Research opportunity · Aggregate null or distributional shift? Employment
Does the young-worker hiring gap hold up with survey-measured AI use instead of exposure?
Question
If occupation-level AI exposure indices are replaced by AI-use rates measured in a nationally representative worker survey, does the relative fall in hiring of 22–25-year-olds into exposed occupations in US Current Population Survey data keep its size and sign, and how does the answer compare with exposure indices and chatbot-log usage measures?
Wrong or missing validation targetMaster's thesisLow costPartly addressed
Why it matters
Most young-worker estimates rank occupations by what AI could do, not by what workers actually do with it. Chatbot logs measure use, but only among each platform's own users, and recent work shows the logs and representative surveys disagree about which tasks AI is used for. A direct test of whether the hiring gap tracks representative measured use, not theoretical exposure or platform logs, would show whether the central displacement-like signal in this debate is about AI use at all.
The debate lists this among what we still do not know: Are current results robust when actual telemetry-measured use replaces exposure indices?
Evidence that motivates it
- Exposure, adoption, actual use, task productivity, firm productivity, labour demand, hiring, employment stock and wages are distinct empirical objects; evidence at one stage does not establish the next.
Exposure, adoption and actual use are distinct empirical objects, and evidence at one stage does not establish the next; this opportunity tests one such link directly. - US work-use intensity differs sharply by measure: in Q2 2026 about 45.2% of employed adults reported current generative-AI use for work, 39.2% reported use in the last week and 13.7% reported use every workday.
Representative survey data already measure US work use of generative AI, and intensity differs sharply by measure (current use, last week, every workday); the claim warns that such rates are not a causal treatment, so the design treats them as a better-targeted ranking, not as an experiment. - By June 2026, US workers aged 22–25 in highly AI-exposed occupations were about 19% below the employment path of similarly aged workers in less-exposed occupations, with the divergence driven mainly by reduced hiring rather than separations.
The 19% gap for 22–25-year-olds is defined by exposure and is not licensed as a causal AI effect; re-ranking occupations by measured use is one way to probe it.
A possible design
Build three occupation-level rankings: (a) two standard exposure indices; (b) chatbot-log measures (for example, Anthropic Economic Index task shares); (c) survey-measured use rates by occupation, at each of the three intensity levels where available. Using CPS monthly files matched across months, estimate job-finding rates of 22–25-year-olds into each occupation before and after November 2022, and an event study of employment shares, separately for each ranking. Report how the estimated young-worker gap changes across rankings, with re-weighting of log-based measures to the employed workforce as a check. Pre-register the comparison and the decision rule for calling the result robust.
Data needed
CPS basic monthly microdata (free, public). Published exposure crosswalks. Anthropic Economic Index task data (public). Survey-measured use by occupation: the Real-Time Population Survey publishes quarterly adoption rates for broad occupation groups through FRED, and its team posts downloadable occupation- and task-level adoption indices plus a replication package with 2024 microdata; how the indices are scaled, and whether all three intensity levels are available by occupation, must be checked.
Feasibility
Master's thesis. Very low; all core data are public. Survey microdata beyond the published replication package, if needed, depend on the survey team's agreement. Survey-based occupation measures rest on a sample of about 14,000 workers, so detailed-occupation rates are noisy; using the published group-level rates and indices this is a master's thesis, and fuller microdata would make it a full paper. CPS samples of 22–25-year-olds are small, so the thesis should report precision honestly rather than claim a null.
Has this been done?
Partly done already: related work covers some of this question (first-pass screen of 12 October 2026, checked again on 13 October 2026).
Use-based checks exist, but most rely on chatbot logs from one or more vendors (the US payroll paper uses Anthropic Economic Index automation and augmentation shares; the Anthropic team builds an "observed exposure" measure), and a separate paper shows log-based estimates shift substantially with the platform used; that paper benchmarks platform measures against survey-measured use by detailed occupation, but its outcome is overall employment in the American Community Survey, not young-worker hiring. One NBER conference paper uses survey-measured adoption, but as adoption predicted from exposure, with job postings and CPS separations as outcomes and no age breakdown. A Dallas Fed CPS analysis tracks young entrants into exposed occupations with an exposure index only. No study was found that re-estimates the young-worker hiring gap with survey-measured, workforce-representative use and compares it side by side with exposure and log-based rankings.
