AI Has an Alibi

Lay the two facts that built that headline side by side, and the verdict writes itself. The share of new hires going to early-career workers has fallen sharply across developed economies since 2022. The drop concentrates in exactly the occupations most exposed to generative AI. A growing research literature points the same way, and every executive who has quietly frozen entry-level headcount this year has felt the logic from the inside. Motive and a body. What else do you need?

A timeline, it turns out, and an honest look at who else was in the room. Because the tidy verdict is, on the evidence, the least defensible of the available explanations. The real version is not “it was AI all along.” It is a lineup, and more than one suspect had a hand in it.

Generative AI was not the only thing that happened to knowledge work after 2020. Remote work also happened, and it affected almost exactly the same jobs.

That is not a loose impression. In a recent working paper, Peter John Lambert and Yannick Schindler measure it directly: the occupations most exposed to generative AI and those most exposed to working from home show a strong correlation (0.77) across more than 680 narrowly defined occupations. Software developers, technical writers, and management consultants sit near the top of both lists. The jobs an AI can most help with are, to a striking degree, the same jobs you can most easily do from your kitchen table.

Once two shocks travel together that closely, any study that looks at only one of them is in trouble. Credit the hiring slump to token predictors while ignoring the disappearing commute, and you are quietly crediting AI with remote work’s effects too, because the data cannot tell them apart. Statisticians call this omitted-variable bias. Leaders know it by a simpler name: blaming whichever suspect you happened to be looking at when the lights came on.

So Lambert and Schindler do the thing the headline studies skip. They put both suspects in the same room, across 243 million new hires and 407 million job postings in the United States, the United Kingdom, Canada, and Australia from 2017 to 2025. Separately, each shock looks guilty: a meaningful jump in either AI or remote-work exposure predicts a fall of around 5 percentage points in the early-career share of new hires by 2025. Estimated jointly, the picture changes. Remote work holds up across every stress test they run. The AI effect shrinks sharply and, in many specifications, becomes statistically indistinguishable from zero.

A careful word, because the temptation is to over-correct the other way. This does not prove that AI has no effect on first-time hiring. It says the statistical case for that effect is far thinner than the headline implies, once you control for the thing standing right next to it. The honest verdict is not “AI is innocent.” It is “the case against AI is more circumstantial than anyone selling the headline has let on.”

Here is what the regression cannot tell you, and what should give every confident headline pause. Walk the timeline.

The sharp decline in hiring for young and upcoming talent sets in around late 2022, with no clear downward trend before it. Yet ChatGPT only arrived in November 2022. For most of 2023 and well into 2024, the tools were impressive party tricks and useful drafting aids, and also confidently wrong often enough that no sane manager would hand them a new hire’s actual workload. The agentic tools that can take a real task and run with it are a 2025 story: Claude Code landed as a research preview in February 2025 and reached general availability only that May. One more soft piece of evidence: “agents” and “agentic AI,” the very capabilities now blamed for the missing jobs, are not covered in Chip Huyen’s AI Engineering, published in early 2025, a serious and current textbook on building with this technology (this also shows how quickly things are evolving in the tech space). The vocabulary of replacement is newer than the slump it is meant to explain.

Yes, ChatGPT and the slump arrive in the same season, and that coincidence is most of why the AI story feels obvious. Coincidence in timing is not capability, though. A cause cannot precede its effect. For LLMs to have driven a decline that set in by late 2022, they would have needed a capability in 2022 and 2023 that they demonstrably did not have until 2025 (I would even debate if they have it now, but that's another topic). Something else was the primary culprit, while the tool that gets the blame was still learning to count the letters in “strawberry.”

With AI alibied out, two suspects remain, both with the timing AI lacks. The first is a hangover. Through 2020 and 2021, with money cheap and everyone suddenly remote, technology firms hired as if the conditions were permanent, fighting hardest for experienced people and padding headcount across the board. When rates rose and the growth math changed in 2022, the correction landed first on the most discretionary line in the plan: the fresh-out-of-school hire who costs more than they produce on day one. That is not AI displacing labor. That is a balance sheet unhiring the optimism of two years earlier, and the timing fits exactly, which is more than the accused can say.

The second remaining suspect is the one the data actually convicts, and it is the most interesting of the lineup, because it is not about cost at all. It is about learning.

A correlation that survives controls still needs a mechanism, or it is just a tidier coincidence. The mechanism is where this stops being an economics paper and becomes an operating problem you can do something about.

A firm hires a graduate who, on day one, rarely produces more than they cost. You hire them anyway because you are not buying today’s output. You are buying the experienced colleague they will become, and that bet pays off only if they climb the curve quickly once they are inside. A first-time hire is an investment in future surplus, and the return depends entirely on the rate of on-the-job learning.

