What Trapped Workers Shows About AI and Wages

What is the brief’s answer about wage tradeoffs?

The Bipartisan Policy Center’s brief, “Trapped Workers: What’s the Cost of Leaving an AI-Exposed Job?”, finds a pattern, not a fixed price for leaving: destination groups with larger modeled reductions in AI exposure often have lower occupation-linked pay, but some lower-exposure routes are associated with higher pay. The brief links wage data to 595,000 observed transitions out of highly exposed occupations. It explicitly treats those moves as a map of where workers landed, whatever their reasons, rather than assuming AI prompted a departure or that workers sought less exposure. Its destination comparisons describe groups and occupation-level wages; they do not tell an individual what an available job would pay, or predict whether a particular role will disappear. That distinction matters because exposure measures possible task susceptibility, while actual workplace adoption and job displacement are separate questions. Read the result as evidence that routes through the existing labor market differ in both exposure and pay, with no universal wage penalty for reducing exposure. For a worker weighing a move, the brief is a reason to compare feasible destinations and their real terms, not a personal reason to resign based on an exposure score.

Sources: Trapped Workers: What’s the Cost of Leaving an AI-Exposed Job? • Bipartisan Policy Center

What did the brief measure—and why are the wage percentages easy to misread?

The Bipartisan Policy Center’s August 31, 2026 brief is a United States description of job transitions observed from 2019 through 2025. It links wage data to 595,000 weighted transitions out of occupations classified as highly exposed to AI. The researchers did not establish that AI caused any particular exit, or that a worker moved to reduce exposure. They count the roles people entered, whatever their reasons, as evidence about pathways available from those origin occupations. The results describe labor-market routes and occupation-linked pay, not the wage cost of an AI-driven decision by an identified person.

The phrase “highly AI-exposed” also has a specific meaning in this analysis. For the destination table, the origin occupation’s AI exposure must exceed 50% under the brief’s aggressive scenario. This is a modeled occupation-level threshold. It does not say that an employer has deployed AI, that a particular worker uses it, or that a job will disappear. The International Labour Organization’s research brief “Workers’ exposure to AI: What indicators tell us – and what they don’t” says indices estimate potential task substitution, vary by method, and omit economic and institutional constraints. BPC uses exposure to sort occupations and transitions; it does not measure workplace adoption or individual displacement.

The brief then reports two wage summaries that sound similar but answer different questions. Its transition outcome split classifies each move by comparing wages associated with the origin and destination occupations. A move is “upward” when the destination occupation’s wage is more than 10% higher, “lateral” when the difference falls within plus or minus 10%, and “downward” when it is more than 10% lower. Across the observed transitions in the brief, the shares are 47% upward, 23% lateral, and 30% downward. Those labels describe the distribution of occupational wage comparisons attached to weighted transitions. They are not a survey of each mover’s personal paycheck before and after a change.

The destination comparison asks something else. For each destination occupation group, BPC compares its median income with the median income of the origin pool associated with workers arriving there. The table expresses these differences in real 2024 dollars, adjusted using the BLS CPI-U, and also shows an exposure reduction and a transition volume. It summarizes the destination group against the represented origin-occupation pool. It does not report a uniform wage change for every person entering that destination, and it should not be read as a salary offer or forecast. The transition-band shares and destination medians have different groupings and denominators; one cannot be substituted for the other.

That distinction explains why the aggregate transition split can show more upward than downward moves even while some destination groups sit below their origin pools on the median comparison. The first summary counts weighted transitions according to occupation-wage bands. The second groups transitions by destination and compares group medians. Destination groups receive different volumes of transitions, and the people reaching one destination may come from different origin occupations. A high-volume destination can therefore shape many individual transition records without its destination median describing each record or the entire distribution. These are complementary views, not contradictory estimates of one average wage change.

Weighting is another reason not to read the totals as a headcount of unique people. BPC says the displayed transition shares are weighted, and the destination chart’s volume represents estimated workers using Current Population Survey weights rather than a count of survey records. A transition is the unit represented in the analysis; a worker may be represented in more than one transition over the period. The 595,000 figure should consequently be understood as the brief’s weighted observed-transition base, not as 595,000 distinct people who all left because of AI. The study maps routes, but cannot identify whether moves followed layoffs, promotions, relocation, personal circumstances, or other causes.

