What Happened to Coder Employment After 2022?
The Federal Reserve coder study examines a specific question: did employment growth in coding-intensive occupations change after the public arrival of ChatGPT in November 2022? The answer is yes, in the sense that the growth path became materially slower, but the answer is not that coder employment simply collapsed. Reading the result correctly requires keeping three things together: how “coder” is constructed, what a change in a growth rate means, and how much of the post-2022 divergence can reasonably be attributed to generative AI rather than other changes around the same time.
That combination matters beyond this occupation. A headline about a highly exposed group can be rhetorically converted into a claim about all software workers, all technical work, or the entire labor market. The study supports none of those shortcuts. It offers a focused test of an occupation where an effect might appear early, then measures whether the trajectory changed relative to a defined past and, later, relative to the industries employing those workers.
A “coder” is a measured group, not a job-title tally
The paper begins with a practical problem. Official occupation categories do not always line up neatly with the work people perform, and a title such as “software developer” can cover design, coordination, testing, documentation, and hands-on programming in different proportions. Counting titles alone would make it difficult to know whether the measured group actually performs the kind of work that language models can assist or automate.
The researchers therefore use O*NET, a Department of Labor-sponsored database that records occupational skills, tasks, and work activities. They focus on the importance of computer programming as a skill, link the information to the occupation categories used in the Current Population Survey, and construct a coding-intensive group. A programming-importance threshold of 2.76 yields a group representing about 3.7 percent of total employment. The group includes obvious programming occupations but also other roles in which programming is sufficiently important. It is designed to be broad enough for a usable employment series and focused enough to reveal a possible occupation-specific change.
This construction improves the question, but it does not make the category perfect. The paper treats software developers as especially important because they are numerous and likely to perform substantial programming work, even though task descriptions alone can understate that activity. The researchers examine self-reported duties and adjust their treatment of the occupation. They also test stricter and looser definitions and a version that excludes software developers. Those choices are not mere technical footnotes: if the estimated slowdown vanished whenever the boundary of “coder” moved slightly, confidence in the pattern would be weaker.
The exposure evidence gives the group a reason to study. The paper compares two different measures of generative-AI exposure: one based on GPT assessments of occupational tasks and another based on observed Claude use in the Anthropic Economic Index. The measures disagree substantially about which occupations belong in the highest-exposure group across the economy. Yet they agree that coding is highly exposed: 99.5 percent of coding employment falls into the high-exposure group under the GPT-based measure and 98.2 percent under the Anthropic measure.
That agreement is useful, but it must be interpreted narrowly. High exposure means that coding tasks are plausibly affected or already represented in observed AI use. It does not mean that every coder uses an AI assistant, that every employer has adopted one, or that exposure mechanically predicts job loss. It is a reason to inspect the employment record, not a treatment assignment that settles causality in advance.
The distinction also protects against a common category error. A task can be highly exposed while the occupation that contains it grows, because productivity can lower the cost of output and expand demand. Conversely, an occupation can weaken for reasons unrelated to AI even if its tasks are highly exposed. The employment data must do the work of showing whether a change in the occupation’s path actually occurred.
The post-2022 pattern is slower growth, not a simple collapse
The study treats November 2022, when ChatGPT was released publicly, as a natural breakpoint for an event-study-style comparison. The authors acknowledge that this date is not the beginning of all coding-related AI: GPT-3.5 tools and GitHub Copilot were available earlier, and firms might not change staffing policies immediately after a product launch. It is nevertheless a useful marker for the period when many managers and business leaders became aware of the productivity implications of large language models.
Before the breakpoint, coder employment in the all-industry sample grew at roughly 4.8 percent annually. The baseline regression estimates a post-break trend change of about 3.9 percentage points in the negative direction. A reader should not subtract those figures and announce a 3.9 percent fall in the number of coders. They describe a change in the slope of the employment path. A group that had been growing quickly can continue to grow, but at a much slower pace.
The raw series has that character. The paper reports that coder employment continued to grow after ChatGPT, even though it moved away from the faster pre-ChatGPT trajectory. In coder-intensive industries, where roughly ten percent or more of workers are coders, the change is more pronounced: pre-break growth was about 6 percent annually and became flat or negative in the post-break specifications. Non-coder-intensive industries show smaller changes. The contrast makes the result more specific than a claim that every part of the economy experienced the same employment shock.
The paper’s preferred counterfactual-controlled results place the occupation-specific post-GPT trend change around 3 percentage points per year across the main specifications. This is the important magnitude for the article’s central question: coder employment did not merely follow the overall path of the industries in which coders worked. It lagged that path. But the estimate remains a change in relative growth, not a direct count of jobs eliminated by AI.
The difference between those descriptions can be made concrete without pretending to know the unobserved alternative history. Imagine an occupation with 1,000,000 workers that had been growing at about 5 percent a year. Under that earlier trend, it might have added approximately 50,000 workers in a year. If its growth rate later slowed to about 2 percent, employment would still rise by approximately 20,000 workers, but the gap from the earlier path would be large. The workers who do not appear in that counterfactual gap are not automatically unemployed: some may enter another occupation, some firms may produce more with fewer new hires, and some of the earlier trend may not have been sustainable. The empirical claim is about the missing growth relative to a benchmark, not a directly observed pile of displaced workers.
