we expect to happen, given the normal way the economy evolves?

Edwards: Those people will look for new jobs, and the vast majority will find them — the average unemployment spell is less than three months long. There's a smaller, fractional set who are permanently long-term unemployed: however hard they've tried, they're not getting a job, and something has to change — the job they're looking for, or the resume they have. That share could get larger, and it has in the past.

But the jobs exposed to AI don't employ enough people for the dystopian scenario to play out. We have 170 million workers, and they don't all work in AI and tech. So even if you have a tech recession with lots of layoffs, it doesn't mean those people are unemployable forever, and it doesn't mean the economy is fundamentally changed, because we have lots of industry-specific recessions that we then recover from. 

It's also about overall aggregate demand, and whether people are trying to buy things. It doesn't happen in isolation.

Newton: The pandemic was close to a rapture for people in leisure and hospitality, and people in the AI world talk about an AI rapture in similar terms. What actually happened for the workers whose jobs disappeared overnight that let the economy reabsorb them?

Edwards: Demand increased again. What happened in the pandemic wasn't a drop in aggregate demand, it was a free fall, because so many businesses shut down. Nothing will ever top that job loss — half of leisure and hospitality lost its job within three weeks. Those 22 and a half million workers were met with a broken unemployment insurance system that Congress tried to fix on the fly, plus stimulus checks, forbearance on loans, protection from eviction. The economy's size recovered relatively quickly; the labor market took three years, as people went out and started to spend again.

I don't think AI could ever do anything as dramatic as March of 2020, in terms of a single month's job loss, and yet we recovered from it. It took public policy that we knew would work, fitted to the situation. It wasn't about the type of job that was lost. It was that people didn't have income. That's what you respond to: the person, not the former employer.

Newton: Let me come at it from another angle: entry-level workers. The Dallas Fed found employment in the most AI-exposed sectors is down about 1% since late 2022, with workers under 25 hit hardest — not through layoffs, but because entry-level jobs simply aren't being created. What do you make of AI's impact being felt through jobs that never appear?

Edwards: It makes a lot more intuitive sense. I start to use a chatbot, my workers are more productive, so I don't have to expand headcount to expand my business, so I just don't hire. But it's not as if the economy is otherwise perfect. There's a lot going on in our labor market that's much more influential on the youth labor market than AI.

In the spring of 2022, the Federal Reserve started raising interest rates. The labor market peaked about three months later and has been falling since. It was at such an extraordinarily strong place that the fall hasn't been enough to trigger a recession, but it looks like one — a drop-off in hiring, in wage growth, in job openings, in quits. The labor market is slowing down.

When that happens, certain things follow. One is upskilling. When you're an employer with a lot of people applying, you can require more. We saw this in the Great Recession — same job, same title, and suddenly it needed a master's degree, suddenly three years of experience. An entry-level job isn't just lost to AI; it can be lost in a labor market where employers have the upper hand, where so many people are looking for work that they can demand more skills for the same job. The opposite happens when the labor market gets tight — that's downskilling, where the job might not even require a college degree anymore.

The youth unemployment rate is by no means at a record high. But the labor market has started to slow, so we'd expect upskilling. And young workers — I hate to say this so bluntly — because they don't know how to