top of page
Glossy 3D 48PEGASUS Logo.png

Job Displacement and Broader Societal Impacts

Executive findings and interpretation

Executive summary

FIELD NOTE The signal is strongest where task exposure, adoption and changes in hiring or hours can be linked. A poll cannot establish job loss.

The public signal

Seventy-one percent of U.S. adults in Pew’s June 2026 survey said they expect AI to lead to fewer jobs in the United States over the next 20 years, up from 64% in the 2024 comparison. Adults aged 18–29 moved from 61% to 73%, the largest observed increase among the published age groups. In 2026, their view was broadly similar to adults aged 30–64. These are expectations, not forecasts of the labor force. [1]

The underlying economic evidence is mixed. Occupational exposure measures where technology may alter tasks. Payroll studies can detect changes in selected cohorts. Neither provides a national count of layoffs caused by AI. The risk is credible even without such a count: employers can reduce junior hiring or training long before net employment falls in a visible way. [2] [3]

What to do now

Preserve the entry ladder while redesigning tasks. Measure hiring, hours, wages, quality, worker exits, and advancement with adoption records where possible. Analyze both aggregate employment and outcomes by tenure and occupation; workers leaving a field can be hidden by sector growth elsewhere. As schools adopt AI, test what students can do independently, not only what they produce with access to a tool. [4] [9]

The policy question is who captures gains and who pays transition costs. Boards should compare productivity to compensation and monitored workload. Leaders should treat public concern as a signal to improve transparency and participation, not validate it as a measured economic impact. The report reviews public evidence dated through November 19, 2026 and offers decision tools, not a prediction of 2046 employment.

A taxonomy of labor market signals

FIELD NOTE Do not interchange exposure, adoption, output or employment. They have different populations and meanings.

Four different measures

Exposure is a modeled overlap between capabilities and tasks. It can signal where to study transitions; it does not show firms use AI or people lose work. Adoption records whether businesses or workers say they use it under a specified definition. Occasional individual use differs from a workflow embedded in production. Productivity links output to inputs; measuring accepted work is better than counting generated items. Gains at one task do not necessarily survive review, integration or demand limits. [2] [7] [10]

Hiring, hours, separations, pay and net employment measure different labor changes. A smaller intake may limit future entry without reducing today’s stock of jobs. Longer hours or wage uncertainty can worsen job quality without a change in headcount.

What attribution would require

To establish causal displacement, ask what would have happened at the same firms without the adoption. Identify dates, compare credible control groups, test pre-adoption trends, and examine other explanations. Announced restructuring is useful for monitoring, but announced jobs do not necessarily correspond to workers actually laid off, or departures that would not otherwise have happened. [3]

Compare like denominators. A population-wide expectation, a firm-level adoption percentage, an occupation’s exposure score and change in a payroll cohort cannot be averaged into a single risk rate. Include dates, geography, industry, baseline, unit and sample or coverage limits with every statistic. If a measure cannot discriminate among the proposed explanations, label it as a signal rather than causal confirmation.

Exhibit 01 · Exposure is not displacement

An occupation can be exposed without widespread adoption, and adoption can change tasks without eliminating jobs. Identifying an effect on employment requires a defensible comparison, not simply a model-capability score.

48PEGASUS Research conceptual exhibit. Not empirical data.

Three mechanisms can act together

FIELD NOTE Productive technology can substitute for tasks, generate demand, and create new work at the same time.

Substitution and support

A writing assistant might reduce basic drafting time, a human may instead spend more time on review, and the output may be good enough for more teams to offer the service. The net employment effect depends on whether additional demand, oversight and new work offset reduced labor per unit. A company can keep headcount flat while changing who performs the remaining work and who gets a first job. [2] [10]

An organization learns more by measuring actual throughput, defect rates, time spent supervising, and the number of independently completed assignments than by multiplying several vendor claims about minutes saved. Include tasks moved to customers or contractors and time spent repairing mistakes.

Scale and feedback

A higher-quality or lower-priced service may attract new demand; existing demand may also be capped by budgets, customer preferences or regulation. Licensed oversight can preserve roles in some work; a product redesign can shift demand to a different occupation, region or supplier. Any forecast therefore embeds assumptions about adoption, labor reallocation, demand, institutions and technology.

Track what is likely to be reversible. Training a worker to review a tool is easier to reverse than dismantling an apprenticeship pipeline. Running a pilot in one channel has lower exposure than making a permanent hiring cut based on assumed utilization. Stage changes so the team can detect unintended outcomes and compare them to the relevant baseline. [3] [14]

Why the first job matters

FIELD NOTE An entry vacancy delivers both current production and a chance to develop judgment.

