The Ground Is Still Sinking. Who Left First?

 

43,633

Between 1989 and 2017, the U.S. federal government bought out 43,633 floodplain properties. A buyout means the government purchases your house, demolishes it, and places the land under a permanent ban on building.

Forty-three thousand sounds like a lot. But those properties are spread across 49 states and 1,148 counties. Per county, the median is eleven. Five hundred sixty counties bought out somewhere between one and ten houses in the space of several decades.

And the median size of a buyout project nationwide is three properties.

One project. Three houses.

This is the largest government-led retreat mechanism the United States has.

One clarification first: those 43,633 are floodplain properties nationwide, not coastal ones. The buyout grew up mainly along inland rivers; the coast is one category within it. How the total splits between coastal and inland, the original study doesn't say — so I won't report a figure.

What managed retreat actually looks like

The term sounds like a column falling back in good order under command.

In practice: three houses on this street come down. The one next door doesn't, because the owner won't sell. The one across the street doesn't either, because it isn't on this round's funding list. What's left is a checkerboard of mown empty lots, and the people still living beside them.

It is slow. From disaster to project closeout takes 5.7 years on average; the longest ran close to seventeen.

And it is fragmenting. From 1989 to 1998, the average project bought out nineteen properties. By 2009–2017, that had fallen to seven.

The scale isn't growing. It's shrinking.

The one time it went furthest

In Terrebonne Parish, Louisiana, there is an island called Isle de Jean Charles. In 2016, the state won a grant through HUD's National Disaster Resilience Competition to move the island's entire population inland — the furthest America has ever pushed the idea of relocating a whole community together.

The new community, The New Isle, was built on roughly 515 acres. Residents began moving in from August 2022. By the fall of 2024 the official count stood at thirty-seven, with nine more units still being delivered. Elapsed time: six to eight years.

Look closely at that number. The official document says "37 residents or families" — whether the unit is people or households, the document itself never specifies. How many moved in total does not appear in the public record.

It did not become a success story

The community on the island is an Indigenous tribe. The trouble began the moment the money arrived: the tribe is not federally recognized, so it could not legally receive the funds directly. The money went to the state, which ran the project.

Then the two sides broke apart. In December 2023 the tribe filed a Title VI complaint. One dispute was over a name — the state had dropped "Isle de Jean Charles" from the name the tribe uses for itself, calling them only the Biloxi-Chitimacha-Choctaw. That, the complaint said, severs the tribe from its homeland in language, and denies its right to name itself.

Another was over a grocery store. The state said it was negotiating with a large chain. The leaders said they wanted the tribe's own community market. The official replied that the chain was what they needed, and that they should be satisfied with it.

Albert Naquin, who pushed for the move for nearly twenty years, never moved into the new community himself: "We were supposed to be a model for other communities. The state took over and messed it all up. Where are the people of Isle de Jean Charles?"

A few households never left at all. The fisherman Edison Dardar put up a hand-painted sign: "We're not moving off this island. Anybody who wants to move can move, but leave us alone."

Scholars note that in the first three years after the state took over, the project's attention went largely to household-by-household outreach, housing options, and site design — not to carrying forward the tribe's existing community institutions.

Money solves the construction. It does not solve "after the move, are we still us?"

The layer that matters more

So far this reads as an expensive, slow relocation that never quite came together. But there is another layer that explains more than all of the above.

Before the project began, most of the island had already gone.

In 2002 there were seventy-eight households on Isle de Jean Charles. By the late 2010s, fewer than twenty-five remained. By household, roughly two thirds had already left over decades of land loss. I won't give a headcount — the public record supports no reliable one.

They weren't moved out. They left on their own. And not all of them left because the sea was coming: some because the road off the island kept flooding and they couldn't reach work; some because shrimping and other fisheries declined in the 1980s.

One point is easy to get wrong, so here it is precisely: some of those thirty-seven had left the island long before the project started. The resettlement plan includes a category for people who moved away before Hurricane Isaac in 2012, letting them take a free lot and build. It came with conditions — proof you could finance construction, and residence in one of several designated parishes as of August 2012. So "those who left first" and "those who were resettled" are not two separate groups. You cannot subtract one number from the other.

What is certain: the dozens of households who left first did so with no project and no funding. They appear on no public list I could find.

Two retreats

The official one has a legal basis, federal funding, a project number, a closeout report. It enters the statistics, gets analyzed and reported. Three houses per project, 5.7 years, and at its most complete, thirty-seven.

The market one is people listing their houses, selling, and going. No project, no name, and no place in any retreat statistic I could find.

A study by Rice University's Kinder Institute and its center for coastal futures used address-level data from more than five hundred buyout counties between 2007 and 2017, tracking over seventy thousand movers. James Elliott and Deborah Banerjee concluded that nationally, the vast majority of movers — roughly fourteen out of every fifteen — were not federal buyout participants at all, but neighbors who moved through ordinary real estate transactions.

