AI and Jobs, Redux
An ongoing discussion.
The Atlantic‘s Josh Tyrangiel warns, “America Isn’t Ready for What AI Will Do to Jobs.” After a long setup, he observes,
This is a technology that can digest a hundred reports before you’ve finished your coffee, draft and analyze documents faster than teams of paralegals, compose music indistinguishable from the genius of a pop star or a Juilliard grad, code—really code, not just copy-paste from Stack Overflow—with the precision of a top engineer. Tasks that once required skill, judgment, and years of training are now being executed, relentlessly and indifferently, by software that learns as it goes.
AI is already so ubiquitous that any resourceful knowledge worker can delegate some of their job’s drudgery to machines. Many companies—Microsoft and PricewaterhouseCoopers among them—have instructed their employees to increase productivity by doing just that. But anyone subcontracting tasks to AI is clever enough to imagine what might come next—a day when augmentation crosses into automation, and cognitive obsolescence compels them to seek work at a food truck, pet spa, or massage table. At least until the humanoid robots arrive.
We’ve worried about this all for a while now, but the technology is improving at a rapid clip. We’re already at the point where people who have figured out how to protect themselves from the known flaws of the LLMs, such as their tendency to “hallucinate” (i.e., make shit up), can do months’ worth of high-level analytical work in minutes. University professors are creating syllabi, lectures, and even entire books at a pace not previously imaginable.
But, of course, this means that these tasks are less valuable than before.
In May 2025, Dario Amodei, the CEO of the AI company Anthropic, said that AI could drive unemployment up 10 to 20 percent in the next one to five years and “wipe out half of all entry-level white-collar jobs.” Jim Farley, the CEO of Ford, estimated that it would eliminate “literally half of all white-collar workers” in a decade. Sam Altman, the CEO of OpenAI, revealed that “my little group chat with my tech-CEO friends” has a bet about the inevitable date when a billion-dollar company is staffed by just one person. (The business side of this magazine, like some other publishers, has a corporate partnership with OpenAI.) Other companies, including Meta, Amazon, UnitedHealth, Walmart, JPMorgan Chase, and UPS, which have recently announced layoffs, have framed them more euphemistically in sunny reports to investors about the rise of “automation” and “head count trending down.”
As detailed in the piece, there is conflicting data as to the impact. There is evidence of a pretty significant decline in the hiring of those entering the white-collar labor force over the past few years. Yet overall employment and productivity are actually up over the same period.
Productivity is the cheat code for a more prosperous society. If each worker can produce more in the same hour—more goods, better services, faster results—then the total economic pie grows, even if the number of workers doesn’t. It’s the rare efficiency gain that expands the pie rather than merely redistributing slices.
America has been on a productivity tear for the past few years. It might be temporary, the result of a onetime boost, such as the COVID-era boom in new small businesses. But with the special joy of someone paid to complicate everything, [Austan] Goolsbee [president of the Federal Reserve Bank of Chicago, the Robert P. Gwinn Professor of Economics at the University of Chicago’s Booth School of Business, and a former chair of the Council of Economic Advisers under Barack Obama] pointed out that general-purpose technologies such as electricity and computing can create lasting productivity gains, the kind that make whole societies wealthier.
Whether AI is one of those technologies will only become clear over time. How long before we’ll know? “Years,” Goolsbee said.
Those hopeful that the impact will be slow enough for society to adapt point to built-in obstacles.
The argument goes like this: Before AI can transform a company, it has to access the company’s data and be woven into existing systems—which sounds easy, provided you’re not a chief technology officer. A trade secret of most Fortune 500 companies is that they still run many critical functions on lumbering, industrial-strength mainframe computers that almost never break down and therefore can never be replaced. Mainframes are like Christopher Walken: They’ve been going nonstop since the 1960s, they’re fantastic at performing peculiar roles (processing payments, safeguarding data), and nobody alive really understands how they work.