The underlying data also require normalization. The authors note that Current Population Survey employment trends and the Current Employment Statistics totals diverged after the pandemic. Rather than allowing that aggregate discrepancy to make many occupations look artificially weak, they use CPS employment shares and convert them to levels consistent with CES aggregate growth. This adjustment does not manufacture a coder-specific result—the paper reports that the main controlled estimates are identical when employment shares are used—but it clarifies what the series is intended to represent. The comparison is about the coder group’s position within the broader employment path, not about one survey’s uncorrected total alone.
The monthly design brings its own constraints. Survey respondents rotate in and out of the CPS, seasonal patterns matter, and the study has 133 observations. The authors use Newey–West standard errors with a long lag allowance to account for autocorrelation, seasonality, and the survey’s rotation structure. Those choices make the reported uncertainty more credible, but they do not create additional independent episodes of AI adoption. The result remains an inference from a relatively short period around a major economic transition, not a randomized comparison of firms that did and did not receive the same technology.
For a worker, the distinction between level and trend has a practical asymmetry. Someone already employed may experience no immediate change in their job, pay, or daily tasks. Someone entering the occupation may encounter fewer openings, higher experience requirements, or a narrower path into the field. The same aggregate series can therefore feel stable to incumbents and deteriorating to entrants. That does not prove an AI-specific mechanism, but it explains why a modest-looking change in annual growth can have consequences before headline employment turns negative.
This distinction between level and growth rate is easy to miss because ordinary language uses “employment fell” for several different events. If 100,000 new coder jobs would have been added under the old trend but only 50,000 are added under the new one, the occupation has experienced a meaningful loss relative to its earlier path even if its headcount is higher than before. For new entrants, the difference can feel like a shrinking market. For existing workers, it may appear as fewer openings or slower replacement hiring rather than mass layoffs.
It also explains why the result can matter before an occupation shows a large negative level change. Hiring is often the adjustable margin. A firm may retain workers who know its systems, reduce new intake, and wait for evidence about tools and demand. Over time, fewer entrants or fewer replacement hires can alter the occupation’s age and experience structure even if the current stock remains substantial.
The study does not claim that this is the only channel operating. Employment can change because of business formation, investment, financing, regulation, or worker movement across occupations. The post-2022 pattern is an empirical signal that needs attribution, not a complete story about individual careers.
Why the pattern is more than a single chart
Several features make the slowdown harder to dismiss as a single arbitrary specification. The authors rerun the analysis after removing the Covid period, whose labor-market disruptions make a local comparison around 2022 difficult. They also remove 2022 and 2023 to allow for uncertainty about when an AI-related response could have begun. The post-break change remains negative in the main comparisons, although its size and precision vary.
They test alternative definitions of the coder group, including stricter and looser programming thresholds, and examine a version without software developers. The signs are generally consistent, though statistical strength is not identical across every variant. That is the right way to read robustness: the core pattern survives meaningful changes, but the exact estimate is not a fixed property of nature independent of classification choices.
The paper also compares coder-intensive and non-coder-intensive industries. This is not a perfect experiment, because the groups differ in many ways. It is still informative. If a broad macroeconomic shock affected every occupation equally, one would expect similar changes across groups after suitable controls. The stronger pattern in coder-intensive settings is consistent with an occupation-related component, though it does not identify AI as the only possible component.
Timing supplies another qualification. The gap between observed coder employment and the industry-based counterfactual widens notably in the middle of 2024 rather than immediately after the November 2022 release. That delay is plausible if firms needed time to evaluate model progress, redesign workflows, and decide whether expected productivity gains justified a change in hiring. It also complicates a simple event-study interpretation. A statistical test identifies other breaks, including one before the ChatGPT release, and the authors caution that Covid-era changes and sampling noise make the exact date difficult to read.
The paper investigates wages as well, but it does not find a visible break in average coder wages around 2022. That result does not cancel the employment finding. Wage averages can be affected by changes in the mix of younger and older workers, experience, location, or specialization. It does, however, reinforce the narrower interpretation: the clearest signal in this evidence is the number and growth of coders, not a confirmed collapse in their pay.
Finally, the authors emphasize that their estimate is not a clean causal effect of ChatGPT. The industry counterfactual removes an important source of confounding, and the historical examples used to assess the method make the approach more credible. But the residual can still contain occupation-specific forces unrelated to language models. The implementation of changes to R&D tax treatment, post-pandemic adjustment in software-heavy industries, and other contemporaneous shifts are discussed as possible alternatives. The estimate is best described as robust evidence of a coder-specific slowdown around the generative-AI era, with AI a plausible and important explanation, rather than as a final accounting of AI-caused job losses.
That calibrated wording preserves both sides of the finding. It would be misleading to say that coder employment was unaffected simply because the raw level continued to grow. It would be equally misleading to say that the study proves a fixed number of coding jobs disappeared. The evidence shows a gap between the path coders were on and the path they followed after 2022. The next question is whether that gap is mainly a symptom of weak industries or evidence that those industries changed their occupational mix—precisely the distinction an industry counterfactual is meant to examine.