The Hiring Paradox: Why AI Can Look Harmless to Firms and Painful to Coders
The subject of this series is the early U.S. labor-market response to generative AI: what happens when firms adopt AI while coding-intensive work becomes one of the technology’s clearest targets. The specific puzzle is that firms can continue to post jobs even as employment growth slows for coders. Understanding that puzzle gives us a better way to judge claims about AI and work: instead of asking whether “jobs” as one giant number are up or down, we can ask which workers, which tasks, which firms, and which time horizon the evidence describes.
That distinction matters because the most visible labor-market statistics sit at different levels. A firm may maintain its overall recruiting volume while changing the mix of roles it seeks. An occupation may grow more slowly even while the industries that employ it continue to expand. A task may become faster or cheaper without an entire occupation disappearing. And a projection about 2034 is not an observation about what happened in 2024. These are not semantic technicalities. They determine whether a worker sees opportunity, whether a manager interprets a hiring signal correctly, and whether a public claim about AI is reassuring, alarming, or simply measuring the wrong thing.
The same technology can produce different labor-market signals
A March 2026 Federal Reserve note examines firms’ job-posting behavior using online postings and measures of AI adoption. It finds no evidence, through its study window, that firms or industries with higher AI adoption posted fewer jobs in total. Some firm-level relationships were small and positive at shorter lags, while the authors treat the overall result as a precisely estimated absence of a negative total-posting effect. That does not mean every firm kept hiring, every occupation was safe, or every worker experienced an unchanged job search. It means that the outcome being measured—total online postings associated with firms—did not show the broad decline one might expect from a simple story of AI replacing labor everywhere at once.
A separate Federal Reserve paper asks a narrower question: what happened to employment in coding-intensive occupations after the public arrival of ChatGPT? It finds that coder employment growth slowed sharply relative to its earlier pace. The paper’s main estimates suggest a lower annual growth trend after November 2022, especially after accounting for the employment paths of the industries where coders work. Yet the raw coder-employment series continued to grow after the break. The finding is therefore not a simple report that “coders disappeared.” It is evidence that the occupation’s path became weaker than it had been expected to be under earlier patterns.
Those two observations can coexist. Suppose a company once planned to open ten coding roles and ten roles in sales, operations, security, or customer support. After adopting AI tools, it may plan to open six coding roles and fourteen roles elsewhere. Total postings remain twenty, but the occupational mix changes. Or the firm may keep its existing coders, use tools to increase their output, and slow replacement hiring when people leave. The total headcount can remain broadly steady while the flow of new opportunities into coding becomes less favorable.
This is a stylized example, not a claim that the Federal Reserve observed exactly this staffing plan inside each firm. Its value is analytical: it shows how an aggregate hiring measure can conceal a distributional change. The same number of openings can carry very different consequences for a new coder, an experienced software engineer, a firm trying to control costs, and a worker whose tasks are being reorganized around AI.
There is another reason not to read the two studies as competing headlines. A posting is an intention to recruit, not a person who has been hired. It can be duplicated, withdrawn, left unfilled, or aimed at a role whose tasks change before the position is filled. Employment data, by contrast, count people observed as employed, but they do not reveal every decision that led a firm to hire, not hire, reassign, or replace someone. The two measures can move at different times even when they describe the same organizations. A manager may leave an existing position open while assessing a new tool, or may post for a different role before the effect appears in occupation-level employment.
The denominators also matter. “Total postings did not decline” is a statement about the sum of recruiting signals across firms and roles. “Coder employment growth slowed” is a statement about the path of a selected occupational group. Neither statement tells us directly how many coders applied, how many offers were made, what happened to wages, or whether a worker who left coding found a closely related role. Those missing links are not minor details; they are where a broad labor-market result becomes a lived career outcome.
Finally, the time windows are not identical. The hiring analysis covers a relatively early period of measured AI adoption and is designed as ongoing monitoring. The coder study compares a longer monthly employment path around a conventional November 2022 breakpoint, while acknowledging that coding tools existed before ChatGPT and that employers may have reacted later. A short-run pause, a slow change in replacement hiring, and a long-run expansion of software demand could all appear in sequence. The opening question is therefore not only “which direction?” but also “on which clock?”