Occupation-level pay comparisons also conceal variation inside occupations. Two people with the same occupational classification can have different wages, seniority, hours, locations, employers, benefits, and specialized responsibilities. Assigning an occupation’s wage measure to a transition makes broad comparisons possible, but it does not preserve every individual’s starting pay or destination package. A category label also cannot show whether a reader meets entry requirements or could obtain an opening. Even an accurate group median can be a poor stand-in for one person’s attainable compensation when their experience or circumstances differ from the group.

The ±10% bands add a further interpretive choice. A move with a small gain or loss is called lateral, so the 23% figure does not mean pay was exactly unchanged. Conversely, the upward and downward categories require crossing the selected 10% threshold; they do not capture every increase or reduction as a distinct category. These cutoffs compress a continuous range of wage differences into three bins. The shares depend on the definitions used. They are not estimates of the probability that a worker will personally receive a raise, retain their current income, or take a pay cut after moving.

Read each number against its actual question: the transition bands show how weighted observed moves compare across origin and destination occupation wages; the destination table shows how destination-group median income compares with the relevant origin pool, alongside modeled exposure change and transition volume. Neither table isolates AI as the cause of a move, measures an individual’s realized wage change, or predicts what a reader will earn. That careful reading still leaves the brief useful: it makes pay part of the discussion about pathways, instead of treating lower exposure as an unqualified benefit. It also blocks two tempting but unsupported shortcuts: that nearly half of moves were upward so leaving is safe, or that reducing exposure costs a fixed share of anyone’s salary.

Sources: Trapped Workers: What’s the Cost of Leaving an AI-Exposed Job? • Bipartisan Policy Center; Workers’ exposure to AI: What indicators tell us – and what they don’t

Where does lower exposure come with lower pay—and where not?

The destination comparison shows a broad tradeoff, not a rule that lower exposure always means lower pay. In the Bipartisan Policy Center brief, some groups with large modeled reductions in occupational AI exposure also have median wages well below their origin pool. Yet other groups combine substantial reductions with pay at or above that comparison point. The pattern matters for comparing routes; it does not identify a suitable job for a particular worker or predict that worker’s pay after moving. | Destination group | Exposure reduction | Median pay vs. origin pool | Weighted transition volume | |---|---:|---:|---:| | Food preparation | 46.1 percentage points | -49% | 14.3 million | | Personal care | 37.1 percentage points | -39% | 5.5 million | | Office and administrative support | 11.4 percentage points | -7% | 61.4 million | | Sales | 18.8 percentage points | -1% | 31.5 million | | Computer and math | 18.6 percentage points | +101% | 13.7 million | | Engineering | 27.3 percentage points | +130% | 4.1 million | These pooled 2019–2025 comparisons use real 2024 dollars. The wage column compares a destination group’s median income with the median for its origin pool; it is not each mover’s paycheck change. Exposure is an occupation-group measure, not a count of tasks removed, proof of employer adoption, or a redundancy estimate. Weighted transition volume represents estimated workers using survey weights, not unique job offers available now. Read the rows as past routes and associated occupation-level measures. Food preparation and personal care illustrate the downside emphasized in the brief: their exposure reductions are large, while destination medians sit well below origin pools. But the table resists a simple “more protection costs more” formula. Construction combines a 46.3-point reduction with median pay 2% above its origin pool; management pairs a 33.3-point reduction with a 127% difference. Computer and math and engineering show high pay comparisons with more moderate reductions. These figures do not establish that a worker can enter those occupations, that the wages are offers, or that exposure predicts future stability. The most common destinations sharpen the distinction. Office and administrative support accounts for 61.4 million weighted transitions, the largest volume here, yet its exposure reduction is 11.4 points and median pay is 7% below the origin pool. Sales has 31.5 million transitions, an 18.8-point reduction, and a near-flat -1% pay comparison. Large historical volume does not mean current openings exist nearby or that a route is reachable from every starting occupation. It also says nothing about whether schedule, benefits, commute, and entry requirements fit. Across groups, the relationship is negative and roughly linear: larger exposure reductions tend to accompany weaker wage comparisons. Exceptions matter, and this is not a personal price list. A worker weighing a move should ask which tasks would change, whether the target role is attainable with current skills and credentials, and whether local compensation clears a household pay floor after benefits and working conditions are counted. Compare feasible routes, not abstract exposure scores. A lower-exposure destination may merit consideration, but the table cannot show it is the best or safest choice for anyone.

Sources: Trapped Workers: What’s the Cost of Leaving an AI-Exposed Job? • Bipartisan Policy Center

Is the wage penalty caused by AI exposure?