Separate stocks and flows

An organization’s headcount is a stock. New openings, applications, hires, transfers, separations and promotions are flows. When organizations cease routine hiring without layoffs, headcount may decline only slowly as people leave. Measures of recent graduate unemployment, underemployment and specific occupation-by-age payrolls capture different parts of that transition. They should be read side by side without treating them as equivalent. [4] [3]

Entry-level opportunities matter for skill accumulation. Routine tasks often provide practice in spotting exceptions, identifying quality and handling feedback from more experienced staff. Automating these tasks may be beneficial if alternative practice develops the same skills, but a better first-pass output is not evidence that novices learned more.

Diagnose before changing headcount

Ask which tasks previously justified junior hiring, how much senior oversight is needed, what skills a novice is expected to master, and how many people advance per cohort. Determine whether a comparable group was affected by a hiring freeze, industry demand changes or an outsourcing decision. Compare duration in training and progression after adoption.

Protect some authentic tasks that increase in complexity, with human feedback and a standard for working without AI. Reinvest a portion of saved time into coaching and paid supervised practice. Compare demonstrated competence, internal transfers and retention after a reasonable follow-up. The case for the program is stronger if it creates a credible internal pipeline, not merely a high completion rate. [12] [11]

How the paper weighs evidence

FIELD NOTE First establish the denominator and the design, then assess what inference it supports.

Primary records

Survey producers support statements about respondent views when wording, sample, field dates and results are available. Official releases support estimates of their stated labor or firm variables, subject to revision. Research trials can support context-specific causal conclusions when assignment and measurement are credible; observational studies identify associations whose interpretation depends on comparison methods. Policy records distinguish a submitted proposal, adopted rule, implementation and measured effects.

The report uses selected public sources through November 19, 2026. It is not a census, replication study or forecast. Dated exhibit captions describe whether a reading is measured, modeled, self-reported or illustrative.

Comparative practice

When a survey uses age groups and another uses generations, do not transfer a ranking without checking cohort definitions and question wording. When a paper revises results, retain the study vintage rather than splicing updates into a synthetic time series. Keep high-income ILO exposure estimates separate from U.S. payroll statistics. Do not combine estimates from unlike denominators into an employment-loss probability. [1] [2] [4]

The examples present possible program designs, not measured success. For consequential decisions re-check the primary figures and learn whether the source’s assumptions, occupation crosswalk, sample or reporting schedule has changed. Identify what evidence would change each decision before interpreting the next release.

A map of societal effects beyond headcount

FIELD NOTE Time, quality, access and bargaining power can change even when employment counts do not.

Worker and household outcomes

An AI tool can reduce repetitive work and make some jobs accessible to people who benefit from translation, speech or alternative interfaces. The same deployment can raise performance quotas, change scheduling, or reduce worker discretion. Quality of employment includes pay, benefits, predictable hours, autonomy, privacy and opportunity for advancement, not merely payroll status.

Families may also face a gap between salary expectations and entry opportunities. Where local employers use different technology or training standards, the burden of adjustment can depend on geography, mobility, household obligations and access to affordable learning. Measure who reaches the new jobs, not only whether they exist. [3]

Education, institutions and trust

Schools must decide whether AI is a supported aid, a supervised tool or excluded for a particular assignment. A tool can give effective practice while weakening later performance when used to bypass the learning step. Evaluate independent competence and accessible alternatives. Public trust also depends on whether workers and students can challenge consequential decisions. [9] [8] [14]

Competitive structure matters because firms may lack effective choices about AI suppliers or share savings unevenly. Productive gains can accrue to consumers through lower prices and reliable service, employees through pay and useful time, or owners and providers through higher margins. The allocation is empirical and institutional, not guaranteed by efficiency.

Read the report as a decision sequence

FIELD NOTE Start with the executive findings, then follow the evidence stream that matches your decision.

For decision-makers

Public officials should start with the survey and political chapter, then compare it with the occupation, hiring and hours evidence. The concentration chapter separates upstream providers from downstream business use and shares of output. The recommendations describe observable triggers and review ownership, not a policy prescription based on a single poll.

Employers should start with the work chapter and use the employer pilot and worker consultation worksheets. Select one workflow and a comparison group. Record quality, hours, employee progression and customer effects. Escalate job changes to the appropriate people and review channels where applicable instead of silently allowing a tool rollout to become a labor decision. [3] [14]

For educators and community leaders

Use the education chapter to compare AI-assisted practice with independent performance, role-based safeguards and access to alternatives. Use the transition worksheets to assess whether a program actually connects people to paid work. Training enrollment alone is not a labor outcome; use placement, retention, post-program earnings and progression where records permit. [11] [13]

References and AI-disclosure notes are at the end. Exhibits distinguish measured statistics from conceptual maps. When citing a statistic, keep the original source, research date and population, alongside any reported uncertainty. No private school, worker or customer records were used for this paper.

Next chapter

Back to the report

Browse: White Papers · Reports · Downloads

bottom of page