The current future of work

What the numbers actually show

The argument for a future of universal income rests on a sequence: AI produces enormous abundance, and that abundance is then distributed. Four measures, laid on the same timeline, show how much of that sequence has happened so far.

Two things emerge that neither side of the public argument tends to mention.

The first is that capital expenditure was flat in 2023 — roughly $162 billion in 2022 and about $157 billion in 2023, with Amazon's own capex falling from $63.6 billion to $52.7 billion. The worst year on record for layoffs happened while infrastructure spending was contracting. Those cuts were a rate shock and an overhiring correction, not an AI story. The AI buildout and the AI-attributed job losses both begin in 2024, and they move together from there.

The second is the amount of universal income paid to date.

Meanwhile the displacement is measurable but modest. Goldman Sachs measured about 16,000 net US job losses per month from AI in April 2026 — 25,000 eliminated by substitution, offset by 9,000 added through augmentation. The Yale Budget Lab found no clear upward trend in AI-task exposure among the unemployed. Unemployment sits at 4.3%.

The problem is directional. The losses arrive first, itemized and personal. The abundance arrives later, hypothetically, and to nobody in particular.

Is the overhiring claim verifiable independently?

Companies say they are correcting for pandemic overhiring. It is fair to ask whether that is true, or whether it is only true because chief executives keep saying it.

Partly yes — and the verification undercuts the conclusion drawn from it.

The hiring surge is real and sits in audited SEC filings. Amazon increased headcount by 93% between 2019 and 2022 according to SEC documents; it doubled its total workforce in 2020 and 2021 alone, and from the end of 2015 through the end of 2021 its headcount multiplied by seven. Alphabet's workforce swelled 62% during the surge. US tech job postings hit 400,000 per month in 2021, the first time that threshold had ever been crossed.

So the skepticism doesn't land on whether they overhired. It lands on whether what followed was a correction, and there the arithmetic fails.

Headcounts at all of them remain above pre-pandemic levels. And the second tell is decisive: if this were a correction, hiring would normalize once the excess was worked off. Instead, hiring at the largest tech employers is now 25% below the 2019 baseline. Headcount above 2019, hiring 25% below 2019 — those two facts together are incompatible with a correction and entirely consistent with permanently smaller teams. SignalFire's assessment is that this is "not a temporary layoff cycle that will reverse when macroeconomic winds shift, but a long-term recalibration of team sizes."

The companies have also stopped bothering with the story. Snap's chief executive said the company would cut 16% of its workforce because of AI-driven efficiencies — not overhiring, not margin pressure. Oracle's simultaneous cuts used similar language.

In fairness, the counterweight exists. The Washington Post reported in May 2026 that layoffs at Amazon, Meta and Microsoft aren't all about AI. Analyst rollups attribute roughly 25% of March 2026 cuts to AI and automation, with about 75% tracing to cost discipline and restructuring. Both things are true at once.

What is not true is that the workforce has been returned to some prior correct size. It has been reset to a new, smaller one.

If the work needed doing, how did going back to the office end it?

Here is a question that exposes the incoherence. Companies said they hired too many people for work that existed. Then they mandated a return to the office. Then people left, and the work apparently stopped needing to be done. Both justifications cannot be describing the same reality.

They aren't. The work never stopped needing to be done. Return-to-office was never a claim about the work — it was a headcount reduction mechanism, and executives have said so on the record.

A BambooHR survey of more than 1,500 US managers found that a quarter of C-suite executives hoped for voluntary turnover after implementing a return-to-office policy, and one in five HR professionals admitted the policy was meant to make staff quit. The report's own conclusion was that these mandates are layoffs in disguise.

The Federal Reserve says the same thing, which is about as far from a corporate talking point as it is possible to get. The Fed's Beige Book — built from interviews with business leaders across all twelve districts — reported that multiple districts "encouraged attrition with return-to-office mandates," and separately noted AI helping organizations silently trim headcount.

On the federal side it was stated outright. Elon Musk and Vivek Ramaswamy said requiring federal employees in offices five days a week "would result in a wave of voluntary terminations that we welcome."