The key relationship is not “AI versus jobs” in the abstract. It is the relationship among AI adoption, the tasks firms need performed, the occupations assigned those tasks, the industries in which those firms operate, and the total amount of recruiting they undertake. A change at one level can be muted, offset, or amplified at another.
Why coding is the revealing case
Coding is a useful place to look for early effects because generative AI is already used heavily for programming-related work. The Federal Reserve coder study does not define its population by casually counting everyone with “developer” in a job title. It links occupational information from O*NET to Current Population Survey categories and uses the importance of computer-programming skill to construct a coding-intensive group. That group represents roughly 3.7 percent of employment, so it is a focused slice of the labor market rather than a proxy for all workers.
The study also compares different ways of measuring AI exposure. The measures disagree substantially about which occupations belong in the highest-exposure group across the economy. But they agree much more clearly about coding: coding employment is overwhelmingly represented among the high-exposure occupations under both the GPT-based measure and the Anthropic Economic Index measure. This makes coding a sensible early test. If generative AI is going to affect employment through the automation or augmentation of language-based technical tasks, coding is one of the places where a signal might appear before it becomes visible in less exposed work.
But exposure is not displacement. A tool can make a worker more productive, reduce the time required for a task, change who performs that task, or make a new product economical. Each possibility affects employment differently. A coder who completes a ticket faster may help a firm expand its software output; the same productivity gain may lead the firm to meet a fixed amount of demand with fewer additional hires. The direction depends partly on what happens to demand, how quickly the organization changes, and whether tasks move across occupational boundaries.
That is why the central question of this series is deliberately narrower than “Will AI take all the jobs?” We are examining a real relationship that can be measured: AI adoption and firm hiring on one side, coding-intensive employment and its industry context on the other. The evidence can tell us more about some margins than others. It can show a change in coder employment growth and a lack of a negative total-posting response during the observed periods. It cannot, by itself, tell us the complete number of jobs displaced, the long-run productivity effect, or how every worker’s tasks will be redesigned.
The judgment this evidence supports
The practical judgment is to resist two symmetrical mistakes. The first is to treat stable total postings as proof that AI has not harmed anyone. The Federal Reserve hiring note explicitly distinguishes total postings from occupations that may be particularly susceptible to automation. It leaves room for concentrated difficulty in a subset of workers even when firms shift their hiring priorities elsewhere. An aggregate can remain stable while the distribution underneath it changes.
The second mistake is to treat slower coder employment growth as proof that AI has already eliminated a fixed number of jobs. The coder paper itself cautions against that translation. Some workers may move into other occupations. Some task changes may make old occupational labels less informative. Some of the observed gap may reflect factors specific to software and technology industries, changes in tax treatment, post-pandemic normalization, or the timing of employer responses. The study’s industry controls and robustness checks make the coder-specific interpretation more credible, but they do not turn the estimate into a complete causal accounting of AI-driven job loss.
The most defensible starting point is therefore a conditional one: generative AI may be changing the occupational mix and the growth path of coding employment without causing a comparable decline in total firm recruiting. That statement is more informative than saying either “nothing is happening” or “the labor market has already been transformed.” It tells us what to examine next.
First, we need a precise measurement map. What does a job posting count that filled employment does not? What can an occupation-level series reveal about tasks inside a firm? Why should a future projection be kept separate from an observed trend? Without those distinctions, the later evidence will look contradictory even when it is answering different questions.
Then we can ask what happened to coder employment after 2022, how much of that change can be separated from an industry slowdown, and what firm-level mechanisms could connect AI productivity to hiring choices. Finally, the outlook can be placed in context: the Bureau of Labor Statistics projects strong growth for some technical occupations and declines for several other task bundles, a pattern that suggests differentiation rather than one uniform AI effect.
The organizing habit is simple: whenever a new number appears, ask who is counted, what outcome is measured, relative to which baseline, over what period, and with what uncertainty. That habit will not eliminate disagreement about AI and work. It will make the disagreement more honest—and make the apparent hiring paradox easier to understand.