This is the second post in a series on AI and the real economy. Last week we laid out the frame: change is certain, progress is not. This week, we address the question everyone asks first: what does AI do to jobs? We answer this question as told through a story of how we’ve argued about it.
Before we go a single step further, here’s why this question matters: Right now, in Washington and in statehouses across the country, people are deciding what rules should govern artificial intelligence, including whether companies must tell anyone what it’s doing to their workforces, whether the tax code keeps rewarding businesses for replacing employees with machines, whether a laid-off worker gets anything more than a shrug.
Those decisions are being made based on what the public believes about AI and jobs, and the strength of public concern around this issue is driven in no small part by the confusing debate around this topic. Unfortunately, much of this debate is being shaped by a handful of men who own the companies selling it.
That’s the stake. Not an abstract argument about technology. Whether your job, your kid’s first job, and the safety net underneath both get protected depends on whether we’re arguing about the right thing. And for three years, we mostly haven’t been.
So, let’s start with a man who changed his mind.
Act One: Doom and Gloom
In May of 2025, Dario Amodei, the chief executive of the AI giant Anthropic, sat down for an interview and said something that made headlines around the world: Artificial intelligence, he warned, could wipe out half of all entry-level white-collar jobs, and the unemployment rate might climb to ten or twenty percent within a few years. He used the phrase “white-collar bloodbath” to describe this future.
Think about how strange that is. This is a man whose company sells the thing he was warning about. Imagine the CEO of Ford holding a press conference to announce that cars might kill us all.
But rewind a little further, because the concern started earlier.
In the Winter of 2022, the generative AI chatbot ChatGPT shows up and everybody’s aunt is suddenly asking it to write a limerick. Almost immediately, the researchers start running the numbers, and the numbers are terrifying. One heavily cited study concludes that eight out of ten American workers have a job where the machine could do at least a tenth of what they do all day. Goldman Sachs floats a figure that gets repeated everywhere: three hundred million jobs worldwide, exposed.
Three hundred million. That’s every job in the United States, twice over.
Naturally, people panicked. This was the era of “AI is coming for your job,” and it felt less like a prediction than a countdown.
But here’s the thing almost nobody noticed at the time. Buried in that scary number is a word doing an enormous amount of quiet work: affected. A job where AI can do ten percent of the tasks is not a job that disappears. It might be a job that gets easier. Your accountant using a calculator is “affected” by the calculator. She still has a job. She may, in fact, have a better one.
Hold on to that. It comes back.
Act Two: The Honeymoon
By 2023 and into 2024, researchers stopped guessing and started measuring. They took real workers—customer service agents, writers, consultants, and computer programmers—gave some of them AI tools, and watched what happened.
And the workers got faster. Noticeably faster. The research found that customer service agents resolved about 14 percent more issues per hour. The programmers finished a coding task 56 percent faster. Even more interesting: the biggest gains often went to the least experienced people in the room. Among those support agents, the novices improved by 34 percent while the veterans barely improved at all. Suddenly, the newcomer with the AI could suddenly keep pace with the veteran.
So, the mood flipped. AI wasn’t a wrecking ball; it was a power tool. Nobody frets that the nail gun will replace the carpenter.
But there were people at the back of the room clearing their throats. Simon Johnson and Daron Acemoglu, both economists at MIT and future Nobel laureates, had just published a thousand-year history of technology called Power and Progress. Their argument cut against the entire honeymoon. There is a comforting belief, they wrote, that better technology automatically means higher productivity, which automatically means better wages for everyone. They gave that belief a name—the “productivity bandwagon”—and then spent a book showing that history mostly doesn’t work that way.
New machines raise average productivity, they argued, while often reducing what an individual worker adds. And who captures the gain has always depended on power, not on the technology itself. Their warning about AI specifically: we are using it too much to automate people’s work and not enough to make people better at their jobs.
Nobody wanted to hear it. The honeymoon rolled on.
Act Three: The Reckoning
Then came the awkward part.
If AI was making everyone so much more productive, that productivity should have started showing up in the national statistics. The economy should have been visibly humming.
It wasn’t. And into that silence, two things rushed in.
The first was a lie that companies discovered they could tell. When a business lays off five thousand people, it has to explain itself. “We hired too many people during the pandemic and now we’re cutting costs” makes the stock price sag. But “we’re restructuring around artificial intelligence”? That makes you sound like a visionary. Wall Street applauds.
It got a nickname: AI washing. Dressing up an ordinary layoff in futuristic clothes. And we’re not speculating here as some of the executives eventually admitted it, and one major survey found that firms that cut staff in a shift to AI showed no corresponding increase in revenue.