The brief does not establish that AI exposure causes a wage penalty. Its authors report that the destination wage ranking barely changes when they compare all observed job transitions, rather than only transitions from highly exposed occupations: management and engineering remain relatively high-paying, while food preparation and personal care remain relatively low-paying. That comparison points to an existing structure in the labor market. The apparent cost is the difference between the wage levels of origin and destination occupations, not a price that the analysis can attribute to AI. That distinction matters because the study observes where workers went, not the alternative each person would have experienced by staying. Its 595,000 weighted transitions record moves out of highly exposed occupations during 2019–2025, regardless of whether the reason was a layoff, promotion, relocation, life circumstance, or ordinary career progression. The analysis does not isolate why a transition happened, compare a mover with their own unobserved stay-put outcome, or show that exposure reduction motivated the move. When a destination group has both a larger modeled exposure reduction and lower occupation-linked pay, those measures travel together across groups; their association does not show that reducing exposure produced the lower pay. Nor does it tell us what the same worker would earn in a specific new role. The all-transition comparison is useful counterevidence to a simple “AI discount” explanation. If lower pay were newly created by the act of leaving an exposed occupation, we would expect the destination ranking itself to explain less of the pattern. Instead, the ranking persists across origins. The brief interprets wage differences as structural features of occupations, while arguing that workers starting in relatively well-paid roles may face a larger drop when they enter lower-paid destinations. That is a reason to take route quality seriously if a move becomes necessary, not evidence that AI has newly devalued every lower-paid occupation or is pushing a particular worker toward one. The International Labour Organization’s “Workers’ exposure to AI: What indicators tell us – and what they don’t” adds a related measurement boundary: exposure indices estimate potential task susceptibility, vary by method, and omit constraints such as economic conditions and institutional barriers. The ILO also notes that exposure measures do not incorporate relative wages. So an exposure score cannot answer whether an employer has adopted a system, whether jobs will be displaced, or whether a lower-exposure destination is financially workable. Future AI-related changes in labor demand could affect wages, but this brief does not estimate that future effect. Its wage ranking is a description of observed pathways and occupational pay structure, not a causal estimate or a personal forecast.

Sources: Trapped Workers: What’s the Cost of Leaving an AI-Exposed Job? • Bipartisan Policy Center; Workers’ exposure to AI: What indicators tell us – and what they don’t

What changed over the brief’s 2019–2025 window?

The clearest change is that upward moves became less common after 2023, while lateral moves absorbed more of the transition share. In the Bipartisan Policy Center’s “Trapped Workers: What’s the Cost of Leaving an AI-Exposed Job?”, upward transitions reached 57.1% in 2023, then fell to 43.0% in 2024 and 41.3% in 2025. Downward transitions were 24.3% in 2023 and 31.0% in 2025; lateral transitions rose from 18.6% to 27.7%. These are shares of weighted transitions from occupations the study classifies as highly exposed, sorted by whether the destination occupation’s wage was more than 10% higher, within 10% either way, or more than 10% lower than the origin occupation’s wage. The figures describe the study’s transition series, not the percentage of individual workers who will get a raise or cut if they move. The pattern matters because the narrowing was not simply a surge in the share of transitions into the downward band. Between 2023 and 2025, the upward share fell by 15.8 percentage points, the downward share rose by 6.7 points, and the lateral share rose by 9.1 points. In other words, more recorded moves ended in occupational wage comparisons inside the study’s broad lateral band, while fewer landed above it. That distinction changes the reading of “cost”: the series suggests reduced access to upward occupational wage steps during those years, alongside a larger middle category, rather than a uniform wave of sharply lower-paid destinations. The brief places the timing alongside broad deployment of large language models in professional workflows, but also notes technology-sector layoffs, federal workforce reductions, inflation, and other concurrent shocks. Its data cannot isolate how much, if any, of the change came from AI. A transition may happen for many reasons, and the study does not establish that a worker left to reduce exposure. The report also notes that inflation-adjusted comparisons make the 2023 upward-mobility advantage smaller, though still the peak in its series; the 2024 and 2025 pattern changes little on that measure. So this is useful historical context, not a forecast for 2026 or a verified trend in every occupation. It says that in this sample, the mix of exits changed over the observed years. It does not tell a reader whether their employer is adopting a tool, whether their role will be redesigned, or what a specific offer would pay. Before moving, compare current local wage information, real openings, benefits, and qualification requirements for the roles actually within reach. The period’s change should prompt that check, not supply a personal wage prediction.

