What the Evidence Does—and Does Not—Say About AI and Jobs
The central question is now answerable in a precise way: how can firms continue to post jobs while coding-intensive employment grows more slowly? The evidence points to different margins of adjustment. Total online postings measure the volume of visible recruiting across roles; coder employment measures the size and trajectory of a selected occupational group; the industry counterfactual measures whether coders kept pace with the workplaces that employed them; and BLS projections describe a possible future distribution of work. Taken together, these measures support a differentiated answer, not one aggregate verdict about whether AI is “good” or “bad” for jobs.
The strongest conclusion is that stable total hiring and slower coder employment are compatible because firms can change the occupational mix, timing, and task allocation of work without reducing their total recruiting. Through the observed periods, the Federal Reserve hiring study finds no evidence of a negative total-posting response among more AI-exposed firms or industries. The Federal Reserve coder study finds that coding-intensive employment growth slowed after 2022 and lagged the path implied by the industries employing coders. Those results are not competing measurements of the same outcome. They are observations at different levels that fit a labor market in which adjustment is uneven.
The answer is a change in composition, not one aggregate collapse
Consider the two findings together. The hiring study looks across firms and asks whether AI adoption is associated with fewer total online postings. Its outcome includes every occupation represented in the postings. The coder study looks inside the occupational distribution and asks whether a high-exposure group continued along its earlier path and along the path predicted by industry employment. One can remain broadly stable while the other weakens.
A firm might keep twenty openings but reduce coding roles from ten to six and use the remaining four for security, product operations, implementation, or customer support. It might retain existing programmers while slowing replacement hiring. It might move some programming tasks to analysts or product specialists who use AI tools. It might make its coders more productive and use the savings to expand a different product. Each example changes the opportunity structure for coding without requiring a decline in total postings.
The aggregate result is therefore not a guarantee that every firm, worker, or occupation is unaffected. The Federal Reserve note explicitly says that its finding does not rule out pockets of workers having a disproportionately difficult job search. A total can remain stable while the distribution underneath it shifts. This is particularly important for people entering an occupation. Existing employees may remain in place because they hold institutional knowledge, while new entrants face fewer openings, higher experience requirements, or a narrower route into the work.
The coder result is also not a simple count of jobs destroyed. The study’s roughly three-percentage-point lower annual growth estimate is a relative trend result after controlling for industry paths. Coder employment continued to grow in the raw series, even as it fell behind the earlier trajectory and the industry-based counterfactual. The missing growth relative to that benchmark may represent slower hiring, unfilled replacement positions, movement into other occupations, changed task boundaries, or some combination. The study does not observe a fixed pool of displaced coders waiting elsewhere in the data.
That is why the most defensible synthesis is stronger than “AI has done nothing” and weaker than “AI has already eliminated a known number of jobs.” The evidence indicates an occupation-specific employment adjustment occurring alongside a non-negative total-posting response.