The source record
Evidence at a glance
Dates identify the source or documented event. A published claim is not automatically a demonstrated outcome.
| Record and date | What it establishes | What it does not establish |
|---|---|---|
| 17 March 2023 U.S. task exposure | GPT research describes potential effects on tasks across the U.S. workforce. [3] | Capability rubric, not a count of dismissals. |
| January 2024 Global exposure and complementarity | IMF staff distinguish exposure from whether technology may complement work. [2] | The framework is not interchangeable with an ILO generative AI index. |
| 4 February 2025 Observed workplace productivity | The published customer-support study measures issues resolved per hour. [4] | One setting cannot establish economy-wide employment effects. |
| 20 May 2025 Global generative AI exposure | ILO publishes a revised occupational exposure framework. [1] | Exposure categories do not supply a dismissal probability. |
| March 2026 revision Early labor outcomes in Denmark | The working paper measures earnings and hours using linked survey and administrative records. [5] | Country, time window and estimated averages constrain interpretation. |
| 2024 survey vintage U.S. Census denominator | ACS B24114 defines a civilian employed population aged 16 and over and includes detailed occupations. [6] | Population estimates do not measure AI causation. |
Research and interpretation
Begin with the original claim
The ILO’s May 2025 index combines task information, worker input, expert review and model-based scoring. It organizes occupations into exposure gradients; 3.3 percent of global employment is in its highest category. The authors judge transformation more likely than wholesale replacement because occupations contain tasks requiring human input. These are estimates of technological exposure, not observed redundancies. [1]
An IMF staff discussion note published in 2024 estimated that almost 40 percent of global employment was exposed to AI. Its framework separately considers whether AI might complement a job. Comparing that percentage with the ILO figure as if one were a revision of the other would obscure differences in the technology assessed and the exposure framework. Neither percentage should be relabelled a worldwide dismissal rate. [2]
A 2023 study by researchers associated with OpenAI, OpenResearch and the University of Pennsylvania estimated that roughly 80 percent of U.S. workers could have at least 10 percent of their tasks affected by GPT-related technologies. Its unit is potential task exposure. Eighty percent is not the share of jobs eliminated, and 10 percent of tasks is not necessarily 10 percent of paid working time. The institutional affiliations are disclosed because a technology provider participated in the research. [3]
Research and interpretation
Separate capability from outcomes
The published 2025 version of Generative AI at Work studied 5,172 customer-support agents. It found a 15 percent average increase in issues resolved per hour, with different effects by experience and skill. That is evidence about productivity in the studied setting. It does not measure a 15 percent economy-wide reduction in staff. Earlier working-paper versions used different figures; quote the version actually consulted. [4]
Humlum and Vestergaard linked Danish chatbot-adoption surveys to administrative labor records. The March 2026 working-paper abstract reports no detectable average earnings or recorded-hours effects, ruling out effects larger than 2 percent over its two-year post-ChatGPT window, alongside changes in tasks. This is evidence against large early average effects in that study, not proof that no individual was affected or that future losses cannot occur. [5]
Our interpretation: an oversight brief should track at least four stages separately—technical capability, actual adoption, changes in work, and labor outcomes. A result at one stage does not establish the next. Before attributing workforce reductions to AI, request the deployment dates, affected functions, headcount and hours series, hiring changes, outsourcing decisions and the employer’s explanation. Test plausible alternative explanations rather than choosing the most compelling headline.
Research and interpretation
Choose a population that matches the question
For a U.S. occupation-based inquiry, the 2024 ACS one-year detailed table B24114 defines its universe as the civilian employed population aged 16 and over. Its metadata identifies graphic designers under estimate B24114_171E and margin-of-error variable B24114_171M. That is a concrete starting point for a national occupation count; it is not a count of everyone who has design skills or every person employed by a design business. This brief does not report an extracted graphic-designer estimate. [6]
For a question such as how many women in a particular occupation could face changed tasks, first specify geography, survey year, employment status, occupation, and the demographic variable. Use the relevant published table where available; otherwise document the PUMS selection and person weights. The Census PUMS documentation supplies variable definitions, code lists and weight guidance. Do not count sample records as if each represented one resident. [7]
Occupation codes need a documented crosswalk. Census explains that classification changes can move workers between categories and limit comparisons over time. Matching labels by eye is insufficient when a study uses a different occupational system. Publish unmatched and split categories and explain allocation assumptions. Otherwise, a precise-looking total may conceal an uncertain mapping. [8]
Research and interpretation
Show the calculation and its limits
Illustration only: suppose a clearly defined population contains 10,000 workers and an explicitly assumed scenario classifies 20 percent as potentially exposed. The multiplication is 10,000 × 0.20 = 2,000 workers in that scenario. Neither input is an actual Census estimate here. The result is not 2,000 predicted layoffs, and it has no forecast year unless a separate model supplies one.
If a published study instead says that 20 percent of tasks within an occupation meet its exposure threshold, multiplying that percentage by workers changes the unit of measurement. The result cannot be described as a count of exposed people without an additional, justified definition. Publish the numerator, denominator, unit, geography, period and assumptions beside every calculation.
Survey uncertainty and scenario uncertainty are different. Census publishes 90 percent margins of error and provides instructions for statistical comparisons. Those margins address sampling variability; they do not validate an assumed AI effect. A hypothetical exposure percentage must remain labelled even when the underlying worker estimate is official. [9]
Research and interpretation
Do not manufacture demographic effects
A group’s share of an occupation can explain how a common scenario distributes numerically across that occupation. It does not establish that the technology affects every subgroup at the same rate. Conversely, applying different rates by sex, race, disability or age requires an evidentiary basis or an explicit scenario assumption. Keep occupational composition, within-occupation exposure and observed outcomes separate.
