Start With the Unit of Measurement: Firms, Occupations, and Tasks
To understand why AI adoption can coincide with stable total hiring and slower coder employment, start with the quantities being counted. This chapter maps the relationship among firms, job postings, industries, occupations, tasks, and projections so that each later claim can be read at its proper level. The practical payoff is simple: before deciding whether a number signals opportunity or displacement, identify who or what the number represents, what outcome it measures, and what period or comparison gives it meaning.
A labor market is not a single gauge. It is a set of linked but non-identical processes. Firms decide what work they want done and whether to recruit. Job postings record some of those recruiting intentions. Workers occupy jobs that statistical systems classify into occupations. Occupations bundle many tasks, and the same task can migrate between occupations as organizations change. Industries provide the production setting in which firms and workers interact. Projection agencies then use models and assumptions to estimate how employment may develop in the future.
The links are real, but they are not one-to-one. That is why apparently conflicting statements about AI and employment may each be accurate.
Six layers of the labor-market picture
1. A job posting is a recruiting signal, not a filled job
The Federal Reserve study of firm job-posting behavior uses the Lightcast database, which collects online postings from more than 65,000 sources and links postings to firms. This creates a large view of how recruiting signals change after firms adopt or plan to adopt AI. It is valuable for the question the study asks: are AI-exposed firms or industries reducing their overall online demand for workers?
But a posting is not the same as a hire. A firm can post the same role more than once, keep a vacancy open, withdraw it, fill it through another channel, or revise its description while recruiting. Online postings also do not necessarily describe the internal work allocation of the person eventually hired. The database is therefore closer to a record of visible recruiting intentions than a census of completed employment transitions.
This distinction changes the meaning of a stable total. If the number of postings does not fall, the evidence weighs against a broad retreat from recruiting among the firms in the sample. It does not show that every vacancy was filled, that wages were unchanged, or that each occupation received the same share of opportunities. A firm can keep its total recruiting volume constant by switching which roles it advertises.
The Federal Reserve authors also caution that job postings are not a good source for wage analysis because wage information is frequently incomplete. That limitation matters because a labor-market adjustment can appear first in pay, experience requirements, or the composition of applicants rather than in the total number of postings. The posting measure is informative, but it is one window onto hiring, not the whole building.
2. Employment counts people, not recruiting intentions
An employment series answers a different question: how many people are working in a defined occupational group at a given time? The coder study draws on the Current Population Survey and adjusts its aggregate path so that it tracks broader employment totals from the Current Employment Statistics survey. It follows a monthly series across the period before and after the November 2022 ChatGPT breakpoint.
That series can reveal a slowdown in the growth of coder employment even if firms continue to post jobs. It may reflect fewer new hires, slower replacement hiring, movement into other occupations, changes in who is classified as a coder, or some combination of these. It cannot by itself identify every individual transition. A person who no longer appears in the coder group may be unemployed, working in another occupation, working under a changed title, or still performing coding tasks that a survey classification does not capture in the same way.
The difference between a flow and a stock is useful here. Postings are a flow of recruiting signals during a period. Employment is closer to a stock of people currently working. A firm can reduce its flow of new coding hires while retaining its existing coding workforce, so the stock declines only slowly or continues to rise. Conversely, a temporary burst of postings may not produce a lasting increase in employment if vacancies are canceled or workers leave.
This is one reason a statement such as “total postings did not decline” should not be translated into “coder employment was unaffected.” The first summarizes a firm-level flow across roles; the second concerns the size and trajectory of a specific occupational stock.
3. An occupation is a bundle of work, not a single task
The word “coder” sounds more precise than it is. The Federal Reserve paper constructs a coding-intensive group using O*NET information about the importance of computer-programming skill and links that information to CPS occupation categories. It does not assume that an official job title alone tells us how much programming a worker does. The resulting group is about 3.7 percent of employment, a focused population chosen because programming is a plausible early site of generative-AI effects.
An occupation is nevertheless a bundle. A software developer may write code, design systems, test, document, meet with clients, interpret requirements, troubleshoot, and coordinate with other teams. Two people assigned the same occupational category may divide those activities differently. One may use an AI assistant to complete routine code; another may spend more time reviewing generated output, making architectural decisions, or translating a business problem into a specification.
This matters because technology often changes task composition before it changes the occupation’s headcount. If one task becomes faster, the firm may ask the same worker to perform more of another task. If a task becomes easy enough for workers in neighboring occupations to perform, its work may move without the original occupation vanishing. If a new product becomes viable, the same productivity improvement may increase the amount of work demanded. The employment effect cannot be read directly from exposure to a tool.
The distinction also explains why occupational statistics can lag organizational change. A survey category may remain stable while the content of the work shifts inside it. Or a worker may move to a different occupational category while continuing to use programming skills. A headcount series is necessary evidence, but it is not a task-level map.
4. An industry is a production context, not an occupation
Coders work across many industries. The largest concentration in the Federal Reserve paper is computer systems design and related services, but the occupation also appears in software, research, manufacturing, finance, government, and other settings. An industry can expand or contract for reasons that affect all of its workers: changes in demand, interest rates, post-pandemic consumption, financing conditions, regulation, or productivity.
