
Job Displacement and Broader Societal Impacts
Employment and causal evidence
From technical exposure to measured displacement
Task, job and worker · Five links in the chain · Exposure is not a probability
What each measure answers
An occupational exposure score asks whether an AI tool could affect some tasks normally performed in that occupation. It can reflect estimated time savings, model ability, or observed use on a provider’s platform. It does not establish whether an employer bought the tool, an employee actually used it, or the use displaced a paid hour. An occupation is a category of related jobs, and a job combines many duties under concrete production and accountability constraints. [38] [39]
Distinguish use from workflow design
A worker who drafts an email with AI is an adopter, but this is not the same as an employer routing all customer interactions through an autonomous tool. Adoption counts measure participation before they measure intensity. Use a second denominator, such as eligible transactions, and record whether human work was supported or bypassed. [40] [41]
Distinguish a staffing choice from its consequences
A company can increase output or quality, idle capacity, hire less, reassign people, lower hours, bring outsourced work in-house, or eliminate positions. National employment reflects further movement across firms and sectors. These outcomes require different records. Payroll shows employment in its coverage universe; departures, vacancies and changes in hours require separate measures. [42] [43]
What qualifies as an attribution claim?
To claim AI caused a job loss, identify the implementation, distinguish it from events such as weak demand or reorganization, show that it preceded the staffing change, and defend a credible comparison for what would otherwise have happened. A press release can support a claim about an employer’s stated rationale, not that comparison. Exposure ranks where to investigate; it cannot substitute for observed adoption, output, staffing and reemployment data. [44] [45]
Global exposure is substantial and uneven
ILO global index · Working paper 140 · Task-based modeling
What the index measures
The ILO’s May 2025 update starts with detailed occupational tasks, asks human respondents and experts about automation potential, and extends its framework to occupations through task scoring and occupational employment shares. The classification separates increasing exposure gradients. Roughly one quarter of workers worldwide are in occupations with some potential exposure, but 3.3% of global employment is in the highest category. Clerical occupations remain highly exposed; newly capable tools also affect some specialized professional and technical tasks. [38]
A large income-level gap
The ILO estimates that about 34% of employment in high-income countries lies in occupations with some potential exposure, compared with 11% in low-income countries. Those are worker shares conditional on occupational structures, not adoption probabilities. As a result, low measured exposure can coexist with serious effects in particular internationally traded services or occupations. [38]
Why job redesign is not job removal
An accounting role may involve producing a draft report, finding documentation, interviewing a colleague, deciding on escalation and accepting responsibility for the conclusion. Drafting and lookup can change substantially even if judgment and verification remain. The ILO considers task transformation more likely than complete replacement in most occupations. This is a qualitative inference, not a statement that employment will never fall. [38]
How to model an exposed occupation locally
List tasks with paid time, repetition, quality threshold, rework, downstream responsibility, acceptable tool use and actual demand. Identify which can be assisted and which, if any, could be delegated while meeting those controls. Compare multiple assumptions about adoption, error correction and output growth rather than multiplying an exposure share by the workforce and labeling the product “jobs lost.” Regional sectors, language fit and institutional arrangements may substantially alter the result. [38] [39]
Gender and geography change the exposure mix
ILO exposure gradients · Employment composition · Distribution matters
Two comparisons, different denominators
For the ILO’s highest 2025 exposure category, its authors estimate 4.7% of female employment and 2.4% of male employment globally, and 9.6% versus 3.5% in high-income countries. A March 2026 ILO brief using harmonized microdata from 84 countries separately compares occupations by gender predominance: 29% of female-dominated occupations were exposed, compared with 16% of male-dominated occupations. The first comparison is worker-weighted within sex groups, while the second is based on occupations. [38] [46]
Occupational composition, not an immutable trait
The ILO links differences to occupational segregation, notably concentration in clerical, administrative and business-support work. This is a statement about jobs and tasks, not ability, preferences, work quality or an individual person's job-loss probability. A within-occupation analysis may reveal further differences in responsibilities, discretion, training and tool access. [46]
A practical distributional test
Compare adoption and staffing by occupation, grade, region and work arrangement before taking demographic averages. For each sufficiently large group, report access to training, task changes, starting pay, hours, promotions, reassignments and departures. Document changes in group composition. Do not use protected-group membership as a proxy for an individual’s automation potential.
