AI Isn't Taking the Jobs. It's Closing the Door.

The Setup

There are two stories about AI and employment, and they cannot both be right. The first is the apocalypse: AI is a general-purpose replacement for cognitive labor, white-collar work is being automated away, and the unemployment line is coming for everyone with a desk job. The second is the dismissal: this is a normal hiring slowdown dressed up in AI language, the same productivity panic that greeted the spreadsheet and the search engine, and it will pass.

The most honest signal that neither camp has the data on its side is that the loudest voices keep switching sides. In June 2025, Sam Altman warned that entry-level white-collar jobs were at serious risk and described customer service roles as headed for total elimination. By May 2026, speaking in Sydney, he said he had been "pretty wrong" and was "delighted to be wrong" about how quickly AI would reshape the job market, conceding that the impact on entry-level roles was smaller than he had expected. Dario Amodei, the chief executive of Anthropic, the company whose Claude AI powers parts of this platform, had predicted that AI could eliminate up to half of all entry-level white-collar jobs within five years and push unemployment to 10% to 20%. He has since softened the framing toward AI expanding the work people do. When the people building the technology cannot keep their own forecast steady for twelve months, the forecast is not where you should be looking.

So this piece does not argue from prophecy. It argues from the data that exists right now, in the spring of 2026, across government statistics, payroll records, and university research. And the data tells a story that is more specific, and more useful, than either camp wants to hear. The apocalypse is not happening to the workforce as a whole. It is happening to the people trying to enter it.

The Apocalypse That Hasn't Arrived

Start with the claim that AI is causing mass unemployment, because the aggregate data simply does not support it. The overall unemployment rate sits around 4.2%, a figure that is unremarkable by historical standards and a long way from anything resembling a crisis. If a general-purpose replacement for cognitive labor had truly arrived, this is the number where it would show up first, and it has not.

The most rigorous look at the question comes from the Yale Budget Lab, which used the Bureau of Labor Statistics Current Population Survey to track whether workers in highly AI-exposed occupations were faring worse than everyone else. Through March 2026, it found no meaningful change in the occupational mix or the length of unemployment for high-exposure workers, and concluded that AI was likely not the driver of the labor market's recent softening. In plain terms: the people whose jobs are supposedly most threatened are not, in the data so far, losing those jobs at an unusual rate.

The Stanford Digital Economy Lab found something even more pointed. Among experienced workers, those aged 30 and older, in the fields most exposed to AI, employment actually grew between 6% and 12% from late 2022 through mid-2025. The senior software engineer, the experienced analyst, the veteran marketer: their employment held up or improved across exactly the period when AI capability exploded. Whatever AI is doing to the labor market, it is not clearing out the experienced ranks.

The dismissal camp has a serious historical argument to stand on, too. Goldman Sachs chief executive David Solomon has pointed out that civilian employment in the United States has grown roughly 145% since 1962 despite wave after wave of automation, and that data center construction alone has added around 200,000 jobs since 2022. The underlying economics were described by the economist Daron Acemoglu years ago: automation displaces some labor, but the productivity it creates tends to generate new demand for labor elsewhere. The lump-of-labor assumption, the idea that there is a fixed amount of work to go around, has been wrong every previous time, and betting against it has a poor track record.

If you stopped reading the data here, you would conclude the hype camp is right. That would be a mistake, because the aggregate numbers are hiding something the averages cannot see.

The Canaries Nobody in the C-Suite Mentions

Disaggregate the labor market by age and the comfortable picture falls apart. The damage is real, it is measurable, and it is concentrated almost entirely on the people at the very start of their careers.

The headline number is the one that has not happened in living memory. The unemployment rate for recent college graduates now sits around 5.6%, well above the 4.2% rate for the workforce as a whole. Four years ago those two rates were nearly identical. Bank of America research notes that recent-graduate unemployment exceeding the national rate is a reversal with no recent precedent: a college degree has historically been a shield against unemployment, not a liability. For computer science graduates specifically, the unemployment rate has reached 6.1%, and for computer engineering it is 7.5%, among the highest of any major.

The most cited research here is the Stanford Digital Economy Lab's paper bluntly titled "Canaries in the Coal Mine?" Working with ADP payroll data covering millions of workers, the researchers found that workers aged 22 to 25 in the occupations most exposed to AI, software developers and customer service representatives chief among them, saw a roughly 13% relative decline in employment compared with less-exposed workers. Entry-level software developer employment for the youngest cohort fell close to 20% from 2022 levels. Note the precision of that finding: the decline lands on young workers in exposed fields, not on older workers in the same fields, and not on young workers in unexposed fields. That is not what a broad economic slowdown looks like. A recession hits everyone. This hits the door.