Who leaves is not random. Baylor researchers tracked more than three million worker profiles across 54 large S&P 500 technology and financial firms and found that mid- and top-level managers show greater attrition than junior staff, that high-skilled employees are more likely to leave than low-skilled, and that the mandates disproportionately affect women. Upwork found 63% of C-suite leaders acknowledged their policies led to a disproportionate number of women resigning. Unispace found 42% of companies with mandates saw higher-than-expected attrition, and 29% struggled with recruitment afterward. Meanwhile 93% of chief executives say they do not go into the office full-time themselves.

Three kinds of displacement that no tracker counts

This matters beyond the hypocrisy, because it identifies a third category of invisible loss. The widely quoted figure of roughly 316,000 jobs cut by companies citing AI since 2023 excludes:

• The never-hired. Employment for workers aged 22 to 25 in the most AI-exposed occupations is shrinking at 3.8% per year and accelerating — but driven by stalled hiring rather than layoffs, which is easy to miss because nobody is being fired.
• Federal separations. Sorted into a separate tracking category. More than 317,000 federal employees left governmentwide in 2025, a 13.7% decrease, with a net reduction near 10.8%.
• The pressured-out. Attrition by return-to-office appears in no layoff tracker at all. It is recorded as a resignation.

Each of these has something in common: the person has no severance, no WARN notice, no unemployment claim tied to the cause, and no administrative trace connecting their loss to the thing that caused it. If you were designing a compensation mechanism, they are the hardest population to even identify.

Every prior technology gave us time

The reassuring version of history says we always adapt. Agriculture went from employing roughly 40% of Americans to under 2%, and the country did not collapse into permanent mass unemployment. Electricity, the telephone, the washing machine, the assembly line — all transformed work, and none ended it.

All of that is true. It also took a length of time that is not available now.

The telephone bar is the one to sit with. Alexander Graham Bell patented it in 1876; landlines did not reach 80% of American households until the 1960s. Three generations of workers had entire careers begin and end inside that transition. A switchboard operator hired in 1920 could retire before the technology that would eliminate her role finished arriving.

And the retraining protocols mostly weren't retraining

There is a common assumption that past transitions came with retraining infrastructure. There was institutional response — but it wasn't what most people picture, and the difference is the whole argument.

What actually absorbed the agricultural collapse was the high school movement. Between roughly 1910 and 1940, American secondary school enrollment went from a small minority of teenagers to the majority. Nobody retrained a forty-five-year-old farmer into a factory foreman. His children went to school longer than he had and entered a different economy. The mechanism was generational replacement dressed up as education policy.

That distinction is everything, because it means the historical success case is not available here. AI displacement is arriving inside a single career, not across three.

And where direct retraining was tried, the record is sobering. Trade Adjustment Assistance has existed since 1962 specifically to retrain workers displaced by trade — two years of paid retraining, extended unemployment benefits, roughly a billion dollars a year at its peak. It is the program people are implicitly imagining when they say we will retrain people.

The nine-year Mathematica and Social Policy Research evaluation for the Department of Labor found that in the final year of follow-up, participants earned about $3,300 less than their matched comparison group while working about the same number of weeks — 33 versus 35. Another study estimated that participating produced a wage loss roughly 10 percentage points greater than not participating. OMB rated the program ineffective in 2003. GAO found about 75% of workers who left found jobs, but many earned far less than before.

The counterweight is real and worth stating. Hyman's 2018 study of 300,000 applicants from 1990 to 2011, using a much stronger design comparing workers quasi-randomly certified against those denied, found large positive effects — roughly $50,000 in higher cumulative earnings after ten years. Wage insurance specifically appears to work well and largely pay for itself.

But read that carefully, because it makes the point rather than rebutting it.

That is the honest price and the honest timeline. The case for deliberate policy is not that AI is more powerful than electricity — electricity was arguably more transformative. It is that electricity gave us forty-five years, and we still needed the Tennessee Valley Authority and a decade of federal investment to finish the job.

Where is this going, and how would we know?