Sources: Trapped Workers: What’s the Cost of Leaving an AI-Exposed Job? • Bipartisan Policy Center

Why should a destination’s next-step options matter?

A destination should be judged by the work it may make reachable after the next move. A role can carry a lower AI exposure score than the one a worker leaves and still lead back to the same occupations. The Bipartisan Policy Center’s October 5, 2026 brief, “Trapped Workers: Which Moves Actually Change a Worker’s Options?”, adds this pathway question to the wage comparison. Its analysis covers weighted transitions out of occupations above a 50% exposure threshold during 2019–2026. It reports that 41% of leavers landed in the same occupational cluster as their starting job; from those destinations, the average exposure of reachable onward jobs fell by just 0.2 percentage points. The 59% crossing into another cluster had reachable jobs carrying 6.5 points less exposure on average. These are group-level network results, not a forecast for an individual or a measure of pay.

A cluster is a group of jobs connected by frequent worker transitions in the study’s network, not an official career family or proof that a person can move among every job in it. The distinction is between changing today’s occupation and changing the network of jobs that may follow. A within-cluster move can lower exposure at the destination while leaving later options similar. Cross-cluster moves were associated with a larger reduction in onward exposure, but the study does not show that crossing caused it, that a particular worker qualifies, or that income is preserved. Its weighted transitions are not counts of distinct people, and its modeled exposure indicates task susceptibility rather than employer adoption, displacement, or an individual’s chance of losing work.

The brief finds that cross-cluster routes are concentrated: 57 bridge occupations, 12% of occupations measured, carried 47% of all cross-cluster transitions and 52% of crossings among leavers from highly exposed jobs. These connectors are pathways, not a universal list of accessible or reliably lower-risk jobs. This October 2026 network analysis extends the August wage brief’s question; it does not revise its 2019–2025 wage estimates. The useful inference is that a route’s value may include options as well as starting pay and current exposure. In comparing a stay-and-redesign path, an adjacent move, or a larger pivot, check whether entry is realistic and whether later roles fit the worker’s pay floor, location, schedule, benefits, health, and family responsibilities. A lower score alone cannot answer those questions. A bridge role may merit investigation if it uses existing capabilities and connects to relevant fields, but the study does not establish that training into one will pay off. Nor is a same-cluster move automatically poor: pay, access, hours, and working conditions also matter. Treat the network as a prompt to inspect what could follow the next move, not as a career ranking or safe harbor.

Sources: Trapped Workers: Which Moves Actually Change a Worker’s Options?; Trapped Workers: A Network Analysis of Worker Mobility in the AI Economy

Does retraining evidence contradict the wage tradeoff?

Retraining evidence complicates the wage tradeoff, but it does not cancel it. The New York Fed’s August 2025 staff report, “How Retrainable Are AI-Exposed Workers?”, studies a different question from the Bipartisan Policy Center’s transition brief. BPC describes occupation-linked pay among people observed leaving high-exposure occupations during 2019–2025, without assuming AI prompted those moves. The Fed examines earnings after publicly funded workforce training and compares AI-exposed trainees with similar people who received job-search assistance. Its result shows that training can be associated with higher earnings for some exposed workers; it does not establish that a particular course or credential will make a knowledge worker’s next move pay. The two findings can therefore both be true: existing destination pay structures shape observed mobility, while some training participants gain earnings relative to a comparison group.

The Fed report assembles more than 1.6 million training participation spells in U.S. Workforce Investment Act and Workforce Innovation and Opportunity Act programs from 2012 through 2023, linking records to occupational AI-exposure measures and earnings before and after training. For workers in AI-exposed occupations, the reported average earnings return to training is about $1,470 per quarter. The study also reports a 29 percent lower earnings return for trainees targeting AI-intensive work than for high-exposure peers who pursued more general training. That comparison is not evidence that AI training lowers earnings in every case: it is a relative result within the study’s participant groups. The report describes the sample as relatively low-income and including displaced workers, and its comparison is matched rather than a randomized assignment to training. The estimate concerns these program participants, not the expected payoff to an individual considering a course or credential.

Timing matters. The New York Fed report says positive earnings returns across groups were driven by the most recent years in its study, when labor markets were tightest. That qualification makes the result dependent on hiring conditions; it cannot be carried forward as a reliable return forecast for a different labor market or a specific local occupation. BPC’s wage map likewise does not say that retraining caused destination pay differences. Together, the sources argue against two shortcuts: “avoid retraining because lower-exposure destinations pay less,” and “take an AI course because training raises earnings.” Before paying or committing time, name the target role and the prerequisite you lack. Check current local openings and pay, training cost and duration, the quality of feedback or practice, and whether the credential is actually required. If the skill can be tested through a bounded project or smaller course, that may be a proportionate first step; if the target requires deeper foundations, match the learning path to that requirement. The evidence supports evaluating training against a reachable job and a real skills gap, not treating education as a wage guarantee.