The hiring data points the same direction. The venture firm SignalFire found that big technology companies cut their hiring of new graduates to roughly half of pre-pandemic levels, with new-grad hiring down about 25% in a single year. The share of technology job postings open to candidates with three years of experience or less fell from 43% in 2018 to 28% by 2024. Aneesh Raman, LinkedIn's chief economic opportunity officer, put it in a phrase that has stuck: the bottom rung of the career ladder is breaking. Coding assistants are absorbing the simple-code and debugging tasks that junior developers used to learn on, and the same pattern is repeating in research, paralegal work, and first-tier support.

This is the evidence the dismissal camp has to explain away, and it cannot. Something specific is happening to entry-level hiring, it is happening in exactly the fields AI touches most, and it is happening to exactly the cohort whose work overlaps most with what the tools now do well.

Why the Bottom Rung Breaks First

Here is where I can speak from the keyboard rather than the spreadsheet. I use these tools every day to build and run this platform, and the productivity gain is not a slide in a vendor deck. It is real, and it is large. But it is worth being precise about what, exactly, the productivity replaces, because that precision is the entire explanation for the data above.

When AI makes me dramatically faster, it is not doing the judgment work. It is doing the first draft of a function, the boilerplate, the debugging pass, the initial literature scan, the summary of a long document, the routine query I would otherwise have farmed out. Notice what every item on that list has in common. None of them is senior work. They are the exact tasks a junior hire used to be given precisely because they were learnable, low-stakes, and a way to build toward the harder judgment that comes later. The on-ramp was made of those tasks.

That is why the effect shows up as a hiring freeze for juniors rather than a firing wave for seniors. A company does not need to lay off its experienced engineers to capture the savings. It simply stops opening the entry-level requisition it used to open, because the senior engineer plus AI now covers the work that a new graduate would have done in their first eighteen months. The experienced worker becomes more valuable, because their judgment now commands a fleet of automated assistants. The inexperienced worker becomes harder to justify, because the thing they offered, cheap capacity for routine cognitive tasks, is the thing that just got cheaper still.

This is the mechanism that reconciles the two halves of the data. Experienced employment in AI-exposed fields grows because AI is a complement to judgment. Entry-level employment in the same fields shrinks because AI is a substitute for the tasks that used to be the apprenticeship. The same tool does both things at once, depending entirely on where the worker sits on the ladder. AI is not taking the jobs. It is removing the bottom rungs people used to climb to reach them.

The $700 Billion Tell

There is a second force at work in the 2026 layoff numbers, and it has very little to do with AI productivity and almost everything to do with accounting. The defining feature of this year's tech labor market is not the layoffs themselves. It is that they are happening at the same time as record profits and record capital spending.

The four largest hyperscalers, Amazon, Microsoft, Alphabet, and Meta, have committed to a combined roughly $700 billion in capital expenditure for 2026, nearly double what they spent in 2025. To put that in context, Goldman Sachs notes that a number that size would rival the peak of the late-1990s telecom buildout, even though AI capex still equals only about 0.8% of GDP, below the 1.5%-plus peaks of past technology booms. This is the largest concentrated infrastructure spend in the history of the sector, and it has to be paid for out of the same income statements that report the layoffs.

Meta is the cleanest illustration. The company cut roughly 8,000 jobs in a round that freed an estimated $2.4 billion in annual operating expense. Against a 2026 capital budget in the range of $125 billion to $145 billion, that saving is about 1.7% of the capex bill. And this is not a company in any distress: it printed $56.3 billionin quarterly revenue, up 33%, at 41% operating margins. The layoffs were not a response to AI making 8,000 workers redundant. They were a way to create a little room on the income statement for an infrastructure bill growing roughly twice as fast as revenue.

This matters because "AI efficiency" has become the most convenient line in a layoff memo. It sounds forward-looking rather than defensive. It tells investors the cuts are strategic rather than a sign of slowing growth. But when the savings from the layoffs are a rounding error against the capex, the honest description is not that AI replaced the workers. It is that the company needed to fund the AI buildout, and headcount was the easiest line to trim. A meaningful share of the 2026 layoffs are a post-2021 over-hiring correction and a capex-financing exercise wearing an AI costume.