If an industry shrinks, coder employment may fall even when the occupational mix inside the industry is unchanged. If an industry expands but hires proportionally fewer coders, that is a different pattern. The coder paper’s counterfactual analysis later in the series is designed to distinguish those possibilities. For now, the measurement lesson is that a sector label such as “information” is not a substitute for an occupation measure. The information sector may move differently from the broader collection of industries that employ coders.
The firm-level hiring study introduces another aggregation choice. It asks whether firms or industries with higher measured AI adoption post fewer jobs in total. A positive or null total at that level can coexist with a change in the distribution of occupations inside the firm. Aggregation is not an error; it is a question. But an answer at one level should not be used as if it answered a question at another.
The diagram’s relationships should be read as analytical links, not as a claim that each dataset contains every other entity. A posting may contain an occupation signal, but it does not provide a complete task inventory. An employment series counts workers in an occupation, but it does not reveal every posting that produced those jobs. A projection describes expected future employment for an occupation, not a realized response already visible in the monthly series.
Exposure, observation, and projection are different kinds of evidence
The coder study uses exposure measures to justify looking at coding closely. Its comparison of an early GPT-based exposure measure with the Anthropic Economic Index finds substantial disagreement about which occupations are most exposed across the economy. Yet the measures agree that coding occupations are overwhelmingly in their high-exposure groups: 99.5 percent of coding employment under one measure and 98.2 percent under the other. This agreement is useful because it reduces the chance that the coder case is chosen from one arbitrary ranking.
It still does not create a causal treatment group. “Highly exposed” means that the tasks or observed uses of AI make the occupation a plausible place to look for effects. It does not mean that every worker received the same tool, used it in the same way, or faced the same employer decision. Exposure can be a selection rule for investigation without being proof that AI caused the eventual employment change.
The evidence also has blind spots. Real-world use data from one model or service may not capture enterprise tools, custom systems, other model providers, or future capabilities. An exposure measure built before later model releases may not perfectly describe current work. The fact that coding remains highly exposed across the measures is reassuring for case selection, but it should not be mistaken for a complete measurement of actual adoption.
BLS projections are different again. The Bureau of Labor Statistics projects employment from 2024 to 2034 for selected AI- and information-technology-related occupations. It expects, for example, strong growth for data scientists, information security analysts, operations research analysts, computer and information research scientists, and software developers. It also expects employment declines in several administrative and service occupations, including customer service representatives. That pattern is consistent with a future in which technology complements some task bundles and substitutes for parts of others.
But a projection is a conditional outlook, not a retrospective finding. It incorporates assumptions about technology adoption, productivity, demand, and occupational structure. It does not tell us whether the coder slowdown observed after 2022 was caused by the same forces, nor does the software-developer projection map exactly onto the coder group constructed from O*NET and CPS data. The projection can broaden the question; it cannot settle the earlier one.
Four questions for every number
A useful reading discipline follows from these layers.
Who or what is counted? Is the unit a firm, an industry, a posting, a worker, an occupation, or a task? “AI adoption” itself may be measured through a business survey response or through AI-related language in a firm’s prior postings. Those are not identical indicators.
What outcome is measured? Is the result about postings, employment, wages, productivity, task content, or a forecast? A null result for total postings is not a null result for wages or for coding opportunities. A projected increase is not a realized hire.
Relative to what baseline? The coder paper compares post-2022 trends with earlier trends and with a counterfactual based on industry employment paths. The hiring note compares firms and industries with different levels of measured adoption while accounting for firm, state, industry, seasonal, and macroeconomic patterns. Without the comparator, a number has no direction.
Over what period? The hiring analysis covers September 2023 through November 2025 and stops after a change in the wording of the Census AI questions created a break in that series. The coder analysis uses a longer monthly path around a November 2022 breakpoint, while acknowledging that coding tools existed before ChatGPT and that employer response may have lagged. The BLS outlook covers a decade. These clocks should not be collapsed into one timeline.
The questions are especially important because the sources do not use identical populations or outcomes. Their findings can be complementary without being mechanically combined. The firm study can tell us that a broad reduction in online recruiting is not visible under its design. The coder study can tell us that a selected occupational group grew more slowly than before and relative to relevant industry paths. Neither can supply the missing task-level and worker-level transitions by itself.
The measurement map therefore changes the apparent contradiction into a set of testable relationships. If total postings remain steady while coder growth slows, look for a change in occupational mix, replacement hiring, or task allocation. If coder employment falls while industries expand, compare the occupation with the industry rather than blaming the sector automatically. If a projection shows software growth alongside administrative decline, read it as a differentiated outlook, not as proof that every technical worker benefits.
With those distinctions in place, the coder evidence can be read on its own terms. The next question is not merely whether coders were “affected.” It is how their employment trend changed after 2022, how large that change appears under different specifications, and what it means to say that growth slowed while employment continued to exist and, in the raw series, continued to grow.