Fourteen out of fifteen. On the very blocks where the government was buying houses, the official program accounts for about one-fifteenth of the people who actually left.

The retreat has been happening all along. It just isn't happening through the program named "retreat."

What the sieve is

If most people leave by selling, then what decides who goes first isn't policy. It's who can sell. The evidence here is not clean:

• Between counties: counties that completed buyouts have more people, higher incomes, more education. But the researchers read income and population as proxies for local administrative capacity — not as proof that wealth causes buyouts.
• Within counties: buyout sites fall on average in lower-income, less-educated neighborhoods.
• On race: the findings conflict. A ZIP-code-scale national study found buyouts concentrated in a county's more disadvantaged neighborhoods. A census-tract study controlling for flood losses and income found that in urban core counties, whiter neighborhoods were more likely to get the program.

The two point in opposite directions. So this piece will not tell you America buys out poor neighborhoods on purpose, and it will not tell you anyone is screening by race. The available evidence supports neither claim.

Outcomes diverge too. Nationally, about 70% of buyout sellers moved to lower flood risk. But in the same data, the lower the income and the more Black and Hispanic the neighborhood, the farther and more scattered owners moved. The Staten Island sample moved almost entirely into higher-poverty census tracts, about a fifth of them into higher flood risk.

All that can be said is this: retreat is never a whole city stepping back at once. It passes through application forms, appraisals, and property titles first — and then it sorts people.

The people who don't even get to sell

Federal law treats the homeowner as a willing seller: you decide whether to sell, and you can negotiate the price. Tenants forced out because their building was acquired are classified as involuntarily displaced — in theory entitled to moving costs, rent differential payments, and relocation counseling.

So tenants are not uncompensated. Nor do they have the owner's choice. An owner can at least negotiate a price; a tenant usually waits for the landlord's decision. And in the data afterward, I found no public tracking of where those tenants went.

In one retreat, the people were counted. In the other, all that remains is a property transaction record. And outside both, there are people to whom even that record doesn't belong.

Where the risk goes

After an official buyout, the land is permanently deed-restricted. The house comes down, the lot stays empty, and the risk exposure at that spot is eliminated.

Market retreat eliminates nothing. You sell, and the risk transfers to the next buyer or the next tenant. The house is still there, someone still lives in it, and the next time the water comes it floods the same house.

One retreat reduces risk. The other only moves it. The larger one is the second.

About "underwater by 2100"

The widely circulated claim about New Orleans runs like this: at a certain sea level the city becomes an island in the Gulf; higher still, no coastal defense works. The claim has a source, but four things need stating precisely.

One: it comes from a Perspective article, published in 2026 in Nature Sustainability. A Perspective offers an assessment, not new observational data. The accurate phrasing is "an assessment paper proposes" — not "a study measured" or "the data show."

Two: the two heights are the ends of one range. The paper gives three to seven meters. At three, New Orleans would be at best a highly exposed island in the Gulf; at seven, no coastal defense system can be expected to work.

Three: that three-to-seven meters is relative sea level. It already includes Louisiana's own land subsidence. It is a local figure, not global mean sea level, and the two cannot be compared directly.

Four: on timing. The three-to-seven-meter figure carries no timetable, and the paper says outright it can't give one. It does make one statement with a year attached: projecting from a 75% wetland loss around 2070, the defended New Orleans area is likely to be surrounded by the Gulf before the end of this century.

Keep these apart: one is a long-term height with no timetable, the other a landscape judgment with a year. Stitch them together and you get "three meters of sea level rise by 2100" — which the paper never said.

For comparison, the IPCC's Sixth Assessment gives global mean sea level rise by 2100, against a 1995–2014 baseline, of 0.38 m under low emissions and 0.77 m under high. That is not the same quantity as three to seven meters, so I won't say the paper's range far exceeds the IPCC projection. Those are two different rulers.

What can be said with a year attached is the wetland line. A 2024 study in Nature Communications, based on thirteen years of observations from 253 monitoring stations along the Louisiana coast, states in its abstract that under the intermediate SSP2-4.5 scenario, submergence of about 75% of Louisiana's coastal wetlands by 2070 is "a plausible outcome." The authors added their own caveat: they were unwilling to convert that into a precise rate of wetland loss.

I go through this in such detail not to look rigorous, but because once these numbers are stitched together they form a causal chain that looks powerful and does not exist in the original research.

Closing

The more distant the number, the easier it is to stitch. The retreat happening now has no number at all.