What the Data Actually Says

Put the pieces together and the answer to the original question comes into focus. Are the layoffs real? Yes, unambiguously: tech layoffs in 2026 have run into the six figures, and a growing list of companies cite AI directly. Is AI causing mass unemployment? No, at least not in the aggregate data we can currently measure. The overall rate is steady, the experienced workforce is intact, and the most careful study available finds no broad AI displacement effect through early 2026.

The truth sits in the gap between those two facts. AI is not firing the existing workforce. It is quietly closing the door behind it, eliminating the entry-level positions that served as the on-ramp into white-collar careers. The measurable harm is concentrated on the young and the new, and it takes the form of jobs that never get posted rather than workers who get dismissed. That is a slower, quieter crisis than the apocalypse narrative, and in some ways a more serious one, because a workforce that stops training its next generation does not feel the damage until years later, when the experienced workers it failed to grow are the ones it suddenly needs.

It is worth being honest about what the data cannot yet prove. The Stanford researchers themselves caution against reading too much into trends that are still small and noisy, and one of the lead authors has declined to state flatly that AI is the cause, noting only that the pattern is consistent with it after other explanations were ruled out. The New York Fed, for its part, attributes a large share of recent-graduate unemployment to the normalization of remote work, which lets employers fill desk jobs with experienced people anywhere rather than training locals. Causation here is genuinely entangled. But the directional signal, young AI-exposed workers faring distinctly worse than everyone else, is consistent across government data, bank research, and payroll records. That is about as much agreement as labor economics ever produces in real time.

What Could Go Wrong

This thesis can be wrong in both directions, and an honest read of the data has to hold space for each.

The apocalypse could simply be slow. Altman walking back his timeline is not the same as the threat disappearing. His own organization's internal planning reportedly assumes significant labor disruption is still coming; he was conceding the pace, not the direction. There is a credible argument that 2024 and 2025 were the investment phase and that the harvest phase, where the capability actually shows up as displaced work, is only beginning. If that is right, the entry-level damage visible now is the leading edge of a wave that reaches the experienced ranks next, and the steady aggregate numbers are a lagging indicator rather than reassurance.

Or the bottom rung could be a cyclical illusion. The opposite case is just as plausible. Entry-level hiring collapsed at the same moment interest rates rose and the post-pandemic over-hiring unwound. Junior roles are always the first cut in any slowdown, because they are the cheapest to eliminate and the least immediately productive. On this reading AI is a convenient scapegoat for an ordinary cyclical contraction, and when rates fall and growth resumes, the entry-level door reopens and the canaries turn out to have been reacting to the weather, not the mine. The remote-work explanation pulls in this direction too.

And the historical base rate favors the optimists. Every prior wave of automation, from the spreadsheet to the ATM to the search engine, was forecast to destroy jobs and instead changed their composition while total employment kept climbing. Acemoglu's point about productivity generating offsetting demand has held for two centuries. Betting that this time is fundamentally different is betting against the single most reliable pattern in labor economics, and that bet has bankrupted a lot of confident forecasters.

The thesis here threads these risks rather than dismissing them: the entry-level signal is real and measurable today, the aggregate apocalypse is not yet visible, and which of those facts turns out to matter more is the genuine open question. Anyone who tells you they know for certain is selling a narrative, not reading the data.

The Bottom Line

The apocalypse-versus-hype debate is a narrative war, and like most narrative wars it generates more heat than light. Both sides are arguing about a single number, the aggregate unemployment rate, that is the wrong place to look. The aggregate is steady. The distribution is not.

What the data actually says is narrow and specific. AI is behaving as a complement to experienced judgment and a substitute for entry-level tasks, and that split is showing up exactly where the mechanism predicts: experienced employment in exposed fields holding or growing, young employment in the same fields falling, and the on-ramp into white-collar work narrowing for the people who need it most. Layered on top is a capex-financing squeeze that is dressing ordinary cost discipline in AI language because the language plays better with investors than the truth does.

For an investor, the useful takeaways are not in the doom headlines. They are structural. The companies funding a $700 billion infrastructure race by trimming labor are telling you something about where they expect returns and how patient they can afford to be. The firms learning to grow revenue while holding headcount flat, the same margin-and-headcount decoupling we traced in Every Company Is a Software Company Now, are quietly demonstrating the productivity the labor statistics have not caught up to yet.

At Wealth Engine Pro, we follow the numbers, not the narrative. The narrative says AI is taking the jobs. The data says it is closing a door, slowly, on one specific group, while the rest of the workforce so far holds steady. That is a less dramatic headline and a far more accurate one, and accuracy is the only thing worth building an investment process on.