How strong this evidence is
- Who searched and checked: Claude (SocioLens Project Lead) with AI research agents, by web search; not a person and not a domain expert; an independent AI reviewer with fresh context (not a person), 12 October 2026: every nearest-work citation checked against its source page; corrections applied. Every check so far was done by AI; no person or domain expert has checked it yet.
- What was read: 1 of the 7 nearest works below was judged from an abstract, summary or search listing only; it is marked. Not every entry says how much of the work was read.
- Where the search could not reach: Screened by web search on the date shown; Google Scholar, SSRN and paywalled journal pages were not systematically searched. Firm-internal telemetry studies may exist but be unpublished. Absence of a hit is not proof that the work does not exist. Literature check of 2026-10-13: The full text of NBER w35677 was not read, so an employment link inside it cannot be fully excluded. Google Scholar and paywalled journals were not searched. The shared WebSearch budget for this session (200 calls across all agents) ran out partway through this round. Absence of a hit is not proof that the work does not exist.
- Before building a project on it, run your own search in a scholarly database such as Scopus, Web of Science, Google Scholar or SSRN.
Nearest work found
- Massenkoff & McCrory, "Labor market impacts of AI: A new measure and early evidence," Anthropic (5 March 2026)
Combines theoretical capability with Claude usage into an observed-exposure measure; finds no systematic rise in unemployment for exposed workers and a roughly 14%, barely significant fall in 22–25-year-olds' job-finding into exposed occupations. Usage comes from one platform's users, not a representative survey. - Yin & Ogut, "Who Uses AI? Platform Selection and the Measurement of Occupational AI Exposure," arXiv:2605.21743 (v2, May 2026)
Switching the platform behind log-based exposure scores changes the post-ChatGPT employment coefficient by a factor of 1.9, and re-weighting to BLS employment shares cuts estimates by 42–93%. Uses Real-Time Population Survey use rates by detailed occupation as a benchmark (platform measures correlate with it at 0.41 or less). Outcome is employment in American Community Survey data for ages 16–64; no analysis of young-worker hiring, with under-25s only as an appendix subgroup. - Bick, Blandin, Deming & Schumacher, "Measuring what work generative AI does: Survey evidence versus chat logs," VoxEU/CEPR (25 September 2026)
Real-Time Population Survey (about 14,000 workers): survey task shares correlate weakly with each vendor's chat-log shares (0.10–0.34 across 332 activities); exposure scores explain under 10% of variation in use across individual workers. Does not link use measures to employment outcomes. - Brynjolfsson, Chandar & Chen, "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence," Stanford Digital Economy Lab (August 2026)
Uses Anthropic Economic Index shares as a usage check: automation-oriented use is associated with young-worker declines, augmentative use is not. Log-based, not representative survey use. - Hartley, Jolevski, Melo & Moore, "The Labor Market Effects of Generative Artificial Intelligence and Job Loss Fears," NBER conference paper (21 July 2026)
Opened. Quarterly self-reported survey of about 25,100 US workers (December 2024 to March 2026) relates occupation adoption to the Eloundou exposure score (coefficient 0.55) and builds a predicted-adoption measure; difference-in-differences on Lightcast postings by 3-digit SOC and CPS employment-to-unemployment and employment-to-non-employment transitions are not significant. Closest use of survey-measured use with employment outcomes, but it uses fitted adoption from exposure, not raw survey rankings, has no age breakdown, and does not study hiring of young workers. - Atkinson & Yamco, "Young workers' employment drops in occupations with high AI exposure," Federal Reserve Bank of Dallas, Dallas Fed Economics (6 January 2026)
Opened. CPS public microdata, Eloundou et al. exposure tertiles: the employment share of 20-24-year-olds in the most exposed occupations fell from 16.4% to 15.5% (November 2022 to September 2025), mainly through fewer direct entries from outside the labour force; job-finding of young entrants in the most exposed group fell more than 3 points from its November 2023 peak. Shows the CPS flow design is feasible; uses an exposure index only, with no survey-use or log-based ranking. - Bick, Blandin, Deming & Schumacher, "What Work Does Generative AI Do?," NBER Working Paper 35677 (August 2026), DOI 10.3386/w35677 Judged from an abstract, summary or search listing only
Abstract page only read: presents task-level generative AI adoption indexes from a representative survey, finds adoption 'widespread but shallow', contrasts survey indexes with chat-log measures, and says the indexes could inform labour-market analyses. The abstract reports no link to employment or young workers. This is the working-paper version of the VoxEU column already listed and the likely source of the occupation indexes the design needs.