So how does that learning happen? A second study, by Natalia Emanuel, Emma Harrington, and Amanda Pallais, observed software engineers at a large firm and measured it down to the level of individual code comments. Sitting near your teammates raises the feedback you receive by 18.3 percent, and the gain is uneven: younger, less-tenured engineers benefit about twice as much as the average. When the offices emptied, that advantage went with them, and the people who lost the most were exactly the ones who needed the feedback most. At the national scale, the same paper finds the fingerprint: in remoteable jobs, young graduates saw their unemployment rise relative to older graduates after the pandemic; in non-remotable jobs, no such gap appeared.

Proximity, it turns out, was quietly underwriting the entire apprenticeship system, and almost nobody had it on a line item. Distributed work raised the cost of developing new talent, and when something gets more expensive, rational firms buy less of it. Fewer first-time hires are not proof that the work has vanished. It is also a sign that the apprenticeship that made those hires worth the bet has become harder and more expensive to deliver.

If AI were truly carving out the bottom of the labor market, you would expect it in the aggregate numbers by now. Mostly, you do not. Apollo’s chief economist, Torsten Sløk, reading high-frequency US employment data, finds no clear AI job-loss signal and argues the buildout is, on net, pulling demand for workers, chips, energy, and the people who install all of it. Cheaper capability expanding total activity rather than shrinking it is the Jevons paradox: making a resource more efficient tends to increase how much of it we use, not less. The aggregate story so far is closer to “AI is creating a great deal of expensive new work” than to “AI is destroying jobs.”

So why does AI get the blame anyway? Because it is the convenient answer for a business that overhired in 2021 and would rather not admit it. “We’re restructuring around AI” is a more flattering line in an earnings call than “we misjudged the cycle.” The bot makes a better story than the binge. And the story has decades of runway: we were handed the machine-takes-over plot long before computers could write a coherent paragraph, from the Matrix to the Terminator to the AI tasked with making paperclips. Vivid, and built on assumptions about intent and autonomy that today’s systems do not possess. A model that cannot reliably book a dinner reservation without supervision is not about to repurpose the biosphere. When a mundane structural shift arrives alongside a technology we have been primed for forty years to fear, the fear collects the credit. The mask fits because we made it long ago and kept it handy. Which does not mean it hides nothing.

Here I want to be precise, because the easy misreading of all this is “see, AI changes nothing.” That is not the argument, and it would be a foolish one to make.

AI will reshape work, and in places it will genuinely remove it. The rules-based, high-volume, low-judgment back-office processing that was offshored in the 2000s is now squarely in range, and that displacement is real and worth honestly planning for. Closer to home, the shift is already visible in early form: the strongest software developers increasingly do not write code line by line so much as direct the agents that write it, reviewing, correcting, orchestrating. The role did not disappear. It moved up a level of abstraction.

That is the shape of it. Not humans pushed out of the equation, but the equation rebalanced. The most powerful configuration available now, and for some time yet, is human plus AI, not AI alone: human intelligence supplying judgment, context, and direction, AI supplying knowledge at scale and tireless execution. The deployments that impress keep people in the loop and let machines extend their reach. The ones that embarrass their owners removed the human to save a salary, then discovered what the human had quietly been holding together.

So if first-time hiring in your organization has quietly contracted, the useful question is not “what is our AI policy?” It is “what happened to our apprenticeship layer?” When teams went distributed, the visible costs (real estate, commutes, coordination) got measured and managed. The invisible one, the ambient feedback that turns a graduate into a senior, did not, because it never appeared on any budget. It simply stopped happening, a comment at a time.

Rebuilding it is not a return-to-office mandate, which manages attendance rather than learning. It is a design problem: who owns the development of each new hire, how feedback gets manufactured on purpose when it no longer arrives by accident, whether proximity is engineered in concentrated doses or replaced with a substitute that actually works rather than simply mourned. The firms that solve this will hire and grow the early-career talent that their competitors have decided they can no longer afford, compounding a decade-long advantage in the cost of building people.

The ladder did not break because the rungs were automated. It broke because we stopped standing close enough to hold it, and then blamed the most futuristic thing in the room.

A note on method: I have not read every underlying paper cover to cover; I read the abstracts, the key tables, and the methods that carry the claims, and my AI assistant helped me pull the load-bearing figures out of the full PDFs so I could check them against each other.

Emanuel, Natalia, Emma Harrington, and Amanda Pallais. “The Power of Proximity to Coworkers: Training for Tomorrow or Productivity Today?” NBER Working Paper No. 31880, November 2023, revised June 2026. https://www.nber.org/papers/w31880

Huyen, Chip. AI Engineering: Building Applications with Foundation Models. O’Reilly Media, 2025.

Lambert, Peter John, and Yannick Schindler. “The Broken Ladder: AI, Remote Work, and Early-Career Hiring.” SSRN working paper, May 2026. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6787638 Accessed 2026-06-06

“Introducing Claude Code.” Anthropic, February 24, 2025. https://www.anthropic.com/news/claude-3-7-sonnet Accessed 2026-06-06

Sløk, Torsten. “Zero Evidence of AI-Related Job Losses.” The Daily Spark, Apollo, May 29, 2026. https://www.apollo.com/wealth/the-daily-spark/zero-evidence-of-ai-related-job-losses Accessed 2026-06-06