Improving Federal Survey Data Collection on AI's Workforce Impacts

This spring, a bipartisan group of nine senators, led by Senators Warner and Hawley, sent a letter urging the Bureau of Labor Statistics (BLS), the Department of Labor (DOL), and the Census Bureau to improve data collection on AI's workforce impacts. The senators are following up on a directive Congress attached to the 2026 federal budget, which called on BLS to track AI's impacts on the economy, "including job loss, creation, and displacement." This directive responds to a gap in public visibility into AI’s workforce impacts, and considerable uncertainty about how these effects might evolve in the future.

The senators' letter pointed to three specific surveys as opportunities for collecting better survey data on AI’s impacts: the National Longitudinal Survey of Youth (NLSY), the Current Population Survey (CPS), and the Job Openings and Labor Turnover Survey (JOLTS).

This post discusses each, offers suggested new survey modules for the NLSY 2027 cohort and for the CPS, and offers an alternative to revising JOLTS that may better address the senators’ underlying goal of tracking occupation-level changes in employment, earnings, hiring, and turnover.

Government surveys alone will not give policymakers perfect insight into how AI is affecting jobs. Data from AI model developers, disclosures from AI-adopting businesses, and independent data collection by academic researchers will all play important roles too. However, revising government surveys may be a particularly useful lever to improve tracking of workforce impacts because they can produce data that is publicly available and consistently defined over time. That consistency is arguably the most important reason to invest in federal data collection now.

Paths to Implementation for BLS and Census

The National Longitudinal Survey of Youth

Add AI questions to the NLSY before Round 1 data collection begins in 2027

The National Longitudinal Surveys are a BLS-sponsored series that tracks the same individuals repeatedly over decades, capturing labor market experiences alongside education, income, and health data. Because they follow the same people over time, adding questions on AI usage to these surveys can reveal how AI usage is related to career trajectories and wages in ways that short-term snapshots cannot. And because the NLSY27 will follow a youth cohort, it would be especially useful for studying early-career adaptation to AI, rather than for measuring AI’s effects on incumbent mid-career and older workers. The upcoming NLSY27 cohort, set to begin data collection in 2027, offers a rare opportunity to start collecting data on AI usage alongside other determinants of future labor market outcomes for a group of American youth. The questionnaire design phase is already underway and the pretest period is happening this year, making any survey edits related to AI time-sensitive. A proposed question set covering AI adoption, usage intensity, and perceived impact is included in Appendix B below.

The Current Population Survey

Add three monthly questions on AI adoption to the CPS

The Current Population Survey is a monthly household survey conducted jointly by the Census Bureau and BLS, covering roughly 60,000 households. It is a primary source for federal labor force statistics, including the monthly unemployment rate. Three short monthly questions (see Appendix C) covering whether workers use generative AI for work, how often, and whether their employer provided it would deliver high-frequency, nationally representative data at low burden. Because the AI items would be part of the same instrument as labor market outcomes, researchers could link adoption to employment, hours, and earnings directly, rather than relying on assumptions about how adoption rates measured in one survey map onto outcomes measured in another.

Add a dedicated annual supplement on AI usage

An annual supplement could go into more detail (See Appendix D), for example, by collecting self-reported estimates of how much time generative AI saves workers. This would allow for nationally representative (though potentially still noisy) productivity estimates that could be broken out by occupation group, demographic groups, and by intensity of use. Proposed survey questions would capture worker-reported substitution of tasks previously done without AI, alongside changes in immediate team size, enabling estimates of how jobs and workflows are changing. By asking non-users why they do not use generative AI, an annual supplement could identify barriers such as training gaps, privacy concerns, employer restrictions, or regulatory frictions, providing diagnostic information necessary for designing policy that can accelerate adoption.

However, even with these supplements, the CPS on its own would not capture firm-level data, measure task-level productivity in a verifiable way, or allow for robust causal identification of employment impacts. The CPS’s sample size would prohibit robust analysis of AI’s potential effects within subgroups of workers in occupations with limited data. Establishment surveys such as the Business Trends and Outlook Survey and Annual Business Survey, larger-scale collection of wage and employment data by occupation, and experimental studies that attempt to estimate productivity impacts, would remain necessary complements.