It isn't that America has no mechanisms. There are federal mitigation grants, community development block grants, a voluntary community relocation program at Interior. There was also a community resilience program, terminated by FEMA in April 2025; late that year a court ruled the termination unlawful and issued a permanent injunction, and it still hasn't actually been restored. Nor is there no precedent: Newtok and Shishmaref in Alaska are underway, and Fiji completed a whole-village move inland back in 2014.

The accurate statement is this: there is no unified, advance, city-scale retreat plan. There are only projects launched after each disaster, counted one house at a time. Added up: decades, and more than 43,000 properties.

And alongside those projects, people are already leaving. They sell their houses and move tens of miles away, or to another state. This has no name. I found no one counting it.

An island lost two thirds of its households before its own relocation project even began.

The retreat we can see is so small it looks like an experiment. The one actually underway, we know only that they sold their houses.

We don't know where they went.

Nothing Needs to Be Said

 

Nineteen hundred yuan. That's the rent one woman pays every month.

She's married. She has a family, and a home. But not far from that home she rented a second, very small place. The rent is about one sixth of her income.

According to a feature report, after she moved out, her husband began coming on weekends to help her set up the furniture.

The same report includes another woman. She moved into her employer's dormitory. After she did, the report says, her husband "straightened out his stance and his attitude."

I'll leave those two cases there without explaining them. I only want you to notice one thing: the same request, raised at home, produced no change the report bothered to record. Raised after she moved out, the report recorded a change.

The same words, said somewhere else, carry weight. Why?

---

A thin book from 1970

In 1970 Albert Hirschman published Exit, Voice, and Loyalty. The problem he set out to solve was simple: when an organization starts to decline, what can the people inside it do?

Only two things, he said. Leave, or speak. Leaving he called exit. Speaking he called voice.

On page thirty-four he writes that when the exit option does not exist, voice is the only recourse — and gives three examples: the family, the state, the church.

By page seventy-seven he is blunter. Exiting the family, the tribe, the church, the state — those most primordial human groups — is usually unthinkable.

Though, he adds, not always entirely impossible.

He kept that second half for himself. That second half is what this piece is about.

---

What's actually on the table

Not some abstract "status within the family." Something you can count in minutes.

China's Fourth Survey on the Social Status of Women (reference date July 1, 2020; thirty thousand valid individual questionnaires) reports that married women average 120 minutes of housework a day, plus 136 minutes on caregiving, tutoring, ferrying children, and looking after elderly or sick relatives. Employed women's total workday labor comes to 649 minutes — 495 paid, 154 unpaid.

On that last figure the report adds one line, which I'll quote as written: roughly twice that of men.

The National Bureau of Statistics' Third National Time Use Survey (fieldwork May 2024) puts daily unpaid labor at 1 hour 52 minutes for men and 3 hours 29 minutes for women.

So the thing being negotiated over is two or three hours a day. It appears on no pay stub anywhere. It gets allocated every single day.

---

Voice is expensive

Start somewhere easier than marriage. You work at a company and you think a process is broken. You raise it. That costs time. You have to pick your moment, accept the risk of being remembered for it, and it may not work anyway.

That's voice. Hirschman's formal definition, page thirty-one: any attempt to change, rather than escape from, an objectionable state of affairs.

Compare exit. You quit. You persuade no one. You argue with no one. You go.

Hence page forty-one: compared with exit, voice is costly, and it depends on how much influence you can bring to bear inside the organization — on how much bargaining power you have.

Whether speaking works does not depend on whether you're right. It depends on how much bargaining power you have.

So where does bargaining power come from?

---

The threat point isn't divorce

An early class of bargaining models treated divorce as the threat point: no deal, and the marriage dissolves, so divorce conditions determine who calls the shots at home (Manser and Brown 1980; McElroy and Horney 1981).

In 1993 Lundberg and Pollak, writing in the Journal of Political Economy, proposed something different. In everyday bargaining between spouses, what's operating usually isn't the threat of divorce — because that threat is too big to be credible. You don't threaten divorce over who does the dishes, and if you say it out loud, the other person doesn't believe you.

The real threat point, they argued, is the noncooperative state inside the marriage: each side retreats to its own role, minds its own business, stops optimizing jointly for the household. Nobody left, but cooperation stopped. They called the model separate spheres.

Now look back at that nineteen hundred yuan.

These women aren't threatening anyone with divorce. What they did was take "each living separately inside the marriage" — abstract, unprovable, impossible to point at — and turn it into something priced, visible, and pointable.

It's right there. Nineteen hundred yuan a month, and an actual apartment.

---

The hard evidence: exit gets cheaper, and the people who stay are the ones who change

Is being able to leave enough, without leaving?

American states adopted unilateral no-fault divorce laws in different years — one spouse can end the marriage without the other's consent. Because the timing varied by state, researchers treated it as an approximate natural experiment. In 2006 Stevenson and Wolfers published the result in the Quarterly Journal of Economics.

