# AI and employment evidence A task exposed to AI is not a job lost. Faster work is not automatically higher pay. Each record keeps the measured outcome, population, source version and limitation beside the finding. Preserve those details when quoting or reusing it. Author: Gene Dai Published: 2026-09-08 Updated: 2026-09-08 Canonical: https://digidai.github.io/research/ai-employment/ ## Method and limits - Selection: a curated set of primary research and methodological updates relevant to this site's AI-and-work coverage. It is not a systematic review, meta-analysis or estimate of AI's net employment effect. - Classify evidence by what was measured, not by the headline. Experiments, administrative-data studies, exposure models and vendor telemetry answer different questions; do not pool their percentages. - Keep paper versions distinct. The published customer-support study differs from its earlier working paper; the Denmark paper was revised and retitled in March 2026. The record specifies the version used. - Checked dates identify editorial source review. Source dates may have year, month or day precision, or be absent. Study observation periods and publication dates are different fields. These entries are not live feeds from the researchers. - Updates replace a claim only with a cited reason and a change-log entry. Conflicting evidence remains visible. For a citation, use the original study plus this record's permanent anchor if the classification is useful. ## Using this register ### Writing a defensible claim Name the measured outcome, affected population and period in the sentence containing the number. Keep the limitation in the same paragraph or table cell. ### Evaluating a workforce decision Check whether a study resembles the work, tools and experience level in your team. Use task trials and actual staffing outcomes; do not turn an exposure estimate into a layoff forecast. ### Comparing apparently conflicting studies First compare outcomes, dates, samples and identification methods. An employment gap and a task-completion-time change can both be real without estimating the same effect. ## Evidence records ### Customer support productivity: Can AI assistance improve measured output at work? Record: [support-qje-2025](https://digidai.github.io/research/ai-employment/#support-qje-2025) The published study reports a 15% average increase in issues resolved per hour after AI assistance was introduced, with different effects across worker experience levels. - Evidence type: Workplace study - Source: [Brynjolfsson, Li and Raymond: Generative AI at Work (QJE, 2025)](https://academic.oup.com/qje/article/140/2/889/7990658) - Source date: 2025-02-04 - Source checked: 2026-09-08 - Scope: Staggered introduction of assistance to 5,172 customer-support agents. This is the published QJE version, not the earlier 5,179-agent working-paper version. - Limitation: Issues resolved per hour is a task-productivity outcome in one setting, not an economy-wide employment, wage or headcount estimate. - Next check: Compare the work and experience mix with your team; count quality and escalations alongside throughput. ### Experienced developer productivity: Did early-2025 coding tools always save time? Record: [metr-developers-2025](https://digidai.github.io/research/ai-employment/#metr-developers-2025) METR's randomized study found that permitting AI tools increased task-completion time by 19% in its sampled experienced open-source developers. - Evidence type: Randomized experiment - Source: [METR: Early-2025 AI and Experienced Open-Source Developer Productivity](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) - Source date: 2025-07-10 - Source checked: 2026-09-08 - Scope: 16 developers, 246 tasks in familiar mature repositories; tools from February through June 2025. - Limitation: This result is not a ranking of current coding products and does not represent all developers, unfamiliar repositories or later models. - Next check: Read the 2026 follow-up before citing this as evidence about current tools; measure accepted changes and review time locally. ### Developer study selection bias: Can the follow-up settle the current speedup? Record: [metr-selection-2026](https://digidai.github.io/research/ai-employment/#metr-selection-2026) METR says its follow-up provides an unreliable estimate of current productivity effects because of participant and task selection, plus measurement difficulties with concurrent agents. - Evidence type: Methodology update - Source: [METR: We are Changing our Developer Productivity Experiment Design](https://metr.org/blog/2026-02-24-uplift-update/) - Source date: 2026-02-24 - Source checked: 2026-09-08 - Scope: Update on the experiment begun in August 2025, separate from the early-2025 randomized result. - Limitation: Raw follow-up estimates should not be advertised as a clean causal estimate or a definitive reversal of the earlier result. - Next check: Look for revised study designs and report who declined AI-disallowed tasks. Count concurrent-agent time consistently. ### Global occupational exposure: Does potential exposure mean jobs disappear? Record: [ilo-exposure-2025](https://digidai.github.io/research/ai-employment/#ilo-exposure-2025) ILO and NASK estimate that about one in four jobs worldwide is in occupations potentially exposed to generative AI; the source distinguishes transformation from replacement. - Evidence type: Occupational exposure model - Source: [ILO / NASK: Refined Global Index of Occupational Exposure](https://www.ilo.org/resource/news/one-four-jobs-risk-being-transformed-genai-new-ilo%E2%80%93nask-global-index-shows) - Source date: 2025-05-20 - Source checked: 2026-09-08 - Scope: Global task-based exposure classification, combining