Did US Worker Retraining Reduce Participant Automation Exposure?
Analyzing 23 million US WIOA job-retraining records (2017-2023), this paper finds the program rarely shifts workers into less automation-exposed jobs, with success driven mainly by wage catch-up.
It is a large-scale, data-driven cautionary note for anyone betting on public retraining programs to cushion AI-driven job displacement.
Julian Jacobs · Jordan Canedy — meet the researchers →
Two readings, equal authority
How to choose: The paper’s words is verbatim — use it when you need to quote, or to judge how they write. Plain language is a paraphrase written for comprehension — use it when you want the idea fast. Neither is a summary of the other; they are two doors into the same room.
“This paper evaluates whether the U.S. Workforce Innovation and Opportunity Act (WIOA) supported American worker resilience to technological automation. Analyzing over 23 million WIOA participation records (2017-2023), we introduce the “Retrainability Index,” which measures program outcomes through post-intervention wage recovery and shifts in Routine Task Intensity (RTI). We show WIOA rarely shifts workers into less automation-exposed work, with a significant portion of participants simply returning to their prior field. Successful outcomes driven mostly by wage gains, possibly due to “catch-up” mean reversion, rather than changes in occupation. Outcomes are moderated by a person’s prior occupational skill set and area of work, as well as their local economy. We find evidence that employer led programs—notably apprenticeships—are associated with the highest incidence of success. This suggests the United States’ existing public active labor market programming can support baseline wage recovery for vulnerable populations, but is not well-equipped to support the large-scale, cross-industry labor transitions.”
The authors study 23 million records of Americans who went through the US government's main job-retraining program between 2017 and 2023, checking whether it actually moved people into work that is less vulnerable to automation. They find most participants just cycle back into similar jobs, so where the program does show gains, those gains are mostly higher wages rather than a real shift into safer, less-automatable occupations. Apprenticeships and other employer-run programs stand out as the one type of intervention that reliably helps.
What this paper defines
Every definition below is the paper’s own sentence, with its locator. The plain gloss is a reading aid and is marked as one.
Retrainability Index
“we introduce the “Retrainability Index,” which measures program outcomes through post-intervention wage recovery and shifts in Routine Task Intensity (RTI)”Abstract / Introduction
In plain terms: A composite score combining how much a participant's wages recovered and how much less routine (automatable) their work became after the program.
Routine Task Intensity (RTI) as automation proxy
“we argue that RTI is the correct proxy for the particular form of automation that American workers were exposed to in the period we study (2017 to 2023)”Section 3: Dataset and Index Construction
In plain terms: A measure of how routine and automatable a job's tasks are, used here to stand in for a worker's exposure to (older, computerization-era) automation.
Incidence (Incidenceg)
“The share of participation periods within the group that achieved a positive outcome (defined as an Index score In > 0). This functions as a “headcount ratio” for successful transitions.”Section 6: Subgroup Analysis
In plain terms: The percentage of people in a group whose outcome counted as any kind of success.
Intensity (Intensityg)
“The average index score among only participation periods that achieved a positive outcome. This measures the magnitude or “depth” of the success for positive outcomes.”Section 6: Subgroup Analysis
In plain terms: Among the people who did succeed, how big that success was on average.
Ashenfelter dip / catch-up dynamic
“This pattern aligns with a classic Ashenfelter dip: workers often enter training programs following a transitory earnings decline, implying that a significant portion of their subsequent wage recovery is likely mean reversion rather than a direct treatment effect of the training itself”Section 5: Characteristics of Program Success
In plain terms: People often enter training right after a temporary dip in earnings, so their later wage 'recovery' may just be bouncing back to normal rather than a real effect of the program.
What they actually did
Each step is a synthesis. Open any step to see the paper’s own sentence it was derived from, with its locator — so nothing here floats free of the source.
- Compile administrative WIOA participation records and restrict to substantive, completed interventions.
Trace this step to the paper
“We compile all available PIRL files from Program Year (PY) 2017, Quarter 4 through PY 2024, Quarter 3, treating each distinct period of participation as a unique observational unit.”Section 3: Dataset and Index Construction
- Assign each participation period to a Local Workforce Development Board to capture local labor-market context.
Trace this step to the paper
“each participation period is assigned to a Local Workforce Development Board (WDB), the primary sub-state governance unit coordinating WIOA service delivery”Section 3: Dataset and Index Construction
- Compute pre- and post-program wage change using an inverse-hyperbolic-sine transform.
Trace this step to the paper
“We compute mean quarterly wages pre- and post-program, ignoring missing wage quarters, and construct the inverse hyperbolic sine (IHS) difference in these means.”Section 3: Dataset and Index Construction
- Merge occupation-level routine task intensity (RTI) measures onto participants' pre- and post-program jobs.