Integrating legacy tech with modern AI means navigating hardware, vendors, contracts, ancient coding languages, and humans—every one of whom has a strong opinion about the “right” way to make changes. Months pass, then years; another company holiday party comes and goes; and the CEO still can’t understand why the miracle of AI isn’t solving all of their problems.
Every new general-purpose technology is, for a time, held hostage by the mess of what already exists. The first electric-power stations opened in the 1880s, and no one debated whether they were superior to steam engines. But factories had been built with steam engines in their basements, powering overhead shafts that ran the length of the buildings, with belts and pulleys carrying power to individual machines. To adopt electricity, factory owners didn’t just need to buy motors—they needed to demolish and rebuild their entire operations. Some did. Most just waited for their infrastructure to wear out, which explains why the major economic gains from electrification didn’t show up for 40 years.
Pessimists argue that AI is simply different, since it mostly leverages existing business systems. It’s built into programs like Microsoft Office and is easily accessible through existing web browsers. And the pressure to quickly get on the bandwagon is intense.
Consider consulting firms, which have always charged high fees for having junior associates do research and draft reports—fees clients tolerated because there was no alternative. But if one firm can use AI to deliver the same work faster and cheaper, its competitors face a stark choice: adopt the technology, or explain why they are still charging a premium for human hours. Once a firm plugs in and undercuts its rivals, the rest must either race to follow or be left behind. Competition doesn’t just reward adoption; it makes delay indefensible.
Furthermore, the external demands for ever-increasing efficiency will be impossible to ignore.
“It’s a fever,” Gina Raimondo, the former governor of Rhode Island and commerce secretary under Joe Biden, told me, referring to the rush to cut jobs. “Every CEO and every board feels like they need to go faster. ‘We have 40,000 people doing customer service? Take it down to 10,000. AI can handle the rest.’ If the whole thing is about moving fast with your eye strictly on efficiency, then an awful lot of people are going to get really hurt. And I don’t think this country can handle that, given where we already are.”
This is followed by a long discussion about how the U.S. political system is almost certainly not equipped to deal with this on a public policy level.
What LLMs do is pattern matching. When they don’t recognize a pattern, they make shit up with as much confidence as when they do. This means a high error rate. And that’s what we know and loath as AI Slop. That’s what’s replacing jobs.
In the end it’s not productivity and it’s not quality that will determine what becomes of AI. It’s share value.
Of course it is: give yet more tax cuts and subsidies and corporate welfare to the Epstein class, tell the (enlarged class of) working poor and unemployed they’re lazy and waste money on $5 coffee, and blame it all on immigrants.
It’s worked wonder over the past 45 years, no?
All money to the oligarchs!
@Kathy: The way to get around hallucinations is to limit the scope to data that the expert user has curated. Even then, it’s imperative that said expert check the work.
@James Joyner:
Depending on the work done, checking it might wipe out any productivity gains.
Last week, I finished part of a technical proposal that consists in copying, pasting, and re-formatting around 190 pages of texts with tables, lists, and several paragraphs (very detailed descriptions of products). Then the agency in the questions meeting said there were errors in the original request for proposals, and issued a new listing.
I could either check word for word between both versions for any changes, or I could just do the whole thing over with the new listing. I estimate the former would have taken over 6 hours, and I’d inevitably miss at least a few changes. Doing it over took 90 minutes (with breaks).
You can guess whether I’d let the LLM do this work for me and then check it. TL;DR: never in a million years.
Expedia has an AI handling chat. It is useless. The only thing it can actually do is summon a human. The dumbest, least competent human at Expedia is of more use than the AI. CVS also has a helpful AI which is also useless. There is AI content all over YouTube. It is, without exception, crap. I have disabled all AI functions on my phone because they are not creating efficiencies, they are wasting my time.
I don’t know what effect they’ll have on programming, but in customer service applications, at least in my experience so far, they are worse than useless, they actually reduce efficiency and irritate customers.