The second thing to rush into the silence was fear, and this is where Amodei’s “bloodbath” warning landed. Notice what had changed, though. The 2022 fear was about what the machines could do. The 2025 fear was about what companies would do with them: use AI as both an excuse and a weapon.
That’s not a technology story anymore. That’s a story about people making choices.
Act Four: The Walk-Back
Now fast-forward one year. May of 2026. Sam Altman, who runs OpenAI, is on stage in Sydney. Someone asks him about the job losses everyone had been bracing for. And Altman says he is “delighted to be wrong.” The damage he’d expected simply hadn’t shown up.
Amodei softened too, reaching for a two-hundred-year-old economic idea about how making things cheaper sometimes means we just buy more of them, not less. The classic example is coal: when steam engines became more efficient, factories didn’t burn less coal, they built so many more engines that they burned far more of it. The argument Amodei is making that that cheaper AI labor might not shrink the number of jobs so much as multiply the amount of work we ask to have done.
Twelve months. Same technology. Same men. Completely different tunes.
Jamie Dimon, who runs JPMorgan Chase and is nobody’s idea of a starry-eyed optimist, wrote to his shareholders that AI will eliminate some jobs, will enhance and create others, and might eventually let people work shorter weeks. Sensible. Measured. Boring, even.
And then a fourth voice—Yann LeCun, one of the founding figures of modern AI—said the thing worth taping to your refrigerator: “Don’t listen to CEOs. They have a vested interest in propping up the power of the products they sell.”
Sit with that, because it cuts in both directions. These are salespeople with two products to sell at once: to investors, a technology powerful enough to remake the economy; to regulators and the public, a technology tame enough to leave alone. So, it’s not surprising that the executives who told us that AI would destroy the world and the executives now telling us not to worry are, in many cases, the same people. In both cases, they were raising money, selling products, or managing how the public feels about them while they did it.
That doesn’t make them wrong. It just means their words are advertisements as much as observations. You wouldn’t take a car salesman’s word on whether you need a new car.
So: were they lying then, or are they lying now? That’s the natural question, and it’s the wrong one. Because the more interesting truth is that the whole conversation has been changing shape, and if you understand how it changed, you’ll understand this issue better than most of the people arguing about it on television.
What actually changed?
Here’s the punchline: for three years we’ve been asking the wrong question. “Will AI destroy jobs?” is a yes-or-no question, and it has no yes-or-no answer. It’s like asking “will the weather be bad?” Bad for whom? Bad for the farmer or bad for the ski resort?
The question that survived the argument is better, and it comes in two halves:
Productivity for whom? Somebody is getting the benefit of all this. Is it the worker who now does her job better, or the shareholder who now needs fewer workers?
Displacement for whom? Somebody is paying the price. And it isn’t spread evenly, as it lands hardest on particular ages, particular occupations, particular towns.
Notice what kind of questions those are. They’re not questions about circuitry. They’re questions about who gets what, which are questions about choices, and choices can be argued with. Nobody votes on whether a technology exists. We can absolutely vote on who it serves.
What this means for you
If you work for a living, the reassuring headlines now bubbling up don’t settle anything about your situation. The national economy can look perfectly healthy while your employer decides it needs three people in your department instead of five. The CEOs’ change of heart was about the big picture. You don’t live in the big picture. You live in your department. And the impact on your and your family’s livelihood depend both on those decisions as well as the actions policymakers are making today to provide a safety net should the worst case arise.
If you’re young, or you have a kid or grandkid starting out their career, this debate is aimed squarely at you. The evidence—which we’ll dig into next week—keeps pointing at the same place: the entry-level jobs. The tasks we hand to the twenty-three-year-old are exactly the tasks the machine finds easiest. That’s the rung of the ladder that’s wobbling, and it’s the rung everyone must step on first.
And if you’re comfortable, don’t be too comfortable. There’s a common assumption that this is a story about workers on a factory floor somewhere. It isn’t. The jobs most exposed to AI are the ones that involve sitting at a desk, reading, writing, and analyzing: lawyers, programmers, accountants, consultants. The corner office is closer to the blast radius than the loading dock is. Meanwhile the electrician, the nurse, and the plumber are, for now, largely untouched.
What can policymakers do?
If you take nothing else from this, take this: we have been arguing for three years about something we can barely see.
While the future may be uncertain, every big company in America knows exactly how it’s currently using AI and exactly what that’s doing to its payroll. Almost none of them will tell us. The public debate runs on press releases, executive interviews, and vibes while the people who actually have the data keep it in a drawer.
Here’s what policymakers can do: Make large employers disclose what AI is doing to their workforces. That wouldn’t settle the argument. But it would let us have the argument with the lights on as we decide what to do next. Given that everyone’s job is the thing on the table, that hardly seems like too much to ask.
So where does that leave us? Next week: what the evidence actually shows.
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.