Job Openings and Labor Turnover Survey

De-prioritize adding AI questions to JOLTS in favor of supporting Enhanced Wage Records collection by state Unemployment Insurance agencies

The senators' letter suggested adding questions on the occupations and wages of new hires and of workers who separate from their employers to the Job Openings and Labor Turnover Survey (JOLTS). The goal would be to understand how hiring, separations, and wages are changing at the occupation level as AI adoption increases.

JOLTS is not well suited to this purpose. It is a business-level survey designed to measure aggregate labor demand and worker flows, not to collect detailed information on individual workers. Its basic survey form asks questions about firm-level employment, such as the total number of hires and separations at the firm within a given reporting period. Expanding JOLTS to collect occupation-specific wage and employment information would substantially increase reporting burden for businesses and could reduce response rates. Even if successful, the resulting data would still come from a relatively small sample of employers (approximately 7,000–8,000 businesses complete the JOLTS survey each month).

An alternative approach would be to expand the collection of Enhanced Wage Records (EWRs) through state Unemployment Insurance (UI) systems. Nearly all employers already submit quarterly wage records to state UI agencies covering almost all wage and salary workers in the United States. These records provide information on individual workers' earnings and employers, but most states do not currently require employers to report workers' occupations, hours worked, or work locations. Adding these fields would create a much richer administrative dataset of “enhanced” records for understanding labor market change.

The principal advantage of EWRs is scale. Rather than collecting new information through a survey that samples from all workers or businesses, policymakers could build on the existing administrative reporting system that covers the vast majority of U.S. workers who are eligible for UI benefits. Occupation-enhanced wage records would support substantially more precise estimates of employment and earnings trends within detailed occupations than are currently possible using household surveys such as the Current Population Survey (CPS). For many detailed occupations, administrative wage records would provide sample sizes that are orders of magnitude larger than those available from monthly survey data.

Because UI wage records are collected by state agencies, federal policymakers have limited direct authority over their contents. However, the federal government could encourage broader adoption of EWRs by: (1) requiring states to collect occupation and hours-worked information through federal legislation, (2) providing funding to help states modernize their reporting systems, or (3) developing model legislation and technical standards that states could adopt voluntarily. Standardization would be particularly valuable because it would improve the comparability of occupation data across states.

If occupation information were collected nationwide through wage records, researchers could combine those data with occupation-level measures of AI exposure and adoption from sources such as the Anthropic Economic Index. This would create a much larger sample size infrastructure for tracking the relationship between AI adoption, employment, wages, hiring, and worker turnover than could be achieved through a JOLTS supplement alone. And because the additional data collection would be a matter of adding only a few fields to existing wage records, this would be far less burdensome than having businesses report on all of their workers through a revised JOLTS survey as well.

Nevertheless, adding fields to wage records would still add some reporting burden on businesses, even if it would be less than some other alternatives like reforming JOLTS or mandating businesses to disclose AI-related job impacts. Mapping payroll job titles to Census occupation codes can be difficult, and impose overhead on firms as well. This could strengthen the argument for developing federal standards around how to report occupation as part of Enhanced Wage Records. A recent House Committee on Appropriations draft for fiscal year 2027 recommends that the Department of Labor develop a set of common data standards for use by states to structure their EWR collections and reporting approach. There is also potential for LLMs themselves to be used to ease the mapping from payroll job titles to SOC codes.

Limitations of Expanding Government Survey Data Collection on AI’s Workforce Impacts

Even with these improvements, federal surveys would still leave several important measurement gaps.

Limited Visibility into Firm-Level AI Adoption

Worker surveys can reveal whether workers use AI, how often they use it, and whether their employer provides access to AI tools. They provide less detail on how firms are changing workflows, redesigning jobs, altering skill requirements, or adjusting hiring plans in response to AI adoption.

Existing establishment surveys can help fill part of this gap. The Annual Business Survey and the Business Trends and Outlook Survey already provide channels for collecting firm-level evidence on AI adoption and use. Expanding or refining these instruments could help policymakers understand how businesses are deploying AI across industries, tasks, and firm sizes. For example, the Annual Business survey could run an annual module on AI adoption at the firm level.