Their estimates: after unilateral divorce laws took effect, female suicide rates fell by roughly 8–16 percent, severe domestic violence against women by roughly 30 percent, and the share of women killed by an intimate partner by roughly 10 percent. All relative declines, not percentage points.

Here's the step I want to make explicit, because it's my inference and not something the paper states directly. The study uses state-level aggregates; it tracks no individual. But the rise in divorce rates over the same period is far too small to account for those improvements. So the reasonable reading is that most of the improvement happened to people who did not divorce.

The law made leaving slightly easier. The change showed up among those who stayed.

Hirschman had written the mechanism down on pages 82–83:

Loyalty postpones exit, but the very existence of loyalty presupposes the possibility of exit. That even the most loyal member can exit is often precisely a key source of his bargaining power. When voice is backed by the threat of exit, its effectiveness as a repair mechanism increases substantially —

whether or not that threat is actually spoken aloud, or is merely understood by everyone present, tacitly, as part of the situation.

Nothing needs to be said. The woman doesn't have to say anything. She only has to actually move into the dormitory, and the threat is already in the room.

One caveat: this evidence is American, from the 1960s through the 1990s. It shows the mechanism exists. It does not show it exists in China at the same strength.

---

But moving out is not the same as having an exit option

The tempting one-line takeaway here — a woman moves out, and she gets bargaining power — doesn't hold. There's a whole class of counterexamples.

A doctoral dissertation at the University of York on "non-cohabiting marriage" in China includes six mothers living long-term apart from their husbands to accompany a child's schooling. All of them moved out — some to a provincial capital, some abroad.

One, forty-six, is abroad with her son during his high school years. The household runs entirely on what her husband earns in Beijing; her visa status is dependent, so she can't legally work full time. Being abroad alone as an accompanying parent, she says, is a loneliness that soaks into your bones and can't be told to an outsider. She later discovered her seventeen-year-old still sucked his thumb, and her first reaction was intense self-blame.

Another, forty-four, gave up a rising career at home. Many nights she lies awake alone, she says; she has become, completely, a live-in maid circling the child and the stove. Her husband keeps accumulating resources back home. Her industry wrote her off long ago.

Every one of them moved out. Not one gained bargaining power from it. The opposite: more economic dependence, an interrupted career, emotional isolation.

So what's the difference? One question — and it's the tool I want to hand you:

Does your leaving impose a cost on the other person?

When the woman who rented a room left, the housework left, the caregiving left, the emotional labor left. Her absence is a cost he has to absorb. That's why he shows up on weekends to help with the furniture.

When the accompanying mother leaves, the housework still gets done and the child still gets raised — just in a different city — while she loses income, career, and social ties. Her leaving doesn't cost her husband anything. It lowers his costs.

Same act of moving out. Exactly opposite directions.

What imposes a cost is an exit option. What doesn't is relocation.

---

Counter-evidence: money doesn't always work

In 2003 Bittman and coauthors published a study in the American Journal of Sociology titled with a question: When Does Gender Trump Money?

Using Australian and American data, they found women's housework time does fall as their share of household income rises — but only up to the point where the spouses earn about the same. In the Australian sample, once the wife's income passes half, her housework time starts climbing again.

Three things to get exactly right, because they're easy to run together. First, in the American sample the famous nonlinear effect actually shows up in the men: those most dependent on their wives' income do less housework, possibly compensating for the gender deviance of economic dependence. In the women's equation the nonlinearity isn't significant. Second, that male effect isn't robust — the authors themselves note it disappears when a few extreme cases are dropped. It's a lead, not a conclusion. Third, the U-shaped curve is the one confirmed in the Australian sample.

And these samples are Australia 1992 and the US 1987–88. They are not evidence about contemporary China.

Chinese studies (Yu Jia 2014; Liu Aiyu et al. 2015; Jian Minyi and He Guangye 2018; Sun Xiaodong 2018) broadly find that the "rebound past fifty percent" turning point does not replicate at the national level. Urban wives' routine housework follows the linear resource-bargaining prediction. The reversal appears mainly in rural samples, and the turning point isn't stable at fifty percent.

So the correct statement is that the resource effect is conditional — on urban versus rural, on which kind of housework, on gender attitudes. Not a uniform national law. But not absent either.

One more. Duncan and colleagues, drawing on a nationally representative British survey plus fifty in-depth interviews, conclude that living apart doesn't necessarily dissolve the traditional gendered division of labor and often comes with its reproduction; economically fragile groups rarely gain empowerment from it. The opposing camp, represented by Upton-Davis, argues living apart lets women sidestep housework and escape male authority while keeping the intimate relationship.

Same phenomenon, two camps reading two opposite directions out of it. Both belong on the table.