Trace this step to the paper
“We then merge occupation-level RTI measures [Acemoglu and Autor, 2011] to characterize the routine task environment of participants’ pre- and post-program employment.”Section 3: Dataset and Index Construction
- Combine wage change and RTI shifts into a single weighted Retrainability Index.
Trace this step to the paper
“The composite score uses a weighted linear combination. Wage change receives 50% weight, while routine cognitive and routine manual task shifts receive 25% each.”Section 3: Dataset and Index Construction
- Reweight the subsample with a calculable index using inverse probability weighting to correct for missing-data selection bias.
Trace this step to the paper
“To correct for this potential bias, we apply inverse probability weighting (IPW) to reweight the calculable subsample so that it resembles the full sample population as defined in Section 3.”Appendix A.2.2: Inverse Probability Weighting
- Train a gradient-boosted decision tree (XGBoost) model to predict positive index outcomes from observable features.
Trace this step to the paper
“To assess the combined predictive power of observable characteristics, we employ a gradient-boosted decision tree model (XGBoost) to predict whether participation resulted in a positive outcome (defined as In > 0).”Section 5: Characteristics of Program Success
- Rank feature importance using Gain, Permutation, and SHAP measures to see what predicts success.
Trace this step to the paper
“We report Model Feature Importance (measured by Gain, which quantifies the improvement in split quality when using each feature), Permutation Feature Importance (which measures the decrease in model performance when feature values are randomly shuffled), and SHAP Feature Importance”Section 5: Characteristics of Program Success
- Estimate causal treatment effects of formal training and registered apprenticeship using propensity score matching.
Trace this step to the paper
“Using nearest-neighbor propensity score matching within exact strata defined by program year, we estimate propensity scores using logistic regression on various participant demographics including pre-program occupation and industry subsector, local workforce board assignment, state, age, and pre-program wages”Appendix D: Causal Estimates of Training and Apprenticeship Effects
Exactly what was run, and how
What they reported — and what they left out
This is a labor-economics paper about a government retraining program, not an AI-model evaluation; it never studies or benchmarks AI/LLM systems, and 'automation exposure' is measured only via an occupation-level Routine Task Intensity proxy, which the authors themselves say does not capture AI exposure.
The numbers they report
Most participants return to work resembling what they left, rather than moving into less automatable jobs.
45% return to prior industry; 27% stay in the same occupation
See it in the paper
“with 45% of all WIOA participants returning to their prior industry of work, and 27% staying in the same occupation.”Introduction (Section 1)
A predictive model of program success performs only moderately, suggesting outcomes depend heavily on unobserved factors.
Accuracy = 72%; ROC-AUC = 0.791
See it in the paper
“The model achieves an overall accuracy of 72% and an ROC-AUC of 0.791.”Section 5: Characteristics of Program Success
Only a small, self-selected share of WIOA participants ever receive formal retraining.
~5% of participants
See it in the paper
“Despite this, only about 5% of WIOA participants receive retraining of any kind, and this sample is highly self-selecting.”Introduction (Section 1)
The typical WIOA participant is middle-aged.
median age = 39
See it in the paper
“Across all WIOA participation in our sample, the median age is 39.”Section 4: Data Summary and Descriptive Analysis
The large majority of participants enter the program already unemployed.
83.9% unemployed at entry
See it in the paper
“Notably, the majority (83.9%) were unemployed at the time of program entry.”Section 4: Data Summary and Descriptive Analysis
Receiving formal training is associated with only a modest bump in the incidence of a positive outcome versus no training.
Incidence: 0.61 (trained) vs 0.58 (non-trained)
See it in the paper
“participants who received substantive training interventions achieved an incidence score of 0.61, representing a modest improvement over the 0.58 baseline observed in the non-training population.”Section 6: Subgroup Analysis
In matched causal estimates, registered apprenticeship shows a meaningfully positive effect on the Index, larger than general training.
ATT = 0.038 (95% CI: 0.032, 0.045); ATE = 0.044 (95% CI: 0.033, 0.054)
See it in the paper
“An ATT of 0.038 is a modest, but notably larger effect in magnitude than the near-zero effects observed for general training receipt.”Appendix D: Causal Estimates of Training and Apprenticeship Effects
In matched causal estimates, receiving formal training in general shows a small negative effect on the Index.
ATT = -0.009 (95% CI: -0.011, -0.008); ATE = -0.007 (95% CI: -0.008, -0.005)
See it in the paper
“the estimates of average treatment effect on the treated (ATT) and the average treatment effect (ATE) for receiving formal training are negative, but small in magnitude suggesting that after conditioning on observable characteristics, receiving formal training is associated with a slightly lower index score relative to not receiving formal training.”Appendix D: Causal Estimates of Training and Apprenticeship Effects
Lower-wage states show higher incidence of 'success' than wealthy, high-skill states, consistent with a ceiling effect.