Gaps in Standardized Private-Sector Data

Federal surveys could also be complemented by more standardized private-sector data on AI usage and deployment. AI developers, cloud providers, payroll processors, hiring platforms, and large employers may have timely information about where AI tools are being used, which tasks they are being used for, and how hiring and skill demands are changing in response.

These data are currently inconsistently defined across providers and often inaccessible to researchers and policymakers. The Department of Labor’s Workforce AI Research Hub could help address this gap by supporting data-sharing partnerships and common standards for anonymized AI usage data from the private sector, as well as employment and wage data from large employers, payroll providers and hiring platforms. This kind of private-sector data would be especially valuable to overlay with large administrative datasets that capture occupation-level trends, such as Enhanced Wage Records, where direct measures of AI usage will not be collected.

Difficulty Identifying AI’s Causal Effects

Better government survey data would still leave researchers with a difficult causal inference problem. Employment, wages, hiring, and turnover are shaped by many forces at once, including macroeconomic conditions, sector-specific shocks, demographic change, trade, and non-AI forms of automation. AI adoption is also likely to be correlated with other firm-level changes, since firms may adopt AI while expanding, restructuring, or changing their business model.

Improved surveys and administrative records can help researchers track correlations, identify leading indicators, and build more granular descriptive evidence about the relationship between AI usage and worker outcomes. Estimating AI-specific causal effects however will often require longitudinal data, natural experiments, and firm- or worker-level research designs that compare otherwise similar groups with different levels of AI exposure or adoption.

The federal survey avenues proposed in this post are therefore best understood as one layer of a broader AI workforce data infrastructure. Their main value is providing public, nationally representative, consistently defined measures of AI adoption and labor-market outcomes over time. To maximize their value, policymakers should pair them with establishment surveys, private-sector data-sharing initiatives, support for state-level collection of EWRs, and independent research designs capable of estimating causal effects.

Appendix A: Preamble/Screener (common to all surveys)

Read to all respondents: "Generative AI is a type of artificial intelligence that creates text, images, audio, or video in response to prompts. Some examples of generative AI include ChatGPT, Claude, Gemini, Copilot, and Midjourney." [Note that this prompt may need to be updated over time as new AI products emerge. The goal should be to understand adoption of the latest and most powerful available AI tools. Today these are broadly understood to be generative AI systems like those listed above, but in the future this may change and this terminology could be outdated.]

Appendix B: Suggested Questions for the NLSY27 Core Questionnaire

Questions 1–3 are asked of all respondents. Questions 4–6 are asked of employed respondents (with Q5 conditional on use of generative AI at work). Questions 7–9 are asked of respondents currently in school or training (with Q8 conditional on use of generative AI for schoolwork). Respondents who are both employed and in school answer both sets.

All Respondents

1. "In the past year, did you use generative AI tools to do your [job/schoolwork] (such as ChatGPT, Copilot, Claude, or Gemini)?" (Yes / No)
2. If yes to Q1: "How often?" (Daily / A few times a week / A few times a month / Once or less)
3. "Do you use generative AI tools outside of [your job/school]?" (Yes / No)

Employed Respondents

1. "Which best describes your employer's stance on generative AI use at work?" (Required or expected / Encouraged but not required / Neither encouraged nor discouraged / Discouraged or restricted / Formally prohibited / Don't know)
2. If respondent uses generative AI at work: "Has the use of generative AI tools changed the type of tasks you perform at work?" (Yes / No / Don't know)
3. "How do you think generative AI will affect your future job prospects?" (Improve them / Worsen them / Don't know)

Respondents Currently in School or Training

1. If respondent uses generative AI for schoolwork: "Which best describes your school or instructor’s stance on generative AI use for schoolwork?" (Required or expected / Encouraged but not required / Neither encouraged nor discouraged / Discouraged or restricted / Formally prohibited / Don't know)
2. "How do you think generative AI will affect your future job prospects?" (Improve them / No effect / Worsen them / Don't know)