Infrastructure can work in both directions
Lower occupational exposure does not mean countries are insulated from competition, nor that higher exposure necessarily means they will lose more jobs. Connection, reliable computing, financing and skills affect whether workers can use AI effectively. Policy should pair reliable access and training with worker input and locally measured outcomes. Comparing countries requires examining different task composition and definitions, not just copying a global exposure percentage into a domestic forecast. [47] [38]
Exhibit 05 · Highest exposure is unevenly distributed
ILO’s 2025 modeled index estimates differences across employment groups. The bands refer to potential occupational exposure to generative AI, not the share of jobs expected to disappear.
International Labour Organization, 2025 index: primary source
The official U.S. baseline as of November 19
BLS October 2026 · November 6 release · Macro context only
Household and establishment surveys
The October 2026 U.S. employment release reported 4.1% unemployment and about 7.1 million unemployed people. Participation was 61.9% and the employment-to-population ratio 59.3%. These come from the household survey. The separate establishment survey initially estimated a 165,000 increase in total nonfarm payroll jobs. These are different survey universes and should not be combined as if each payroll job represented a unique person. [42]
Vintages matter
The same release revised August’s gain to 186,000 and September’s to 60,000. Initial estimates change with additional responses and seasonal revisions; document the release date whenever the numbers are republished. In October, professional and business services, health care and transportation and warehousing added jobs; information showed little change. None of those industry totals reports AI attribution or what happened to a particular occupational grade. [42]
The flows beneath net payrolls
The latest JOLTS release by the cutoff, for September 2026, reported about 7.5 million openings. Openings are a stock on the last business day; hires, quits and layoffs are flows during the month. A flat net payroll count can coexist with substantial hiring and separation, and it can hide falling entry-level offers if senior hiring rises. [43]
How to use the baseline
For national context, compare revisions, participation, hours and hires with the geographic or sectoral question being asked. For causality, add measures of actual adoption and a credible comparison design; the national series themselves do not identify AI as a cause. October payroll growth does not refute localized displacement, and disproportionate weakness in an exposed subgroup does not establish economy-wide contraction.
Realized job loss versus announced cuts
BLS Displaced Worker Survey · [*REDACTED*] job-cut announcements · Different populations
A realized-displacement measure
The BLS Displaced Worker Survey estimated that 7.4 million U.S. workers were displaced in 2023–25, including 3.3 million whose lost job had lasted at least three years; 66.1% of those long-tenured workers were reemployed in January 2026. The measure covers job losses from closings or moves, insufficient work, or abolished positions or shifts. It is about all specified reasons, not AI. It also excludes someone whose first job was never offered. [48]
What reemployment does not tell us
The survey measures employment status at interview, not whether wages, hours, benefits, location or career trajectory were restored. A full measure of transition costs examines unemployment duration, earnings and occupational mobility, disaggregating by tenure and age. Group differences in reemployment are not, without further evidence, attributable to technology. [48]
A much timelier announcement series
[*REDACTED*] logged 617,189 announced cuts by U.S.-based employers through October 2026. AI was cited for 121,401, or about 19.7%, while market/economic conditions were cited for 130,531. In October alone, 1,265 of 43,994 announced cuts cited AI, or about 2.9%. Use the wording “AI cited in announced cuts,” not “AI laid off 121,401 workers.” The source tracks reported announcements and reasons through its own methodology; timing, implementation and attribution require verification. [44]
Why they cannot be summed
An announcement can occur long before an implementation or never materialize; the BLS survey looks back on observed worker experiences. Employers also may cite technology alongside demand or reorganization. These series complement each other as signals of different things; their counts are not interchangeable and cannot be added to estimate either layoffs or AI-caused displacement.
Exhibit 06 · AI-cited cut announcements are not confirmed layoffs
January–October: 121,401 AI-cited of 617,189 announced cuts (19.7%). October: 1,265 of 43,994 (2.9%). The periods overlap. [*REDACTED*] counts employer announcements, not realized layoffs, verified AI causation, or all separations.