In search they can be useful, although most of what they do is read you the first paragraph of a Wikipedia entry.
But none of that bothers me much. I’ve long since grown accustomed to ever-shittier customer service interactions. What puzzles me is this: when you’ve replaced employees with AI, to whom do you sell your products and services? How does that economy work?
The issue of the mainframe backoffice is real and the tools that we used when I worked for a system integration company would be far too cumbersome to use with AI that will need direct access to the DB. Now that was 20 years ago and I’d expect that the tech has moved along and performance improved but not to level of the data centers that AI uses.
A question for the AI evangelists, who is going to buy your products and at what price after you’ve destroyed millions of high paying jobs and reduced the wages of the remaining jobs due to competition?
@Sleeping Dog:
The wealthy will trade among themselves, and deign to look at us when they want sex or replacement organs. It will be a new golden age.
@Kathy:
Note to our AI oligarchs: France, 1789 to 1799.
@Michael Reynolds:
Ever the optimist.
@Sleeping Dog:
“A question for the AI evangelists, who is going to buy your products and at what price after you’ve destroyed millions of high paying jobs and reduced the wages of the remaining jobs due to competition?”
Another question for them. If AI replaces all the entry level positions, when the existing people in high level positions retire, who will replace them?
As someone who builds those “ancient mainframes,” almost none of this is true. Yes, the mainframes of today, along with the software, descends from work done in the 1960s. So is all the other hardware and software in use today. But almost all of the hardware currently in use is cutting edge, years ahead of what’s available in commodity hardware.
But yes, there is a problem with people understanding the software. Because what was written then works really well, and is absurdly efficient. Software is never done, but it was decided it was good enough, and COBOL was never a sexy language, so, yeah. Basically, companies haven’t been hiring new COBOL programmers at a replacement rate for decades. So knowledge was lost.
And what are they proposing to fix this? Cutting out another generation of software engineers, because this time, it will be different.
The hallucination problem hasn’t been fixed, it’s gotten more subtle, and requires expertise and focus to notice. Both of those things are in short supply now, and it’s unclear to me where it’s supposed to come from in the future. And all AI is either being either given away or sold below cost, and every time a company has tried raising prices to a sustainable place, people won’t pay. OpenAI is trying to sell ads and do adult chat because they have cash flow issues.
Machine Learning can be really valuable. You want to spot patterns in data, it’s great. But LLMs are not the future, though it’s apparently going to be another thing people are only going to learn via touching the stove.
No one’s actually produced production level code for any meaningful programs. LLMs are good at writing documentation from already existing code, but you need an expert familiar with the code to review the documentation. Which is always the problem. Every AI company says you can’t trust the output, but unless you trust the output, there’s no savings if you want to use the output safely. Because reviewing completely brand new code takes longer than reviewing code you’re familiar with, and is more tiring.
LLMs are analogous to Elon Musk: “He talked about electric cars. I don’t know anything about cars, so when people said he was a genius I figured he must be a genius.
Then he talked about rockets. I don’t know anything about rockets, so when people said he was a genius I figured he must be a genius.
Now he talks about software. I happen to know a lot about software & Elon Musk is saying the stupidest shit I’ve ever heard anyone say, so when people say he’s a genius I figure I should stay the hell away from his cars and rockets.”
They’re bullshiters, pure and simple.
@Michael Reynolds:
And I doubt the search engine guys pay WIKI anything.
@Kevin:
Anyone who thinks colonizing Mars is either feasible or makes any sort of sense whatsoever is clearly and unambiguously a shithead.
Anyone who believes this should spend an hour with one of the popular LLMs grilling it on a subject they already personally know about. It’s not very good. It’s like the guy at the end of the bar who thinks they know everything, except obsequious instead of argumentative, and with less risk of vomiting. (An improvement over the barfly, to be sure)
It can generate code at the level of an intern, requiring large amounts of supervision. Typically, interns are useless on net, and the internship is just a months long job interview looking to see if the little rugrat can show signs of growing past that.