Recommended reporting rule: show both the estimated number and the percentage of each group’s relevant workforce. State whether categories overlap, explain missing or suppressed cells, and assess whether apparent differences exceed sampling uncertainty. Census statistical guidance supports testing comparisons rather than ranking estimates by size alone. [9] Without these steps, a demographic headline can confuse a larger population with a higher individual risk.
Keep the distinctions visible
Claims, findings and unknowns
A global exposure percentage tells us how many local jobs will vanish.
What the evidence supports: An exposure estimate describes its defined population and threshold. [1] [2]
What remains open: Local adoption, task mix, employer responses and future employment outcomes.
Higher productivity means the same percentage fewer workers.
What the evidence supports: The cited workplace study measures output per hour. [4]
What remains open: Whether demand, hours, staffing, pay or service quality change afterward.
An official Census denominator makes a scenario a forecast.
What the evidence supports: Official population data and an assumed effect remain separate inputs. [6] [7] [9]
What remains open: The validity of the effect assumption and its transfer to the selected population.
Recommended oversight questions
Questions for decision-makers
These questions are our recommendations. They identify the records needed to test a claim or evaluate a decision.
01What precisely does the headline percentage measure?
Why ask: A task share, a worker share and a job-loss rate are different quantities. [1] [2] [3]
Who can answer: Study authors, analysts and communications teams
Records to request: Original study, version, threshold definition, numerator and denominator.
Link to this question →02Which population was used for the local calculation?
Why ask: An estimate must match the geography, occupation, year and employment status in the question. [6] [7]
Who can answer: Workforce agency or commissioning organization
Records to request: Census table and cells or PUMS query, filters, weights and extraction date.
Link to this question →03How were occupation codes connected between datasets?
Why ask: A crosswalk can change who is counted and conceal ambiguous matches. [8]
Who can answer: Research team and data contractor
Records to request: Code versions, crosswalk, split-category allocations and unmatched records.
Link to this question →04What evidence connects deployment to the reported labor outcome?
Why ask: An outcome following an announcement is insufficient to establish its cause. [4] [5]
Who can answer: Employer leadership, worker representatives and independent researchers
Records to request: Deployment chronology, comparison group, headcount, hiring, hours, wages, separations and outsourcing records.
Link to this question →05Are demographic differences measured or assumed?
Why ask: Group counts, within-group rates and differences in treatment should not be conflated. [7] [9]
Who can answer: Researchers and public officials using the estimates
Records to request: Subgroup definitions, unweighted sample sizes, weights, margins of error, statistical tests and scenario assumptions.
Link to this question →Source register
Sources reviewed 14 November 2026. Dates below identify the cited publication or version; undated documentation is labelled. Links open the original record.
- 01
International Labour Organization · 20 May 2025
Generative AI and Jobs A Refined Global Index of Occupational Exposure ↗
Working Paper 140. Original institutional research; occupational exposure, not observed job losses.
- 02
International Monetary Fund · January 2024
Gen AI Artificial Intelligence and the Future of Work ↗
Staff Discussion Note 2024/001. Exposure and complementarity framework; staff analysis rather than an employment forecast.
- 03
OpenAI · 17 March 2023
GPTs are GPTs An early look at the labor market impact potential of large language models ↗
Original research summary and paper link. Provider-affiliated authorship disclosed; prospective task exposure.
- 04
The Quarterly Journal of Economics · 4 February 2025
Generative AI at Work ↗
Brynjolfsson, Li and Raymond. Published version used; differs from earlier working-paper sample and estimates.
- 05
National Bureau of Economic Research · March 2026 revision
Still Waters Rapid Currents Early Labor Market Transformation under Generative AI ↗
Humlum and Vestergaard, Working Paper 33777. Danish early-outcomes study; revised from Large Language Models Small Labor Market Effects.
- 06
U.S. Census Bureau · 2024 survey vintage
2024 ACS B24114 variable definitions ↗
Detailed occupations for civilian employed people aged 16 and over. Graphic designer estimate and margin-of-error identifiers verified; no count extracted for this brief.
- 07
U.S. Census Bureau · 2024 documentation selected
American Community Survey PUMS Documentation ↗
Variable definitions, user guidance, weights and accuracy documentation organized by release year.
- 08
U.S. Census Bureau · Reviewed 14 November 2026
About Occupation ↗
Explains occupational coding and limits to comparisons across changing classifications.
- 09
U.S. Census Bureau · 2024 instructions
Instructions for Applying Statistical Testing to ACS Data ↗
Methods for margins of error and statistical comparisons. Survey uncertainty is distinct from AI-scenario uncertainty.
Method and corrections history
Edition 1, reviewed 14 November 2026. Selected original research and official statistical documentation were checked for population, unit, outcome and version. This is a source audit and calculation guide, not an exhaustive literature review or independent replication of the studies. The numerical example is hypothetical. No local or demographic job-loss count is claimed. Analysis and recommended questions are labelled; future revisions should identify altered sources, study versions and calculations. Corrections history: initial edition, no prior corrections.
- 14 November 2026 · Edition 1
- First publication of this brief, its connections companion and matching PDF and Word editions. No subsequent corrections are recorded in this edition.
New evidence and substantive corrections will be dated here. Read the corrections policy →