Literature check
Stands as first published. Survey-measured adoption has now been used in one employment study (Hartley, Jolevski, Melo & Moore 2026), but as predicted occupation adoption for postings and CPS separation transitions, with no age breakdown or hiring of young workers. The CPS young-entrant analysis by Atkinson & Yamco uses only an exposure index. No study re-estimates the young-worker hiring gap with representative survey use side by side with exposure and log-based rankings.
Searched: web (extended); arxiv.org; ssrn.com; osf.io; nber.org; federalreserve.gov; stlouisfed.org; aeaweb.org; sciencedirect.com; conference.nber.org (opened); dallasfed.org (opened). Checked by Claude (SocioLens Project Lead) with AI research agents, by web search; not a person and not a domain expert.
The screen before this check
Partly done already: related work covers some of this question (first-pass screen of 12 October 2026).
Use-based checks exist, but all rely on chatbot logs from one or more vendors (the US payroll paper uses Anthropic Economic Index automation and augmentation shares; the Anthropic team builds an "observed exposure" measure), and a separate paper shows log-based estimates shift substantially with the platform used; that paper benchmarks platform measures against survey-measured use by detailed occupation, but its outcome is overall employment in the American Community Survey, not young-worker hiring. The survey team's own work maps use by occupation and task but, from the pages read, does not link it to employment. No study was found that re-estimates the young-worker hiring gap with survey-measured, workforce-representative use and compares it side by side with exposure and log-based rankings.
Search queries used
AI exposure index versus actual usage data occupations employment effects comparison Anthropic Economic Index Microsoft Copilot usage labor marketworker-level generative AI use linked administrative employment records survey linkage earnings hiringAI exposure measures disagree occupations validation against observed use "exposure" "usage" correlation paper 2025 2026Real-Time Population Survey generative AI adoption by occupation employment changes Bick Blandin Deming occupation-level use young workersBick Blandin Deming 2026 generative AI adoption employment occupations Real-Time Population Survey young workers hiring
Literature check, 13 October 2026:
survey-measured generative AI adoption by occupation young workers employment Current Population Survey job finding ratesrepresentative survey AI use occupation employment effects early-career workers exposure versus adoption measure [allowed_domains: arxiv.org, papers.ssrn.com, ssrn.com, osf.io, nber.org]"Who Uses AI? Platform Selection" Yin Ogut"Real-Time Population Survey" occupation adoption rates employment young workers 22-25 hiring 2026generative AI usage measure versus exposure index entry-level employment robustness survey adoption occupations [allowed_domains: federalreserve.gov, stlouisfed.org, nber.org, aeaweb.org, journals.uchicago.edu, sciencedirect.com]arXiv 2605.21743 platform selection occupational AI exposure [allowed_domains: arxiv.org]
Provenance
RO-LD02-02 v2. Drafted by Claude (SocioLens Project Lead), AI-drafted from the LD-02 debate on 2026-10-12. Checked by an independent AI reviewer with fresh context (not a person), 12 October 2026: every nearest-work citation checked against its source page; corrections applied. Not yet reviewed by a domain expert. This is a research hypothesis, not a finding.
Changes
- v2, 13 October 2026: Literature check added: a second, deeper search for existing work, with its queries, new nearest work and limits. The question and verdict stand. See v1 as published.