The question everyone asks is when it breaks. What year does the economy stop being able to absorb this?

There is no such year, and looking for one is the mistake. Economies do not have a cliff edge. What they have is an adjustment rate — and the rate is the whole story. The same displacement spread across forty years is absorbed quietly by retirement. Compressed into five, it is not — because a fifty-two-year-old program manager cannot be replaced by her own grandchild.

Why the headlines will be the last to know

There are roughly 160 million employed Americans, and in an ordinary month several million change jobs. Against that churn, two hundred thousand annual layoffs is a rounding error. This is exactly why the Yale Budget Lab found no clear trend, why Goldman projects the AI effect will not push unemployment more than half a point above trend, and why the unemployment rate sits at 4.3% while every chart above points the direction it does.

The unemployment rate is not lying. It is answering a different question than the one that matters.

What the data describes is not a level change — fewer jobs — but a shape change. Entry-level positions in AI-exposed occupations are shrinking 3.8% a year and accelerating. Middle management is being cut at roughly double the rate of the workforce overall; manager headcount fell 6.1% over three years against overall reductions of 3.5%, and middle management job postings dropped 42% between 2022 and 2024. Workers aged 35 to 40 are growing at 2%.

Put those together and the future state is not mass unemployment. It is an economy with roughly the same number of jobs and no ladder between them. The bottom rung is being removed while the middle rungs are being removed, and the people already standing above both are staying put. A labor market can hollow out this way for a decade without the unemployment rate moving much, because everyone who has a job still has one. The loss falls on the people who would have gotten one, and on the people who would have moved up.

And it compounds on a delay. The middle managers eliminated in 2026 are the senior leaders who do not exist in 2030. The junior hires never made in 2025 are the mid-career workers missing in 2035. The damage is scheduled, and it arrives long enough afterward that nobody will connect it to the decision that caused it.

The pattern that would signal real trouble is the middle two moving together — long-term unemployment climbing among mid-career workers while young people's entry rate falls. That is the hollowing surfacing in national statistics rather than hiding in payroll files.

But notice what that requires. Somebody has to be watching two specific series, over years, and know to read them against each other.

There is no threshold that trips, no alarm that sounds, no morning where the number crosses a line and the country notices. That is not a gap in our monitoring. It is the shape of the thing itself — a change in who gets to work, distributed slowly enough and unevenly enough that it never becomes an event. And a compensation mechanism that was always going to wait for an event before it arrived.

What other countries decided

Other wealthy countries have AI too. They are not producing displacement at the American rate. The reason is less flattering than "they are handling it better," and the distinction matters.

European employers are not sitting this out: 71% have reassessed or are reassessing job responsibilities because of AI — 79% in Italy — and more than a quarter have reduced hiring or cut jobs as a direct result. What differs is the mechanism and the speed, and that comes from law rather than virtue. Works councils hold consultation rights, and employers who implement AI-driven changes without adequate consultation face legal challenges and delays. The EU AI Act adds another layer, restricting employers' ability to make decisions about workers using AI and requiring collaboration with worker councils beforehand; those provisions take effect August 2026, and fewer than 20% of organizations describe themselves as very prepared.

The result is not fewer reductions. It is reductions that take eighteen months, get negotiated, and arrive with notice.

China went furthest. An appellate court in Hangzhou ruled that companies cannot use AI as a reason for reducing headcount — Chinese labor law permits layoffs only when businesses face circumstances beyond their control, and the court held that AI adoption does not qualify. It is the most significant legal precedent anywhere protecting jobs against AI displacement, from the country moving fastest on AI capability.

Singapore built the architecture first

Prime Minister Wong said the thing out loud that no American official has: that "there is no economic law that guarantees" new jobs will replace lost ones, and that governments "cannot leave this to the market."

Concretely: S$1 billion committed to National AI Strategy 2.0. The SkillsFuture Mid-Career Enhanced Subsidy covers up to 90% of course fees for citizens aged 40 and above. A Mid-Career Training Allowance pays up to S$3,000 per month for workers over 40 pursuing approved full-time courses. Credit top-ups of S$4,000 for the same group. From late 2026, enrollees in selected AI courses receive six months of free access to premium AI tools, on the reasoning that fluency comes from use rather than instruction. It is not aspirational — more than one in two eligible Singaporeans have used their SkillsFuture credits since 2015, with over S$1 billion claimed.