Sources: How Retrainable Are AI-Exposed Workers?; Trapped Workers: What’s the Cost of Leaving an AI-Exposed Job? • Bipartisan Policy Center

How should a reader compare the routes with their own pay floor?

Compare three routes on the same practical terms: redesign the current role, move to an adjacent role that uses experience you already have, or make a larger pivot with explicit checks on prerequisites and training. The Bipartisan Policy Center’s destination figures describe occupation groups, not your offer, household budget, or ability to enter a role. Use them to frame questions, then check the conditions that govern your decision. A route that lowers modeled exposure but breaks a non-negotiable income floor may not be feasible; a route with a higher occupational median may still require credentials, a commute, or a schedule that does not fit. Start with a compact comparison for each route. Write down dependable net pay and benefits, not just a headline wage; the skills and credentials required and which you already hold; openings you can realistically reach in your location or remotely; the time and cost to qualify; and the effects of hours, commute, health, and family responsibilities. Add two work questions: which tasks would change, and what decisions, review, or accountability would remain part of the role? Finally, ask what other occupations become reachable after this move. The October 2026 BPC report, “Trapped Workers: Which Moves Actually Change a Worker’s Options?”, examines onward pathways as well as immediate destinations. Its network findings are a reason to consider future options in the comparison, not proof that any particular route is open to you. For a quick screen, use one row per route and the same columns: pay floor and benefits; entry requirements; local access; time and cost; fit with health and household constraints; task mix and accountable work; and plausible next steps. Mark unknowns instead of filling them with guesses. BLS Occupational Employment and Wage Statistics offers national and geographic occupational wage estimates that can help establish a benchmark. Those estimates summarize occupations; they do not tell you what an employer will offer you, whether an opening is available, or the value of benefits and hours in a particular job. Check those details directly before treating a route as financially workable. The current-role option deserves a concrete test, not an assumption that staying is automatically safer. Ask a manager which deliverables, review duties, and decisions you are expected to own as tools change. Then look for evidence in actual assignments: whether your work is shifting toward checking outputs, resolving exceptions, coordinating handoffs, or taking responsibility for a result. If there is a credible adjacent opening, compare its actual terms with the current role. Reserve a larger pivot for a defined target whose prerequisites and learning burden you can realistically meet. The order can change when circumstances change. A confirmed closure, a viable offer with terms you have checked, or a health or safety need may make a move more urgent. Otherwise, make the comparison bounded and reversible: identify the biggest unknown for each route, verify it, and avoid committing substantial income or training time on the strength of an exposure score. The free task checker can help organize task-level change-pressure signals; it does not estimate your probability of displacement. The paid career roadmap compares scenarios and stated constraints, but it cannot guarantee employment or income. Use either only if it helps structure a decision you have already framed around your pay floor and real access.

Sources: Occupational Employment and Wage Statistics; Trapped Workers: Which Moves Actually Change a Worker’s Options?

What should you do next?

Use one bounded work session to compare options, not decide your whole career. List recurring tasks in your role that are changing, then name one redesign test, one adjacent role that could use your existing knowledge, and a larger pivot only if relevant. For each route, ask whether likely pay meets your household floor, what benefits, hours, commute or location would change, which skills or credentials are required, and how much time and money qualification takes. Include health and family constraints, plus whether local openings exist. The Bureau of Labor Statistics’ Occupational Employment and Wage Statistics provides pay estimates; use them as context, not as a personal offer or full compensation comparison. Check requirements and openings separately. The Bipartisan Policy Center’s “Trapped Workers” brief shows why lower exposure alone cannot choose a route: group wage patterns do not establish what you would earn or whether AI caused a move. A role closure, concrete viable offer, or non-negotiable health or safety need could change the order of options. Otherwise, start with a small redesign test and verify what work you would own. The free [task checker](/ai-job-risk-checker) organizes task-level change-pressure signals; it does not estimate your probability of displacement.

Sources: Occupational Employment and Wage Statistics; Trapped Workers: What’s the Cost of Leaving an AI-Exposed Job? • Bipartisan Policy Center

Questions readers ask