Appendix C: Suggested CPS Questions for Monthly Survey

Monthly Survey Additions

1. "In the past month, did you use generative AI tools (such as ChatGPT, Copilot, Claude, or Gemini) for work-related tasks?" (Yes / No)
2. If yes: "How often?" (Daily / A few times a week / A few times a month / Once or less)
3. If yes: ""Does your employer provide generative AI tools or paid subscriptions for work use?" (Yes / No / Don't know)

Appendix D: Suggested CPS Annual Supplement

Annual Supplement: Generative AI Module

Section A: Core Usage (all who pass screener)

1. "Do you use generative AI for your job?" (Yes / No)
2. "Do you use generative AI outside your job?" (Yes / No)
3. "Which of the following types of generative AI tools have you used?" (select all): A standalone AI chatbot accessed via a web or mobile app (e.g., ChatGPT, Gemini, Claude) / An image, audio, or video generator (e.g., Midjourney, DALL-E) / An AI system that can attempt to carry out multi-step computer projects independently (e.g., Claude Cowork, Claude Code, or OpenAI’s Codex) / AI features embedded in other software you use (e.g., in email, office, or design applications) / Other

Section B: Frequency and Intensity (employed genAI work users)

1. "How many days did you use generative AI for work last week?" (0 / 1 / 2–4 / 5 or more)
2. "On days you use it, approximately how much time do you spend?" (Less than 15 min / 15–59 min / 1–4 hours / 4 or more hours)
3. "In a typical week, approximately how many hours do you work?" [respondent enters X]
4. "If you did NOT have access to generative AI, how many additional hours per week would you need to complete the same amount of work?" (Less than 1 hour / 1 hour / 2 hours / 3 hours / 4 or more hours)

Section C: Task Domains (employed genAI work users, last week)

1. "For which of the following tasks did you use generative AI at work last week?" (select all): Writing communications / Administrative tasks / Interpreting, translating, or summarizing / Searching for information / Generating new ideas / Coding or programming / Customer or coworker support / Data analysis or visualization / Tutoring or educational assistance / Other
2. "Of the tasks you selected, which ones did generative AI help with the most?" (select up to three)

Section D: Workplace Context (employed respondents)

1. "Does your employer provide generative AI tools or paid subscriptions for work use?" (Yes / No / Don't know)
2. "Has your employer provided any training on how to use generative AI?" (Yes / No)
3. "Which best describes your employer's stance on generative AI use at work?" (Required or expected / Encouraged but not required / Neither encouraged nor discouraged / Discouraged or restricted / Formally prohibited / Don't know)

Section E: Perceived Impact (employed genAI work users)

1. "Compared to one year ago, has the use of generative AI tools changed the type of tasks you perform at work?" (Yes / No / Don't know)
2. "Compared to one year ago, has generative AI changed the quality of your work output?" (Improved / About the same / Decreased)
3. "How do you think generative AI will affect your future job prospects? (Improve them / No effect / Worsen them / Don't know)

Section F: Substitution and Organizational Change (employed genAI work users)

1. "In the last year, did you use generative AI to perform tasks that you previously did without it?" (Yes / No / Don't know)
2. If yes: "How would you describe the number of tasks where generative AI replaced work you previously did without it?" (A small number / A moderate number / A large number)
3. "In the last month, did generative AI replace other software or digital tools you previously used for work?" (Yes / No / Don't know)
4. "To use generative AI at work, have you or your employer made any of the following changes?" (select all): Received training / Developed new workflows / Employer purchased new software or subscriptions / Employer hired AI-skilled staff / Changed data practices / Other / None
5. "In the last six months, has the use of generative AI led to any change in the number of people in your immediate work team (the coworkers you directly work with on a regular basis)?" (Increased / Decreased / No change / Don't know)

Section G: Reasons for Non-Use (employed non-users of genAI for work)

1. "What is the main reason you do not use generative AI for work?" (single select): Not relevant to my job / Don't know enough about AI / Employer discourages or prohibits / Accuracy or reliability concerns / Privacy or data security concerns / AI is not a mature enough technology / Too expensive / Concerns about bias / Laws and regulations prevent or restrict use / Tried but didn't meet expectations / Other