Notice the design choice. The subsidies are weighted toward people over 40. That is a deliberate answer to exactly the problem the agricultural comparison exposes: the historical mechanism was generational, and this transition does not allow for it. Singapore is attempting to buy mid-career transition directly, at 90% subsidy, because it knows the cheap generational path is closed.

Whether it works is unproven. Singapore's headline unemployment held at 2.0% in the first quarter of 2026, but long-term unemployment rose to 0.8% — small in absolute terms, and exactly the indicator listed above. That is the hollowing forming under a tight headline number, in the country trying hardest to prevent it.

And why Denmark does not transplant

Denmark spends roughly 2.5% of GDP on active labor market policies and runs tripartite training councils where employers, unions and government jointly set reskilling priorities. But that system depends on high trust between those three parties, on collective bargaining determining training priorities, on social insurance funded through high payroll taxes, and on cultural norms that legitimize large transfers for individual skill development. The United States has essentially none of those preconditions. Pointing at Denmark without saying so is not a serious argument.

Elsewhere, smaller versions: Germany has committed roughly €1 billion to AI research and skills. South Korea has limited automation tax incentives specifically to fund worker transitions. Belgium is running targeted reskilling for displaced steelworkers in Liège. Estonia is putting AI training into vocational education, reaching 38,000 students in a phase starting this year.

Japan deserves a caveat rather than credit. With an aging population and a shrinking workforce, automation there fills positions that would otherwise sit vacant. That is the generational mechanism happening by accident, and it is not replicable in a country with a growing labor force.

One more honest complication: the protection cuts both ways. Highly regulated labor markets slow AI adoption and buy workers transition time, but the same regulations make it harder for workers to move into new industries and roles. Slower firing is also slower hiring. And Europe's smaller displacement number is partly just a smaller tech sector — the buildout is overwhelmingly American, so the American labor market is where the conversion of payroll into capital expenditure is actually happening.

It is not that America has no response. It is that several countries decided this was a thing requiring a decision, and built an institution to hold it. We published a framework.

What exists for an individual right now

If you read that the Department of Labor announced $30 million in AI training grants and wondered where to sign up, the answer is that you cannot. None of that money goes to individuals.

The $30 million goes to state workforce agencies — up to $8 million each — to create funds that encourage employers to develop training. Employers then apply for partial reimbursement per employee. The related Industry-Driven Skills Training Fund awarded more than $86 million to 14 states in September 2025, and it trains "current employees or newly hired workers." A displaced worker is not in the population. The $98 million in pre-apprenticeship funding is YouthBuild, ages 16 to 24, awarded to governments and nonprofits. The $81 million RESTART initiative serves formerly incarcerated individuals.

So the answer to how this is being communicated is that it isn't, because people are not the audience. It is communicated on grants.gov to grant writers.

There is a door. It is an entirely different building.

The system that funds an individual's training is the Workforce Innovation and Opportunity Act, and it long predates any of this. It runs free through roughly 2,300 American Job Centers nationwide. Start at CareerOneStop's job center locator, or call 1-877-US-2JOBS. In Georgia it operates as WorkSource Georgia, administered through the Technical College System of Georgia, with local boards by region.

The mechanics that matter:

• You do not need to be low-income. There is a Dislocated Worker stream for anyone terminated or laid off, or holding notice of layoff. The income test applies to the Adult stream, not that one. Many people self-disqualify here and should not.
• Money moves through an Individual Training Account. An ITA is a voucher tied to your name that pays an approved school directly. It can stack with a Pell Grant.
• The program must be on your state's Eligible Training Provider List. If it is not on the list, the voucher will not pay for it, however sensible the training.
• Authorization comes first. Your local board must verify eligibility, confirm the program is active, and authorize funding before you start. Enroll and pay first and you may have no recourse.
• Veterans receive priority of service in all Labor Department funded employment programs, by statute.
• Ask about supportive services early. Transportation, childcare, tools, books and exam fees are often covered, and they are what actually derails people who have already solved tuition.

