AI Is Changing Jobs. Our Data Need to Catch Up.
Exploratory, evidence-informed thinking on how AI is reshaping work, productivity, and economic opportunity from the OpenAI Economic Research team.
AI Is Changing Jobs. Our Data Need to Catch Up.
What our labor statistics miss and how to build a clearer picture of a changing workforce
Labor-market data have long struggled to keep pace with changes in the actual work that people do. This problem predates AI, but AI is making it more visible and more urgent.
As a labor economist and the Head of Labor and Workforce Economics on the OpenAI Economic Research team, I study how technology is changing jobs, skills, and hiring. That work relies on federal surveys and government data. These tools allow us to compare the labor market over time, but they often fail to capture how legacy roles change over time and how new roles emerge.
Consider three questions that should be straightforward: Are layoffs rising across U.S. occupations? Is employment growing or shrinking in AI-related roles such as data scientists and product managers? Are wages changing for early-career software developers? Today, no single dataset can fully answer these questions.
Three measurement problems stand in the way: job categories lag behind changes in actual work, major datasets capture different parts of the labor market and cannot easily be connected, and detailed estimates often rely on samples too small to support firm conclusions.
These measurement challenges are pervasive and frustrating. They are also fixable. Addressing them will require job categories that update more quickly, consistent classifications that work across datasets, and survey samples large enough to support detailed analysis. With those changes, researchers could identify sooner and with greater precision which roles are growing or shrinking, whether wages are changing among early-career workers, and where new kinds of work are emerging.
New jobs don’t fit old labels
Job categories change slowly by design. The Bureau of Labor Statistics (BLS) typically revises its formal job categories about every eight to 10 years. That stability helps researchers track long-term trends, but it can leave new roles hidden inside outdated categories for years. As AI reshapes what people do at work, that lag becomes more consequential.
The role of product managers illustrates this problem. The Standard Occupational Classification (SOC), which federal statistical agencies use to categorize workers based on the work they perform, does not recognize product manager as a distinct occupation. Depending on their responsibilities, these workers may be grouped with marketing managers or other management occupations. As a result, researchers cannot consistently track how many product managers there are, how their wages are changing, or how AI is affecting their work.
The same lag affects industry classifications. Rachel Lipson of the Harvard Kennedy School and Allison Schwartz of Anthropic describe how jobs related to quantum computing may be classified under electrical component manufacturing or support activities for oil and gas operations. Some work involved in building and operating data centers may appear under companies that lease nonresidential buildings.
Broad or outdated categories obscure changes in new jobs.
Although today’s emerging roles may be difficult to identify, the creation of new kinds of jobs is not unique to the AI era. Research by David Autor and his coauthors finds that more than half of current employment is in job specialties introduced since 1940. Official records eventually captured occupations such as airplane designer, mental-health counselor, wind-turbine technician, and cybersecurity analyst.
The challenge is to recognize these new areas of work early enough to measure them while they are still emerging.
Labor data are fragmented
Classification is only part of the problem. Even when an occupation has its own SOC category, researchers cannot necessarily identify it in every federal dataset. Data scientists, unlike product managers, have their own SOC category, but the Current Population Survey (CPS) groups them with other mathematical science occupations and does not indicate which respondents are data scientists. Researchers can therefore estimate employment and wages only for the combined group, not for data scientists alone.
Limited occupational detail also makes it difficult to study layoffs. The Bureau of Labor Statistics’ Job Openings and Labor Turnover Survey (JOLTS) reports monthly job openings, hires, quits, and layoffs by industry, but it does not collect workers’ occupations. It can show that layoffs increased in an industry, for example, but not whether the affected workers were software developers, accountants, or salespeople.
These sources provide different pieces of the picture. Household surveys such as the CPS connect occupation with age and earnings, but they do not measure employer hiring and layoffs as JOLTS does. JOLTS provides timely measures of worker flows, but not by occupation or career stage. No source brings these elements together, making it difficult to see how changes in demand, employment, and wages are affecting particular occupations and groups of workers.
Sample size limits detail
Another constraint on using data to analyze changes in the workplace is sample size. The CPS surveys about 60,000 households each month. That sample size may be helpful to measure national employment and unemployment, but the sample thins quickly when researchers divide it across hundreds of different occupations, age groups, and other characteristics.
Consider whether wages are changing for early-career software developers. Answering this requires data that combine occupation, earnings, and some measure of career stage. The Occupational Employment and Wage Statistics (OEWS) program reports wages by occupation but does not include workers’ ages or career stages. The CPS combines occupation, age, and earnings, but only a small number of younger software developers appear in a monthly sample. Given this, an apparent wage change may reflect sampling variation, but it may not indicate a meaningful shift in the labor market.
Declining participation in these federal surveys makes the constraint more severe. CPS response rates have fallen from about 92% in January 2010 to roughly 65% in early 2026. When fewer households respond, the survey may underrepresent groups that are less likely to participate. If response rates vary by income, education, race, or gender, the resulting estimates may be biased. Survey administrators should explore more dynamic, engaging survey techniques and other innovative methods to encourage responses.
What should change
These gaps point to one priority: a labor-data system that can show how employment, wages, hiring, and layoffs differ across specific kinds of work without sacrificing comparisons over time. Achieving that goal will require three related changes.
First, occupational classifications should combine stability with greater flexibility. Each job could retain a primary occupation code that supports longitudinal analysis, while a second, more frequently updated label captures emerging roles and specialties. A software developer, for example, might retain that primary classification while also being identified as an AI infrastructure engineer or AI application developer.
Second, federal agencies should carry occupational detail more consistently through the data products that researchers use. The CPS should distinguish more of the occupations already recognized in the federal taxonomy. JOLTS should explore collecting or publishing information about hiring and separations by occupation. OEWS should provide more timely estimates for detailed occupations. Where datasets cannot use identical categories, agencies should publish clear crosswalks that allow researchers to connect them.
Third, agencies should improve survey participation and maintain samples large enough to support detailed analysis. These improvements would help ensure that federal statistics adequately represent the workforce and produce reliable estimates for specific occupations and groups of workers.
Are layoffs rising in particular occupations? Are roles involved in developing and deploying AI growing or shrinking? What is happening to early-career software developers? The federal statistical system should be able to answer these questions sooner and with greater precision.
Better data would not establish whether AI caused every change to specific roles or types of work, nor would it predict what comes next. But it would give researchers, workers, employers, and policymakers firmer ground for understanding changes already underway. Much of the necessary work is unglamorous: revising classifications, redesigning surveys, expanding samples, and making datasets work together. Modernizing these systems is not a technical footnote. It is essential to understanding how work is changing as AI adoption grows.
– Alex Martin Richmond (Labor Economist)
Subscribe to keep up with the Economics of AI.
SubscribedLearn more about the OpenAI Economic Research Team’s work.
© 2026 The Economics of AI
San Francisco, CA