occupational tasks, expert input, AI scoring and labor microdata. - Limitation: Potential technical exposure is not observed adoption, feasible whole-job automation or a forecast of layoffs. - Next check: Pair exposure with actual adoption, wages, employment and transitions in the geography being discussed. ### Interpreting exposure indicators: What can an exposure indicator predict on its own? Record: [ilo-indicators-2026](https://digidai.github.io/research/ai-employment/#ilo-indicators-2026) ILO's 2026 methodological brief warns against reading exposure measures alone as predictions of job losses or labor-market outcomes. - Evidence type: Methodology update - Source: [ILO: What AI exposure indicators reveal about jobs](https://www.ilo.org/resource/news/new-ilo-brief-explains-what-ai-exposure-indicators-reveal-about-jobs) - Source date: 2026-04-17 - Source checked: 2026-09-08 - Scope: Interpretive guidance on exposure indicators, rather than a new measured job-loss count. - Limitation: The warning does not imply no displacement; it specifies what this type of indicator cannot establish. - Next check: Identify which result is an exposure score and which uses actual labor-market outcomes before combining studies. ### Young workers and employment: Are employment changes concentrated in particular groups? Record: [stanford-canaries-2026](https://digidai.github.io/research/ai-employment/#stanford-canaries-2026) Stanford's August 2026 revision reports a 19% employment gap for young workers in highly AI-exposed occupations relative to less-exposed peers, while reporting no widespread economy-wide displacement. - Evidence type: Observational study - Source: [Stanford Digital Economy Lab: Canaries August 2026 update](https://digitaleconomy.stanford.edu/news/canariesaug26/) - Source date: 2026-08-12 - Source checked: 2026-09-08 - Scope: ADP payroll evidence through June 2026. Workers aged 22 to 25 in the two most-exposed quintiles are compared with similarly aged workers in the three least-exposed quintiles, relative to November 2022. - Limitation: The authors describe these as non-causal patterns, with education, pre-trends and representativeness caveats. The gap is not AI-caused layoffs, an unemployment rate or 19% of all young workers losing jobs. - Next check: Check the paper revision, cohort definition, comparison group and robustness checks before making a causal or national-population claim. ### Earnings and recorded hours: Do reported task benefits show up in average labor outcomes? Record: [denmark-labor-2026](https://digidai.github.io/research/ai-employment/#denmark-labor-2026) The March 2026 NBER revision estimates null effects on earnings and recorded hours at worker and workplace levels, ruling out effects larger than 2% in its study horizon. - Evidence type: Observational study - Source: [Humlum and Vestergaard: Still Waters, Rapid Currents (NBER 33777, March 2026 revision)](https://www.nber.org/papers/w33777) - Source date: 2026-03 - Source checked: 2026-09-08 - Scope: Danish adoption surveys linked to administrative records, studying the first two years after ChatGPT. Previously titled Large Language Models, Small Labor Market Effects. - Limitation: Null average earnings and hours effects do not mean no task change, no benefit for individual adopters or no future employment effects elsewhere. - Next check: Use the revised title and estimates; distinguish the measured horizon from the publication date and the earlier version's estimates. ### Observed Claude usage: What does usage telemetry say about employment? Record: [anthropic-cadences-2026](https://digidai.github.io/research/ai-employment/#anthropic-cadences-2026) Anthropic's June 2026 report changes its sampling and classification pipeline to study hourly use and outputs, with separate reporting for Claude conversations and first-party API traffic. - Evidence type: Vendor usage telemetry - Source: [Anthropic Economic Index: Cadences](https://www.anthropic.com/research/economic-index-june-2026-report) - Source date: 2026-06-26 - Source checked: 2026-09-08 - Scope: Vendor-observed use of Claude products and first-party API; methods changed from earlier seven-day report samples. - Limitation: Claude usage is not a representative sample of all workers, a verified task-success census or a causal measure of jobs created or lost. - Next check: Read the methodology appendix before comparing report editions; keep platform use, self-reported benefits and administrative employment outcomes separate. ## Related reading - [Cursor vs GitHub Copilot: evaluate the workflow](https://digidai.github.io/2026/02/08/cursor-vs-github-copilot-ai-coding-tools-deep-comparison/) - [Open-weight model selection](https://digidai.github.io/2026/03/13/open-weight-ai-war-llama-mistral-deepseek-qwen/) - [AI hiring procurement evidence](https://digidai.github.io/research/ai-hiring/) ## Update log - 2026-09-08: Initial eight-record register. Includes the March 2026 Denmark revision, METR's February 2026 methodology warning and Stanford's August 2026 update; no cross-study effect is calculated. ## Formats and corrections [JSON](https://digidai.github.io/research/ai-employment/data.json) | [CSV](https://digidai.github.io/research/ai-employment/data.csv) | [Review worksheet](https://digidai.github.io/research/ai-employment/review-template.csv) | [Research updates RSS](https://digidai.github.io/research/feed.xml) | [Corrections](https://digidai.github.io/contact/) The review worksheet contains unanswered questions and blank fields for your evidence. It is not a completed audit or test result. These are editorial summaries, not a license to redistribute the linked publications. Cite original sources and preserve limitations.