Alabama 0.67, Mississippi 0.63, Arkansas 0.65 vs New York 0.46, Massachusetts 0.47
See it in the paper
“states with lower baseline costs of living and historically lower wages—such as Alabama (0.67), Mississippi (0.63), and Arkansas (0.65)—often achieve higher incidence rates than wealthy, high-skill economies— such as New York (0.46) and Massachusetts (0.47)”Section 6: Subgroup Analysis
Younger participants succeed far more often than older participants, an 'age penalty.'
Incidence 0.75 (age 15-19) declining to 0.50 (age ~60)
See it in the paper
“Incidence rates peak among early-career participants—reaching 0.75 for 15-19-year-olds and stabilizing around 0.65 for those in their early 20s—before declining steadily to 0.50 by age 60.”Section 6: Subgroup Analysis
Registered Apprenticeship and Customized Training have the highest incidence rates of any training service type.
Incidence = 0.71 for both; InW = 0.10 and 0.17 respectively
See it in the paper
“We again find evidence that Registered Apprenticeships and Customized Training record the highest incidence rates (0.71 for both) alongside strong wage intensity (InW = 0.10 and 0.17, respectively).”Section 6: Subgroup Analysis
Low-income participants show both higher incidence and much larger wage gains than non-low-income participants, consistent with a catch-up dynamic.
Incidence 0.63 vs 0.56; InW 0.16 vs 0.08
See it in the paper
“Low-income participants record a higher incidence rate (0.63 vs. 0.56) and double the wage intensity (InW = 0.16 vs. 0.08) of their non-low-income counterparts.”Section 6: Subgroup Analysis
Structural program factors predict success slightly better than participant demographics do.
Average feature-importance rank: 6.2 (program/structure factors) vs 10.3 (demographics)
See it in the paper
“The analysis indicates that structural program factors possess slightly higher predictive power than demographic characteristics, with average ranks of 6.2 and 10.3 respectively.”Section 5: Characteristics of Program Success
Formal trainees see a clearly larger wage-growth premium than non-trainees.
InW rises from 0.10 (non-trainees) to 0.17 (trainees)
See it in the paper
“the “training premium” remains distinctly associated with a stronger wage trajectory (InW rises from 0.10 for non-trainees to 0.17 for trainees).”Section 6: Subgroup Analysis
Even trained participants see only a small reduction in how routine/automatable their work is.
Share with reduced routine-task intensity rises from 0.22 to 0.39 (cognitive) and 0.29 (manual)
See it in the paper
“Even with formal training, the share of participants who experienced a reduction in the routine nature of their work increased only slightly from 0.22 to 0.39 and 0.29 for routine cognitive and routine manual tasks respectively.”Section 6: Subgroup Analysis
What they assert, beside what they showed
Left is the claim in the paper’s own words. Right is the data offered for it. Where the two do not fully meet, a gold band names the distance.
WIOA rarely moves workers into less automation-exposed work; most just return to their old field.
“We show WIOA rarely shifts workers into less automation-exposed work, with a significant portion of participants simply returning to their prior field.”
“with 45% of all WIOA participants returning to their prior industry of work, and 27% staying in the same occupation.”
Introduction (Section 1)Employer-led programs, especially apprenticeships, are associated with the most successful outcomes.
“We find evidence that employer led programs—notably apprenticeships—are associated with the highest incidence of success.”
“We again find evidence that Registered Apprenticeships and Customized Training record the highest incidence rates (0.71 for both) alongside strong wage intensity (InW = 0.10 and 0.17, respectively).”
Section 6: Subgroup AnalysisSuccessful outcomes are driven mainly by wage gains, likely reflecting mean-reversion 'catch-up' rather than a real change in occupation.
“Successful outcomes driven mostly by wage gains, possibly due to “catch-up” mean reversion, rather than changes in occupation.”
“This pattern aligns with a classic Ashenfelter dip: workers often enter training programs following a transitory earnings decline, implying that a significant portion of their subsequent wage recovery is likely mean reversion rather than a direct treatment effect of the training itself”
Section 6: Subgroup AnalysisThe predictive feature-importance results should not be read as showing any feature causes better outcomes.
“We emphasize again that this exercise is predictive rather than causal and do not claim that any feature identified as important by the model necessarily has a causal effect on retraining outcomes.”
“The model achieves an overall accuracy of 72% and an ROC-AUC of 0.791.”
Section 5: Characteristics of Program SuccessReceiving formal training, once matched on observables, is associated with a slightly lower Index score than not receiving it.
“the estimates of average treatment effect on the treated (ATT) and the average treatment effect (ATE) for receiving formal training are negative, but small in magnitude suggesting that after conditioning on observable characteristics, receiving formal training is associated with a slightly lower index score relative to not receiving formal training.”