[*REDACTED*], October 2026 Job Cut Report: primary source
The early-career employment gap
Stanford and [*REDACTED*] · Data to June 2026 · Descriptive benchmark
What the 19% means
The August 2026 revision of [*REDACTED*], [*REDACTED*] and [*REDACTED*]’s working paper analyzes U.S. administrative payroll data through June 2026. Employment among ages 22–25 in the two most exposed occupational quintiles stood about 19% below what the authors estimate it would have been had it followed growth in the three less-exposed quintiles for the same age group. The authors characterize this as a descriptive indicator. Slower hiring, rather than higher separation, was the principal margin of divergence in their sample. It is neither a 19% fall from the exposed group’s own starting level nor a fraction of jobs known to have been removed by AI. [45]
The changes behind the relative comparison
From November 2022 to June 2026, employment in the more exposed group for ages 22–25 fell about 11% in levels, while the less-exposed group rose about 10%. Their ratio, roughly 0.89 divided by 1.10, illustrates why the relative shortfall is about 19%. This arithmetic compares group trajectories, not the fate of individual workers. [45]
Why the difference is informative
Experienced workers showed no comparable gap; the paper highlights early-career workers in software development and customer support and stronger weakness in occupations where use is characterized more as automation than assistance. Its payroll sample follows a subset of the U.S. economy, and exposure is associated with other features that can shift demand. The paper reports no evidence of widespread displacement in its data, an important limitation on extrapolating the subgroup signal. [45]
Monitor rather than declare a universal effect
Publish the age bin, occupation grouping, start and end dates, and denominators. Track hires, separations and starting pay across observable age and role bands. Triangulate against job postings and representative worker surveys. Avoid presenting estimates from different paper versions as comparable monthly points without verifying that their samples and definitions match. [45] [49]
Exhibit 07 · An early-career gap in exposed occupations
Illustrative indexing of the authors’ reported descriptive comparison. Employment of workers aged 22–25 in highly exposed occupations was about 19% below a benchmark where it kept pace with young workers in less-exposed occupations. No estimate of the number of people laid off or a causal effect is implied.
[*REDACTED*], [*REDACTED*] and [*REDACTED*], August 2026 study revision: primary source
Is a smaller junior share fewer junior jobs?
[*REDACTED*] and [*REDACTED*] · 41-country working paper · Composition versus levels
The international evidence
A November 5, 2026 working paper uses 1.25 billion job postings and 154 million employment records across 41 countries to study foreign affiliates of firms flagged as adopting AI, using AI-related hiring among its measures. The authors’ instrumented analysis finds that the junior share fell by about 1.9 percentage points relative to comparison affiliates; most of the shift reflects growth in senior employment. The main estimate for junior employment is not a statistically precise decline. The data totals describe the research corpus, not numbers of affected workers. [50]
Distinguish share, level and hours
If a team expands from 100 workers (25 junior) to 110 workers (still 25 junior), junior share falls from 25% to about 22.7% without any junior job loss. If the junior category is defined by tenure, experience and job duties, the values will differ from a fixed age band. A national subgroup comparison and foreign-affiliate analysis need not produce the same estimate. [50] [45]
Limits of the design
Identification relies on the event-study comparison and assumptions about the adoption instrument. AI hiring can miss firms deploying vendor tools without hiring specialists, and the results need not generalize to domestic-only firms, small employers, or informal work. Results are still useful: they show how employers can change staffing mix while total employment and some senior roles grow. [50]
Career pipeline implications
A falling junior share can thin the number of future experienced workers even without an immediate fall in junior levels where senior demand keeps rising. Track junior openings and actual starts per cohort, mentoring exposure, promotion eligibility and senior fill sources. Test prospective skill shortages against real hiring demand rather than drawing a mechanical inference from a falling share.