---

The line: down 42.7% in a single year

On January 1, 2021, China's divorce cooling-off period took effect.

Divorces registered with civil affairs bureaus, from the Ministry of Civil Affairs' bulletins: 4.047 million couples in 2019; 3.736 million in 2020; 2.141 million in 2021 — down 42.7 percent in one year.

Then the rebound: 2.10 million in 2022, 2.5937 million in 2023, 2.622 million in 2024, 2.743 million in 2025.

Demand didn't disappear. It was postponed, or rerouted.

(One caution: registered divorces and legally processed divorces are two different numbers — the latter also includes court judgments and mediated divorces. They can't be added or used interchangeably. Everything above is registered divorces.)

So what is a cooling-off period doing? Hirschman wrote about this in 1970, on pages 79–80:

Specific institutional barriers to exit can often be justified on this ground: they can call forth voice in organizations that are deteriorating but are still salvageable… This seems to be the most defensible reason for making divorce procedurally complex and for making people spend time, money, and emotional wear on it — even if that often wasn't the direct intent behind its design.

Making exit harder in order to force negotiation: he wrote the logic down more than fifty years ago.

But there's another half to the book, and it usually gets ignored. Page forty-three: the availability of the exit option can cause the art of voice to atrophy. That, he says, is one of the book's central arguments.

On page eighty-three he puts the halves together and states the condition:

Exit lowers people's willingness to develop and use voice, but raises their ability to use it effectively. … Exit should be possible, but it should not be too easy, and not too attractive.

So the complete proposition is a range, not a direction. Exit too easy, and people don't bother talking — they just go. Exit too hard, and talking carries no weight, because nothing stands behind it.

That nineteen hundred yuan lands inside the range. Affordable, so the threat is credible. Not enough to buy out an entire relationship, so she isn't simply walking away.

That's not a coincidence. The price is the range. And what institutions do is move its edges.

I'm not going to judge the cooling-off period here. I only want to say: it is pricing exit, and pricing changes how the people who stay negotiate.

Worth noting alongside it, because it's the same move pointed the other way: a 2022 Supreme People's Court provision states that applying for a personal safety protection order is not conditional on filing a divorce suit. The lawmakers performed an unbundling — pulling immediate protection of personal safety out of dissolution of the marriage, so the first no longer waits on the second.

But article nine of the same provision adds a limit in the opposite direction: in divorce cases, a party asserting domestic violence solely on the ground that a protection order was once issued still has to have the matter assessed comprehensively.

The piece that was pulled out doesn't automatically accrue back to the main case. The unbundling runs one way.

---

And one more thing: almost none of this is recorded

While preparing this, I went looking for a number. How many married women in China rent a place of their own? How many couples live apart?

I couldn't find it. And the more I looked, the clearer it became that this wasn't carelessness. In publicly released statistics, at least for anyone under sixty, that number isn't there.

Layer one: the census. It registers people by current residence; each person registers in one place. The National Bureau of Statistics states plainly that spouses living apart each register at their own usual residence, and someone registered to a household but not living there isn't counted in that household's size. A couple living in two places is two households.

The effect: the census's unit for measuring families is people who live together, not people joined by marriage. So "married but living apart" has no slot in what the census outputs.

Not to overstate it — the census does record marital status and relationship to the head of household, so in principle a person with a spouse elsewhere can be identified. What's genuinely missing isn't the data. It's the indicator. No table aggregates it into a number and tracks it year after year. It isn't uncountable. It just hasn't been treated as something worth counting regularly and showing to people.

Layer two: who gets asked. The Bureau's annual population change sample survey has a "living arrangement" item, with options including living with spouse and living alone. Combined with marital status, you could in theory compute the share who are married but not living with a spouse.

That item is only asked of people aged sixty and over. So was the equivalent in the Seventh National Population Census, and in the dedicated surveys of older adults. It isn't that nobody asks the question. It's that when it's asked, the respondents are always the same group — and the items usually sit under the heading of elder care.

Layer three: the courts. Marriage registration statistics record two instants, marrying and divorcing; the long continuous stretch in between goes unrecorded. And on the judgment side the door is shut: a 2016 Supreme People's Court provision lists divorce litigation, and cases involving custody or guardianship of minors, among the judgments not published online.

Three layers add up to a living arrangement that, before you turn sixty, no statistical table treats as its subject.

What we get to see is the divorce rate rising and falling — the most sluggish indicator in the whole picture, since it only moves once the thing is already over.

Which is why you won't read here that "more and more women are doing this." I don't know how many there are. And that fact is itself part of the point.

---

Two questions

This isn't only about marriage. Any long-term relationship — a job, a partnership, a lease, a family — can be measured with two questions.

One: if I leave for a while, what does it cost me?