Three things to know honestly

The caps are small. Local boards set ITA limits and they commonly land in the low thousands — enough for a certification, not a career change. Training must point to a locally in-demand occupation, which is a reasonable rule producing an unreasonable result: it steers funding toward whatever is already on the list.

And that is the mismatch worth naming. The training menu is built around healthcare, skilled trades, commercial driving, manufacturing, IT support and office administration. It was designed for plant closures. The displacement documented above is hitting college-educated mid-career white-collar workers, entry-level knowledge workers, and federal program staff — people often over-credentialed for everything the voucher will pay for.

The one federal system that puts money in an individual's hands is pointed almost exactly away from the population currently being displaced. Not out of malice. Because it was built in 2014 for a different problem, and nobody has redirected it.

What to actually do

The instinct at the end of a piece like this is to say call your representative. That advice fails because it is unfalsifiable. A staffer logs "constituent concerned about AI," and nothing happens.

Be specific instead. As of this summer at least eight federal bills address this, and several are bipartisan. None has moved out of committee. That is worth knowing before you call — and worth saying to the office, because a bill sitting in committee is a bill that still responds to constituent pressure.

The four to lead with:

• S.3339 — AI Workforce PREPARE ActSen. Jim Banks (R-IN), introduced 12/03/2025, with Hassan, Hickenlooper and Husted. Senate HELP Committee. Would require employers to disclose when AI was a substantial factor in a mass layoff, specifying the type of AI and the estimated share of job loss attributable to it. Would also use AI-informed projections to update the state in-demand occupation lists that determine what federal retraining money can pay for — the direct fix for the mismatch described above.
• S.3108 — AI-Related Job Impacts Clarity ActSen. Josh Hawley (R-MO) with Sen. Mark Warner (D-VA), introduced 11/05/2025. Senate HELP Committee. Would require major companies and federal agencies to disclose AI-related layoffs, hiring, unfilled positions and retraining quarterly, by industry code, with the Labor Department publishing the data. Solves the counting problem.
• National Workforce Transition Fund — Sen. WarnerIntroduced 07/21/2026 as part of "A Framework for America's AI Future," a package of separate bills spanning data centers, model safety, workforce and national security. Would limit bonus depreciation on AI data center infrastructure and dedicate the revenue to worker retraining, portable individual training accounts, employer redeployment grants and tuition assistance — governed by a board split evenly between employer, labor and government representatives, the tripartite structure Denmark uses. The mechanism taxes the buildout to pay for the workers the buildout displaces. Too new to have a congress.gov number at the time of writing.
• H.R. 7576 — AI Workforce Training ActRep. Josh Gottheimer (D-NJ-5) with Rep. Mike Lawler (R-NY), introduced 02/13/2026. House Ways and Means. Note what this one actually is: a tax credit amending the Internal Revenue Code, which means it subsidizes employers who train staff. It puts nothing directly in the hands of a displaced worker — the same pattern as the $30 million in Labor Department grants. Worth supporting, worth not mistaking for individual assistance.

Four more worth naming if the conversation goes further:

• Economy of the Future Commission ActSens. Warner and Mike Rounds (R-SD), March 2026. Bipartisan. Would convene policymakers and industry experts to identify concrete steps on workforce training and worker support.
• Workforce of the Future ActReps. Larson, Cleaver and McIver, December 2025, with a Senate companion from Blunt Rochester, Hirono and Schiff.
• Investing in American Workers ActSen. Warner. Aimed at tax incentives, on his own reasoning that the tax code "often makes it easier for companies to invest in machines than in people." That is the argument of this entire briefing, stated by a sitting senator.
• Data Center Tax Accountability and Disclosure ActAlso from the July 2026 framework. Would require large AI data centers to publicly disclose energy use, water use, emissions and backup generation. The bill to raise at a county zoning hearing.