Test competing explanations for youth outcomes
New York Fed · Remote work and vacancy evidence · Confounding matters
Overlapping shocks
Early-career hiring has been affected by the changing interest-rate environment, sector cycles, post-pandemic recruitment, and changes in workplace arrangements. The New York Fed’s June 2026 analysis estimated that remote work could account for 64% of a particular increase in unemployment among young college graduates under its model, arguing timing makes it a plausible explanation for much of that increase. This is an analytical estimate about an observed episode, not a statement that remote work causes 64% of all graduate unemployment or that no later AI effects exist. [51]
Check the denominators
The NY Fed’s third-quarter 2026 update reported 5.2% unemployment and 42.5% underemployment for recent college graduates. Underemployment is a measure of employed graduates in jobs that typically do not require a degree, unlike unemployment, measured among labor-force participants. The feature defines recent graduates as ages 22–27 with at least a bachelor’s degree. Neither number is the unemployment rate for Stanford’s exposed 22–25 subgroup. [52]
Other evidence counts vacancies, not outcomes
New York Fed staff’s May 2026 job-postings analysis includes occupational and aggregate time comparisons by AI exposure. Postings can provide timely signals about employers’ advertised demand, but do not prove hires. Duplicated listings, deleted ads, and shifts across recruiting platforms matter. A payroll comparison and vacancy comparison examine different margins. [49]
Design a stronger comparison
Check pre-existing differences in compressed hiring and remote-work feasibility. Hold workforce grade, industry and location constant where feasible; test changes across different dates and exposure measures; measure visibility and coverage of informal job opportunities. Report how estimates change with those choices. Not all competing explanations must be absent for AI to matter, but a credible analysis should quantify or bound their potential role.
Adoption is widespread but shallow
U.S. worker-reported use · Real-Time Population Survey · Occupations versus tasks
What worker surveys add
[*REDACTED*], [*REDACTED*], [*REDACTED*] and [*REDACTED*]’s nationally representative survey analysis links reports of generative AI use to occupations and work tasks. The authors find at least 20% of workers report using AI in more than 80% of occupations and on more than 40% of job tasks. At least half of workers use AI in only 40% of occupations and for fewer than 3% of job tasks. Occupation use is not task coverage, and breadth at a low threshold is consistent with shallow intensity. [40]
No direct replacement measure
Someone marking a task as AI-assisted can still perform the task, review its outputs or work longer to repair mistakes. Conversely, a company can use automated tools without every employee personally prompting them. Worker reports enable study of use, but replacement needs operational records about human work eliminated, shifted, reassigned or preserved. [40] [41]
Why platform logs differ
Provider logs cover a provider’s users and can map one input to several possible work tasks. A taxonomy can overweight general writing categories shared across occupations, while employees’ survey answers may be inaccurate or sensitive to employer policy. Logs, survey answers and approval records each answer different questions; none captures complete economy-wide work use. [40] [39]
An internal tracking design
Separate organizational authorization, user access, active use, transaction penetration, severity, tasks affected and quality-adjusted performance. Use the same reference period and deduplicate people across tools where practical. Offer private reporting and avoid using monitoring data for unrelated disciplinary evaluation. Where permissions and privacy prevent individual matching, publish aggregate coverage and acknowledge blind spots instead of treating a lack of observed use as zero adoption.
Firm adoption is not the share of workers using AI
Census BTOS supplement · November 2025–January 2026 · Weighting changes meaning
One survey, multiple denominators
The Census Bureau research working paper finds 18% of firms used AI in a business function in the supplement period, or 32% when firms were weighted by employment. Measures for worker task use at firms are 23% and 41% on those respective weighting bases. The employment-weighted rate places greater weight on large employers; it does not identify individual users within adopting firms. [41]
Limited functional breadth
Among adopters, 57% reported applying AI in three or fewer functions. The same paper reports self-reported AI-related employment decreases in about 2% of AI-using firms (firm-weighted). This measures the share of firms reporting an outcome rather than the number of positions; an outcome can be rare among firms and concentrated by size or deployment mode. Examine tables and questionnaire before relabeling a denominator. [41]
Wording matters over time
The Census Bureau broadened the regular question in November 2025 from use in producing goods and services to use in any business function. The levels before and after the wording change are not directly comparable. A conversational assistant used in marketing is plausibly included under the broader wording even when the underlying production system remains unchanged. [53]
Implications for risk monitoring
Pair headline adoption with paid work hours affected, business functions, system version, assisted or autonomous use, spending, and quality. Add geography and establishment size only where sample coverage and support allow. Treat adoption as a process measure; a rise in one or two office uses is not proof of lower staffing requirements or a particular economy-wide output effect.