Not what divorce costs. Not what quitting costs. What partial leaving costs. Nineteen hundred a month, or you can't come up with a cent? That number is your weight at the table.

Two: does my leaving impose a cost on the other person?

If the answer is no, it doesn't matter how far you move. The accompanying mother crossed half the globe, and her bargaining power still fell.

Both questions are concrete. Both can be worked out.

I'm not urging anyone to divorce, and not urging anyone to endure. This isn't a life plan. It's a ruler.

As for whatever number you measure — that's your own business.

Two 5.6%s: Six rulers, one graduating class

 

One week in April 2026, a new graduate who had just sent out her résumé could read two contradictory headlines on the same screen.

One said hiring plans for the class of 2026 were up 5.6% from last year — grads finally catching a break. The other was a Bloomberg cover story: this class is walking into the toughest entry-level market in years.

Neither is clickbait. Neither is made up. Behind each stands a legitimate data source.

Set them side by side for a day and you reach a natural conclusion: one of them has to be lying, or at least one of them quietly picked the number that suited it.

There's another possibility.

Neither is lying. The two are simply not measuring the same thing. One measures how many people employers plan to hire; the other measures where job seekers actually ended up. The first is a plan, the second an outcome. The first is a survey some companies filled out in February; the second is where a whole generation stands in April. A bathroom scale and a thermometer never contradict each other.

The same graduating class is a single negative. And in 2026 there are at least six developing solutions for it.

---

Six rulers

The employer survey. The national association of colleges and employers asks 185 employers a simple question: how many new graduates do you plan to hire next year? Fielded mid-February to mid-March 2026. Not probability sampling — members and non-members answer if they feel like it. It measures intent.

A Chinese job platform's posting database. It counts new postings published on its own site; if the title or description hits the keyword, it counts. The report says so itself: it only reflects what happens on the platform.

PwC's barometer and Stanford's payroll study — the real pair here. PwC is built on a billion online job postings. Stanford uses administrative records from the payroll system: twenty-five thousand firms, 4.6 million workers. The first counts postings that went up. The second counts people who actually exist on a payroll.

Each has to assign every occupation an "AI exposure" score, and those two exposures are not the same thing. PwC's is ability-level: break the occupation into abilities one by one, then assess whether AI could in theory cover them. Stanford's is task-level, and folds in actual usage logs. One measures what's possible in theory; the other, what's already happening. The graduations on these two rulers cannot be converted into each other.

Handshake, connecting posting volume on one end to student and recruiter surveys on the other. It counts postings and also asks people whether they're afraid.

The New York Fed's monthly series. It doesn't ask firms, doesn't count postings, doesn't ask about attitudes. It measures one thing: after searching everywhere, how many people still haven't landed.

Six teams, each standing in front of its own ruler and nothing else. The people running the survey can't see the payroll. The people counting postings can't see beyond the platform. The people measuring outcomes don't know how many firms meant to hire.

Anyone standing in front of any one of these rulers will honestly read out the direction they can see.

---

Changing rulers

Start with the smallest switch, so small it hardly looks like a problem: that +5.6% is reported as a median.

In the fall 2025 round, the same organization, the same employers, originally reported the mean — +1.6%. Which is where many stories got the line "a rebound from 1.6% to 5.6%," a steep recovery curve.

Recalculate that fall batch as a median and what you get is −2.4%.

This needs no third party to interpret. It's written in the data source's own asterisked footnote. That steep recovery curve has a mean at one end and a median at the other. They aren't even on the same ruler.

And even aligning the statistic, the problem only moves — because +5.6% measures a plan, not hiring that has happened.

That same spring, the sixth ruler gives: March 2026, unemployment for recent graduates 5.6%; for all workers 4.2%.

The same number landing on the same class. One says employers plan to hire 5.6% more. The other says 5.6% of those who searched still haven't landed. There is no arithmetic relationship between them; the coincidence is pure. But it's enough to show one thing: a number, on its own, carries no direction. The direction comes from the ruler.

Two more lenses follow.

Flow and stock. Postings are flow; people on payroll are stock. "Entry-level jobs are increasing" can be true and false at once depending on which you measure: more postings and fewer people employed, or the reverse.

Relative and absolute. Stanford's most-cited number is 16% — a relative decline against low-exposure occupations. In absolute terms the same data gives −6%. Both are correct; one asks "how much worse than everyone else," the other "how much less than before." Slip from relative to absolute and the picture in the reader's head changes completely. For how slippery this is: a paper citing that study wrote it down as 13%.

There's a quieter slip too. That 16% comes with a 95% confidence interval — but the original doesn't publish the interval's upper and lower values. It only draws a shaded band. What was citable was never a point; it was a band. By the third retelling the band is gone and only a number is left.