That second ask is the one nobody is making. Every bill on the list is oriented toward measurement or general training funds. None does what Singapore did, which was to look at who is actually being displaced — mid-career, mortgaged, forty-plus, too credentialed for the voucher and too young to retire — and design for that person specifically.

Congressional offices count contacts by bill number. A hundred constituents asking about S.3339 by name is a data point that moves a vote. A hundred constituents concerned about AI is a folder.

If you are in a state with a data center fight — Georgia, Virginia, Texas, Ohio, Arizona — you have leverage the rest of the country does not. Your county commission is already negotiating with these companies over water, power and tax abatements. Active opposition groups more than doubled in a single quarter, from 396 at the end of 2025 to 833 across 49 states by March 2026, and roughly $130 billion in projects stalled in the first quarter of 2026 alone. The Warner proposal says out loud what those negotiations already imply: the buildout can pay for the transition. That argument works at a zoning hearing as well as it works in Washington.

The gap to flag while the language is still soft

There is one problem with these bills that almost nobody is raising, and it could hollow out the fund before a dollar moves.

Money scoped to workers "displaced by AI" requires somebody to decide who qualifies. Run the findings in this briefing through that filter:

• Cut to protect margin against a depreciation schedule. The employer records it as restructuring. Not AI.
• Never hired, because the role was never posted. No job was lost. Not displaced at all.
• Pushed out by a return-to-office mandate. Recorded as a resignation. Voluntary.
• Separated in a federal reduction executed with AI tooling. A policy decision. Not AI.

None of them clearly qualifies. And under S.3339 the determination rests with the employer, who self-attests whether AI was a substantial factor — a party with every incentive to say no, since saying yes invites scrutiny it gains nothing from.

So a fund scoped to AI displacement, gated by employer attestation, would systematically exclude nearly everyone this briefing documents. It is the same counting gap that leaves federal separations, stalled hiring and pressured resignations out of every published layoff total — except written into law, where it decides who gets paid.

The fix is not complicated. Eligibility could attach to displacement rather than to cause — anyone separated, not backfilled, or unable to re-enter after a qualifying period — with attribution used for data collection rather than as a gate on assistance. Tie the money to what happened to the worker, not to what the employer says caused it.

The Warner fund was introduced in July 2026 and sits in committee. Bill text is most malleable in exactly this window. A specific, well-argued comment sent to that office and to the Senate HELP Committee now is worth more than a hundred general expressions of concern later.

What you will hear back

Four objections come up, and they are not made in bad faith. Better to be ready.

"Retraining programs don't work."

The strongest objection, and the Trade Adjustment Assistance record supports it. The honest answer is that the average program performed badly, but the best-identified study found large gains after ten years at roughly $20,000 per person per year, and wage insurance specifically appears to pay for itself. The counter is not that retraining works. It is that we know what expensive retraining costs and what it returns, and we have never funded it at that level.

"Singapore is six million people."

True, and per-capita translation is daunting. Point at the denominator on the other side: $725 billion in a single year of hyperscaler capital expenditure. The question is not whether the country can afford it. It is whether it will decide to.

"Workforce training is a state function."

Also true — the system runs through states, and this cuts against federal mandates. But the occupation lists that constrain state spending are shaped by federal data and federal guidance, which is precisely what S.3339 targets.

"Government can't pick which jobs will be in demand."

Fair, and the record supports the skepticism. But government is already picking — the Eligible Training Provider List is a list of government-approved occupations, written in 2014. The choice is not between picking and not picking. It is between an old list and a current one.

Appendix: letters you can send

Send through the office's own web form rather than by email. Most congressional offices filter direct email from outside the district; the web form runs an address check that marks you as a verified constituent. Physical mail to Washington is still irradiated and delayed — use the district office if you mail. Verify bill numbers at congress.gov before sending, since new bills receive numbers on introduction.

To a senator

To a representative

Two-minute call script

One customized sentence beats three paragraphs of argument. The bracketed personal line is the part a staffer actually reads and quotes upward. Everything else establishes that the writer knows what they are asking for; that line is why anyone cares.