Productivity experiments establish local effects
Customer support and writing · Two primary studies · Different outcome units
A deployed customer-support assistant
[*REDACTED*], [*REDACTED*] and [*REDACTED*]’s peer-reviewed 2025 study of a staggered rollout finds a 15% average improvement in issues resolved per hour, with effects varying considerably across workers. Lower-skilled and less-experienced workers gained more in this setting. Use the published edition’s 15% rather than silently mixing it with the 14% in the earlier working-paper abstract. This is evidence about assisted support productivity in the studied setting, not 15% fewer positions. [54]
An incentivized writing experiment
[*REDACTED*] and [*REDACTED*]’s randomized experiment with 453 college-educated professionals performing defined writing tasks found a 40% reduction in completion time and an 18% increase in evaluated output quality. The measures reflect selected assignments and the study’s quality evaluation. They are neither annual labor productivity growth nor a reduction in the labor needed for entire jobs. [55]
Unit conversion changes the headline
Time per task is not tasks per hour. If a qualifying task takes 60% as long, repeating only that task without any other costs would raise its theoretical throughput by roughly two-thirds. A queue with review, meetings, compliance, downtime, training and different job duties does not match that assumption. Similarly, a higher score on a rubric cannot be added to a percent time saving.
Translate cautiously into operations
Measure quality-adjusted output, not just how fast a draft appears. Include repeat contacts, escalation, senior oversight, rework, service quality and fixed integration costs. If more output can be sold or if unmet demand is large, output may expand. If demand is inelastic and error control inexpensive, remaining paid hours may shrink. The staffing decision requires evidence about those additional conditions.
Performance varies by task and tool generation
Expert coding · Review and rework · External validity
Coding when context is expensive
In METR’s randomized early-2025 study, 16 experienced open-source developers worked on 246 real tasks in repositories with which they were familiar. AI access lengthened completion time by an estimated 19%, with a confidence interval of 2% to 39% longer in the subsequent methodology update. The study does not represent all developers, repositories or software tools. It illustrates how verification and contextual requirements change measured productivity. [56] [57]
The 2026 follow-up cautions against an easy reversal
METR reported that developers disinclined to work without AI increasingly opted out of its next experiment. It also flagged payment changes and difficulty timing parallel activity. Its authors judge newer tools likely faster in that setting but emphasize weak evidence about the magnitude. Neither paper supports a timeless statement that AI always slows or accelerates coding. [57]
Improvement and failure can appear together
[*REDACTED*] and colleagues’ field experiment with consultants finds improvements on selected tasks at which the tool was capable, while a task outside the tool’s effective range yielded worse performance. Staffing forecasts based on tool speed in an easy benchmark ignore challenge-case error rates and inadvertent reliance on polished but wrong results. [58]
A two-stage procurement experiment
First randomize eligible tasks or time windows while preserving quality review. Then retest on harder task strata, expert users, new entrants, different offices and common edge cases. Count downstream repair and reviewer time even when they occur outside the software. Estimate effects separately for each major task, model version and oversight configuration. Report whether workers can identify a bad result without cues and whether quality holds at higher volume.
Task change can precede pay and hours
[*REDACTED*] and [*REDACTED*] · Danish linked records · Revised March 2026
A precise finding with limited scope
[*REDACTED*] and [*REDACTED*] link large adoption surveys with Danish administrative records. Their revised working paper finds little effect on earnings and recorded hours on average, ruling out effects larger than 2% over the initial two years after [*REDACTED*]’s launch in the settings analyzed. It documents new tasks and reorganization in generation, oversight and integration. The result does not rule out every small effect, every subgroup change, or larger impacts later. [59]
Why it does not contradict group-specific U.S. evidence
Worker earnings and paid hours at organizations are not entry-level employment by age and exposure group. A company can reshape tasks and pay largely the same staff while reducing new recruitment later. Different institutions, exposed occupations and time windows further limit direct comparison. The Danish paper and the U.S. payroll paper can both identify accurate early features of their respective labor markets. [59] [45]
The missing organizational step
Some tasks are not measured on payroll. Saved time can go to quality work, learning, informal coordination, increased output, or changes in work intensity. A tool can also move costs from a junior task queue to a senior verification queue. To detect those transfers, join process measures to grade-specific staffing, training, case mix and supervisor workload.