---

Three layers

The tempting thought at this point: if it's all measurement basis, every contradiction dissolves the same way.

It doesn't.

Layer one: all correct, nobody is wrong.

Employer hiring intentions are recovering. One analysis found firms with the highest AI investment intensity grew total headcount 10.2% and entry-level headcount 12% in the two years after adoption. And this class really is having a hard time. All three measure different things — intent, internal firm stock, individual outcome — and hold at once. That analysis also wrote its own caveat: correlation isn't causation; intensive AI adopters may already have been larger, more technology-intensive and faster-growing beforehand. Cite the number, cite the warning with it.

Layer two: this one really is an error.

The claim circulating on the Chinese internet — "AI campus-hiring postings up twelvefold" — comes from a job platform's report for January–February 2026. But that figure describes experienced mid-to-senior hiring in the new-economy sector: AI's share of new postings rose from 2.29% to 26.23%. The extremely low base is the main source of the multiple. The same platform has a separate campus-hiring figure of an entirely different magnitude.

That report is about experienced hiring from beginning to end. On another page, where it writes that postings requiring under one year of experience fell about 20% year over year, the source itself adds a parenthesis: experienced hires only.

The qualifier was written by the data source. The one who dropped it was whoever repeated it.

There's a thinner layer still: same platform, same indicator — stretch the window from two months to four and twelvefold becomes 8.7-fold.

Layer three: this one is a real fight.

PwC and two other researchers — Lambert, of Warwick and the LSE, and Schindler, of the Ellison Institute in Oxford — used the same underlying data vendors. The same job-postings database, the same hiring records.

PwC read out: entry-level jobs have been "seniorized."

Lambert and Schindler read out something else: what raised the step wasn't mainly AI — it was that nobody is in the office.

Their material is 243 million hiring records and 407 million job postings across the US, UK, Canada and Australia. Estimated separately, generative AI exposure and work-from-home exposure each predict about a five-percentage-point decline in the junior share of new hires by 2025. Two suspects, and each one fits. But put both into the same model and let them compete: the AI coefficient collapses, often becoming statistically indistinguishable from zero, while the work-from-home term holds.

Here a passage is still missing, and leaving it out would be unfair.

Stanford considered the entanglement. They pulled computer occupations out of the sample entirely, and computer-related firms too — the conclusion held. They split occupations into teleworkable and non-teleworkable and looked at each; the more exposed group grew more slowly in both.

But that second test rests on a premise worth spelling out. Telework and AI exposure are already highly correlated — so correlated that once the sample was split the cells ran short, and the two lowest quintiles had to be pooled just to assemble comparable groups. In other words, the test of "looking at them separately" was itself run under the condition that they don't separate very well. And that is precisely the starting point of the later working paper.

It was also exactly when discussing the non-teleworkable group that they left one word in their own concluding sentence. The original says the results for that group indicate their findings are not driven by outsourcing or work-from-home disruptions — at least not solely.

Those words are themselves half a concession to the later criticism. They were written in a 2025 paper, when Lambert and Schindler's did not yet exist.

One more point, and Stanford wrote it into their own text: in the test extending the sample backward, their two exposure measures give different pictures. Under one, the most exposed quintile had already begun growing more slowly from around 2020; under the other, that doesn't appear. They wrote that sentence on their own initiative. Nobody picked it out for them.

So the difference isn't about who was careless. It's two identification strategies: Stanford's is to separate and set aside — remove the suspect, or split into groups and see whether what's left holds. Lambert and Schindler's is confrontation — both suspects stay in the model and you see whose coefficient survives.

Both are legitimate, and they reach different conclusions.

Smoothing it over as "they weren't actually measuring the same thing" would be tidier, but false. This isn't a difference of units. It's a head-on conflict between two causal readings of the same data.

Saying "statistics can lie" is easy. The hard part is pointing at one specific case and saying which kind of error it is. In the first layer nobody was wrong. In the second the retelling was wrong. In the third, two serious research teams genuinely disagree.

---

What's left

Not the truth. A sentence far narrower than the original, but one that holds:

PwC's analysis shows that in the United States, within the most AI-exposed tier of entry-level jobs, the ones it classifies as "seniorized" grew 35% in postings between 2019 and 2025, while the rest shrank 10%.

Count the hats that sentence wears: what a posting says is not who got hired; the most exposed tier is not all entry-level jobs; the United States is not the world.

In that same report, PwC added a note to its own chart: this chart is not saying AI caused these effects; the structural characteristics of the most exposed tier, and other shocks, may also be at work. The people who produced the number narrowed the causal opening themselves first.

The much-quoted "sevenfold" deserves the same treatment. It compares the most exposed entry-level jobs against the least exposed ones — not against ordinary jobs. And that chart's sample is only Canada, Singapore, the UK and the US. The sevenfold is real; the range it governs is far smaller than the sevenfold that got passed around.