And then there’s the cost. None of the current systems are priced at a spot where they can make a profit — they’re just trying to undercut prices of humans and then will jack it up (the Uber model, where they originally cost far less than taxis and now cost more).
And the data centers require so much energy that they destabilize local energy markets.
But, that’s all assuming that AI is being used as a real tool and evaluated on those metrics. I worked at a startup decades ago, and quickly realized that what we were ostensibly doing was not the real goal — we weren’t really making a product, we were selling a dream to the investor class, telling them that we would be revolutionizing the internet and that they could get in on the ground floor. The actual users were just a means to an end, and their happiness was almost irrelevant, so long as a non-expert investor could look at it and think “yes, stupid peasants might like this.”
Here the dream is straightforward, and the dream of every C-suite executive: get rid of employees. It doesn’t have to be successful, it just has to be convincing to non-experts that long-term it will be successful. And even if all it does is make employees nervous and “grateful to have a job”… that’s part of the dream.
@Michael Reynolds:
There are many customer service applications where the goal is to prevent the customer from actually accessing the product. I think it could be very helpful for health insurance companies, for instance.
There are lots of tasks that need to be done, but don’t need to be done well. Generative AI is great at those.
On that note, I wonder if the image on this post is AI generated or just awful. I kind of hope it’s AI, as I would feel bad for an artist who had to create it. There are lots of things about it that scream AI, but there are also some bad artists that AI was trained on, and the low resolution and blur… Either way, it has a “What if WPA, but bad?” quality.
@charontwo:
Riddle me this, Charon: what is the age of consent on Mars?
It’s the age-old libertarian fantasy of having somewhere current government laws and regulations don’t apply, so they can be free (and fuck children). There’s a long history of sea steading failures, with new countries created on abandoned oil platforms — this is just a more modern version. They usually fail before the children can be brought over.
It seems like an oversimplification, but Mars is being pushed by Elon Musk, the man who kept emailing Jeffrey Epstein trying to get invited to pedophile island for the parties, so I’m pretty confident in that simplification.
@charontwo: I see Musk has backed off Mars. Now he’s hot to colonize the Moon. Stealing a line from I forget who, hopefully not here, it’s a matter of efficiency. He can not go to the Moon six or eight times in the time it would take to not go to Mars once. Musk claims, among other things, to be an engineer. He has no engineering degree. And he apparently never learned the basic rule, EMMS, Engineering Must Make Sense.
I like the idea that an AI could do certain kinds of rudimentary research for me (say helping update simple data sets or basic facts). I would even be willing to pay for such work, but as long as I have to worry about hallucinations, it seems like a cost not worth incurring.
If I have to go back and make sure that, say, all the electoral data I needed was properly procured and entered, what’s the point of paying to save my time?
OTB commenters seem really negative on “AI” and also seem to be rooting against it (though that is more of an impression).
It sounds like folks have some experience with public facing LLMs, which they find wanting. I agree that they are imperfect… frustratingly so.
I’m curious what, if any, other sources people are accessing to assess “AI” present and future. Who are folks reading/watching/listening to to keep up-to-date on this?
@Mimai:
100% this. I like AI, but I am not paying for it. I liked FB, but I ain’t paying for it. FB is now a dumpster fire scroll and put down. AI is on that same path but it consumes way more power, requires stronger chips etc….for what? Every “expert” on AI has a stake. It feels like a perpetual circle jerk. We HAVE to adapt it etc….they already seeing the slow adoption and pull back of investment in AI. It’s a great new thing, but way overhyped and is fortunate to come about when tech is flush with cash and investment. It is being oversold
@Mimai:
Discussed at LGM yesterday, lots of links in the comments.
“LGM”
“Marcus”
“Farrell“
@charontwo: Gotcha, thanks. I don’t read LGM so it wouldn’t have hit my radar.