Watch for level shifts after broader adoption
A short staffing lag is plausible but not guaranteed. If a deployment matures, look for changes in approved staffing budgets, contractor spend, attrition replacement, average workweek and vacancy composition. Ask whether the firm can absorb or sell additional output. Treat a close-to-zero estimate as a scoped result, not the arithmetic midpoint of all possible long-run outcomes.
Occupational forecasts and exposure ranks
BLS 2025–35 projections · Baseline not an AI scenario · Growth and vacancies differ
The decade baseline
BLS projects total U.S. employment to rise from about 170.3 million in 2025 to 176.2 million in 2035, an increase of 5.9 million or 3.5%. It incorporates many structural forces, including demographics, productivity and technology. It does not hold all other conditions fixed to isolate AI, and the projection universe should not be directly spliced onto the monthly payroll series. [60]
Roles do not move in lockstep
Data scientists are projected to grow about 35%, and the combined group of software developers, QA analysts and testers about 10%, from 2025 to 2035. BLS also anticipates pressure from continued workflow automation on several office and administrative-support occupations. Those are role-specific forecasts under many assumptions, not a list of roles safe from AI. [61] [62] [60]
BLS publishes a separate AI exposure product
Its four relative exposure categories combine external theoretical exposure and observed-use evidence. BLS explicitly warns users not to read a high category as a job-loss forecast, wage forecast, worker replacement estimate or adoption likelihood. A highly exposed occupation can grow because it derives economic value from using the technology, while low-exposure work can decline for unrelated reasons. [39]
Openings are not employment growth
An occupation with little net growth can still post many openings as workers retire or move. A fast-growing, small occupation may add fewer roles than a slow-growing, large one. For training decisions, compare annual openings, net level changes, local employers, entry requirements, typical pay and whether placements actually materialize. Show forecast vintages and use alternative scenarios for major uncertainty. [62] [60]
Exhibit 08 · The nationwide baseline is not an AI forecast
The 2025–2035 BLS baseline projects about 5.9 million net additions economywide. It is not a no-AI comparison, not a forecast for intervening years, and not a count of jobs that AI will create. Values rounded.
U.S. Bureau of Labor Statistics, 2025–2035 Employment Projections: primary source
Gross churn cannot be read as net job loss
World Economic Forum · Survey-based 2030 estimates · All macro drivers
The often misquoted global figures
The World Economic Forum’s 2025 employer survey and formal employment baseline underpin estimates of 170 million jobs created and 92 million displaced by 2030, leaving a net increase of 78 million. The gross estimates reflect several concurrent macro trends, including demography, trade, climate and economic conditions alongside technology. They are not observed events, a forecast for generative AI alone, or identical to counts in the global workforce including informal activity. [63]
A positive net can still hurt
Positive net growth does not imply displaced workers can immediately fill new roles. Time, wages, credentials, geography, child care and access to trusted advice shape transitions. A new computing role does not replace a clerical role one-for-one from a worker’s perspective merely because both sit on a spreadsheet. [63] [48]
How to construct useful scenarios
Start with an employment baseline and separately vary the amount of work technically eligible, adoption and utilization, quality-adjusted labor savings, demand response, and the rate at which worker hours are retained or redeployed. Show gross staffing reductions and gross additions and then net them. Record whether the unit is roles, people, paid hours or full-time equivalents. Avoid creating apparent precision by adding global results from incompatible products.
Interpret a forecast as conditional
Label assumptions near each figure and show which direction the result changes when an assumption moves. Include the possibility of fewer entry roles alongside growing total employment. Track official releases and operational usage against clear thresholds. Update upon measured shifts rather than treating a 2030 employer-expectations exercise as a live national employment report.
From a task trial to a labor-market causal claim
A practical attribution audit · Identification before extrapolation · Observables and checks
Define the intervention and the outcome
Record the workflow, version, responsible team, deployment dates, users and transaction penetration. Predefine labor outcomes: headcount, entry hires, hours, wages, transfers and departures. Do not let an unspecified dashboard label such as “productivity gain” replace a measurable outcome. In a controlled trial, randomizing AI access can establish a task effect; randomizing shifts or units and measuring rework can improve the local production estimate. [55] [56]
Use a credible comparison
If randomization is infeasible, a phased rollout can support a comparison across matched teams. Check parallel trends and changing case mix, seasonality, demand, outsourcing and worker transfers. Confirm untreated teams did not already use equivalent tools. If a firm adopts AI in response to an earnings shock, before-and-after layoffs can reflect the shock, the AI decision or both. [59] [50]
Show what could falsify the story
If entry hires fall identically in low-adoption teams, investigate a hiring freeze. If reviewers’ hours rise enough to offset task savings, document where time moved. If a key result disappears under a pre-specified alternative comparison or revised baseline, report it. Measure outcomes long enough to distinguish pilot effects, learning curves, workforce adjustments and subsequent customer demand.