The camps, laid out:

Three pieces of evidence say it is AI. PwC — postings, ability-level exposure. Stanford — payroll, task-level exposure. And the newest: a study of 62 million résumés across 285,000 US firms finding that AI-adopting firms cut junior hiring 9–10% within six quarters, driven entirely by reduced hiring rather than layoffs, with senior hiring unaffected.

But be careful with "three pieces." What they share is only a common direction, not stackable evidence. The two that genuinely corroborate each other are payroll and résumés. PwC's measures posting text and cannot be converted into the other two — it stands on the same side, but it isn't a third vote.

On the other side, the strongest is Lambert and Schindler's, and its strength is that it's the only one here doing causal identification head-on. Three further clues point the same way: in Handshake's data, entry-level tech postings fell about 15% and healthcare about 12%, with no sign that more exposed categories fell harder; in employers' own accounts, AI ranks fifth among reasons for hiring less; and one research institute notes the hires rate fell 0.8 percentage points, with a frozen market as the main cause.

On that employer-account point the claim has to stay narrow: only 19 firms answered. Nineteen firms can't hold up a statistic. It's enough only to say that even among the dozen or so cutting back on purpose, those putting AI ahead of budget and business demand were a minority. Narrow, and therefore solid.

There's also a study formalizing the chain "the step breaks → senior talent runs dry" into a dynamic game-theoretic model. Its reasoning is complete, but it's reasoning, not observation.

All of it is locked to one variable: which grade of evidence you accept. Postings basis or payroll basis; employers' accounts or administrative data. Loosen that lock and the whole chain recalculates.

---

One number that needs no exposure measure

Every argument so far has revolved around which occupations are more exposed. One set of numbers doesn't touch that question at all.

An annual report on new graduates shows: among those who worked while in school, 82% landed a job. Among those who didn't, 41%.

By that report's account, this is the same graduating class split in two by whether they'd stepped on the first rung — and the gap is double.

It can't prove AI did anything. It shows one thing: the first step itself has weight.

As for what happens when that step gets higher, two accounts are alive at once. A radiologist wrote on a public blog that the ability to read complex scans isn't independent of volume on easy ones, and that once AI absorbs the easy ones, residents inherit the hard cases but lose the substrate that calibrates judgment. A paralegal said in a public class that whoever learns to use AI early will leapfrog everyone. Both are public posts, neither an interview, neither identity independently verified. They aren't evidence — only two accounts alive at the same time. And voices like these turned up in only three trades: legal drafting, radiology, junior programming. Three trades, not every trade.

Has there been a precedent? We looked.

America's emergency shipbuilding in the Second World War: the gap surfaced within one to two years of expansion, and in-plant training averaging about six months turned trainees into machinists. It worked — but the boundary conditions were hard. Britain's apprenticeship system went through several steep declines taking roughly one working generation; the modern apprenticeship reversed the direction but never restored the scale. Japan's successor internships in traditional crafts: 1,785 applied over three years and 63 received placements, against an apprenticeship that runs eight to ten years. For aircraft maintenance and printing, no completed precedent of successful repair was found.

So history offers neither "an inevitable break" nor "it will fill in on its own." It offers a dividing line: skills that can be decomposed and standardized can be replenished in months; judgment that only grows from long presence on the ground takes a recovery period close to its original formation period.

And that line falls precisely on a question with no answer yet — the kind of judgment that was supposed to grow slowly on the first rung: if the height of that rung has changed, where does it grow now?

History offers no precedent here. It tells us only that once questions like this appear, developing them takes a working generation.

---

Back to you

Six rulers, two directions. Plans recovering, outcomes poor. Postings rising, headcount falling. A 16% relative decline, a 6% absolute one. Not one of these numbers is false.

For China, mechanism only, not figures — because within what's publicly available there is no cross-tabulation matching that administrative payroll data by age, occupation and exposure. The official basis since 2024 splits into 16–24 and 25–29, excluding enrolled students. That change of basis is itself what shows how hard cross-country comparison is: even on who counts as young, the two sides aren't using the same ruler.

One thing can be set side by side, though — not a number, but the same sentence.

One side's postings database read out "seniorized." In the other side's platform report, "stripping out the junior tier" is the phrase they chose themselves: of new postings in the first two months of 2026, those requiring three or more years of experience made up 73.34%.

Two markets, two sets of statistics, two different phrases, describing the same shape. They can be set side by side because both are about what the hiring side wrote down as requirements — the same ruler. Their unemployment figures cannot, because those are two different rulers.

Another batch of reports is coming. It will disguise itself again as an objective report requiring no choice from you.

The version you believe — which ruler did you read it from?