Limit what gets extrapolated
A team can prove a process result without proving national displacement. For a macro claim, document adoption coverage, composition weights, aggregate demand, new-job creation, transfers and credible alternative explanations. Publish uncertainty around causal estimates and the conditions under which an estimate applies. For administrative descriptive studies, say “associated with,” “attributed by the employer,” or “observed relative to a comparison” as appropriate. [45] [44]
Protect the first rung of the career ladder
Recruiting, learning and promotion · Measure throughput and skill · Illustrative employer response
Do not treat all novice work as expendable
Junior tasks often train workers to notice anomalies, navigate internal systems, understand customers and document decisions. A tool can eliminate repetitive steps while incidentally reducing opportunities to develop judgment. A customer-support rollout showing outsized performance gains for less-experienced employees instead suggests one potentially useful design: provide guided assistance and keep progressively more complex cases under human responsibility. [54]
A concrete redesign
For a help-desk, replace rote drafting with diagnostic practice, compare staff-written solutions with model suggestions, and rotate junior employees through human-supervised escalation reviews. For analysts, preserve independent problem framing and a period of tool-free verification; then allow AI where it helps retrieve or summarize source material. Define which outputs juniors must defend in their own words.
Test whether the redesign works
Track completed practice cases, corrected error patterns, time to independent assignment, quality for hard cases, mentoring hours and promotions. Watch whether senior workers spend more time supervising than predicted and whether promotion increasingly depends on prior experience unavailable to new hires. Report accepted offers and job starts, not just listings. These are recommended controls, not promises of improved employment. [49] [50]
Keep the feedback channel available
Solicit reports of being asked to rubber-stamp uncertain outputs. Make clear which staff have authority to slow production or escalate. Use age- and grade-appropriate comparisons without inferring an individual’s skill from age. Preserve staffing records alongside workflow records so future evaluation can tell whether a falling junior share reflects fewer opportunities, senior-team growth, or role reclassification.
Exhibit 09 · The first rung of the career ladder
If AI absorbs some routine first assignments, the employer still needs a way to build judgment and progress workers. Structured practice and review are examples, not claims that every employer or occupation follows this pathway.
48PEGASUS Research conceptual exhibit. Not empirical data.
Build an employment dashboard with clear guardrails
Federal, project and worker indicators · Quarterly review · Escalate evidence, not headlines
A tiered public record
List labor market status from the latest BLS releases and label reference month and reporting vintage. Add official flows for hiring and layoffs, graduate outcomes, and dated measures of occupational exposure and usage. Tag each series by definition, coverage, unit, geography and whether it reflects observed outcomes, forecasts, intentions or modeled exposure. A dashboard that seems comprehensive but combines incompatible series can mislead more than a brief summary. [42] [43] [52]
A tiered employer record
Use three groups of indicators: operations (volume, quality, time, review), staffing (vacancies, starts, hours, pay, moves and exits), and equity (performance and access by sufficiently large relevant groups). Include worker feedback about workload, discretion and learning that an HR event log may not capture. [41] [46]
Interpret the pattern before acting
If adoption rises but opportunities and quality stay steady, maintain observation while screening small groups for concentrated losses. If entry hiring falls as volume rises and other evidence supports automation, review junior pathways and budget for transition. If adoption rises and error rates rise, pause automated action while preserving the assisted pathway and gather evidence. Follow records rather than forcing a single score to summarize incompatible outcomes.
Documentation and transparency
State what is known, unknown and scheduled for revision. Assign each measure an owner and source, and establish a challenge process for workers or managers who believe the numbers hide changed workloads. Avoid unsupported forecasts. National indicators, research results and employer pilots are complementary: none by itself establishes a position eliminated by AI or determines how the benefits will be shared.
Browse: White Papers · Reports · Downloads