“We caution that the matching diagnostics for both analyses are mixed—while balance improves substantially after matching, residual imbalances persist for several covariates and propensity score overlap is limited—and these estimates should therefore be interpreted cautiously and in conjunction with the broader observational evidence presented in Section 5.”
Appendix D: Causal Estimates of Training and Apprenticeship EffectsWIOA functions mainly as a 'second chance' wage-floor program rather than an upskilling engine for the labor transition.
“we argue that WIOA acts primarily as a “second chance” program, offering the most vulnerable participants a potential wage floor.”
“it serves chiefly as a kind of ’second-chance’ intervention aimed at re-introducing participants to stable employment, rather than a means of supporting upskilling.”
Section 4: Data Summary and Descriptive AnalysisHow they frame it, and what they want next
Their framing
The authors repeatedly separate what their data can and cannot support: predictive feature-importance and matching results are labeled non-causal or 'suggestive rather than definitive,' and the paper's headline finding (wage gains without real occupational change) is explicitly attributed to a mean-reversion 'catch-up' dynamic rather than a genuine treatment effect. They deliberately bound their conclusions to WIOA as it currently operates rather than to active labor market programs in general, and connect their backward-looking, computerization-era findings to the AI transition only as a speculative extension, not a direct finding.
Register: The authors write cautiously throughout, routinely qualifying causal-sounding claims with words like 'suggests,' 'possibly,' and 'may,' and explicitly labeling their predictive and matched estimates as non-causal or suggestive rather than proven.
Where they hedge
“We emphasize again that this exercise is predictive rather than causal and do not claim that any feature identified as important by the model necessarily has a causal effect on retraining outcomes.”Section 5: Characteristics of Program Success
“We caution that the matching diagnostics for both analyses are mixed—while balance improves substantially after matching, residual imbalances persist for several covariates and propensity score overlap is limited—and these estimates should therefore be interpreted cautiously and in conjunction with the broader observational evidence presented in Section 5.”Appendix D: Causal Estimates of Training and Apprenticeship Effects
“We insist that these findings describe WIOA as it currently operates, not ALMP or retraining programs in general.”Section 7: Implications for the AI Transition
“While RTI effectively captures computerizationdriven automation, it does not capture exposure to AI.”Section 3: Dataset and Index Construction
What they say it means
- Governments should back private-sector retraining partnerships and apprenticeships over general classroom training.
the paper’s words
“our research suggests governments should support private-sector retraining partnerships and apprenticeships, which have consistently demonstrated the highest incidence of success in both our work and in prior literature.”Section 7: Implications for the AI Transition
- AI itself could be used to help participants navigate WIOA's complex menu of program options.
the paper’s words
“there may be a role for AI itself in helping WIOA participants navigate the system and identify best-fit interventions.”Section 7: Implications for the AI Transition
- Some workers may be better served by non-retraining interventions, such as direct income transfers, rather than by ALMPs at all.
the paper’s words
“whether a person would benefit more from a non-ALMP intervention, such as direct income transfers, or programs that do not require the existence of continued work.”Section 7: Implications for the AI Transition
What they call for next
- Retraining providers should run experiments to isolate the causal effect of programming from self-selection.
the paper’s words
“retraining providers should invest in experiments that can help us better understand the causal impact of programming on outcomes.”Section 7: Implications for the AI Transition
- More research is needed on the potential of future active labor market programming beyond WIOA as it currently exists.
the paper’s words
“More research is needed to understand the potential for future ALMP programming.”Section 7: Implications for the AI Transition
Limitations they state
“It drops individuals who abandon the program, remain persistently enrolled, or exit into unemployment—plausibly the worst outcomes. Consequently, this selection mechanism almost certainly biases our index results upward, meaning the aggregate incidence of success detailed in this paper likely inflates the true share of WIOA participants experiencing positive outcomes.”Section 3: Dataset and Index Construction
“While RTI effectively captures computerizationdriven automation, it does not capture exposure to AI.”Section 3: Dataset and Index Construction
“we are reticent to make causal determinations about formal training impacts on program outcomes using this observational dataset.”Section 4: Data Summary and Descriptive Analysis
“A methodological limitation is that subsector-level routine task intensity is only weakly correlated with the occupation-level task intensities of the specific jobs held by WIOA participants.”Appendix A.1: Subcomponent Details
Moves worth stealing
States the paper's most damaging limitation (survivorship/selection bias that inflates its own headline results) in the methods section itself, immediately after defining the sample, rather than saving it for a caveats paragraph at the end.
“Consequently, this selection mechanism almost certainly biases our index results upward, meaning the aggregate incidence of success detailed in this paper likely inflates the true share of WIOA participants experiencing positive outcomes.”
Directly engages a concurrent competing paper's choice of automation metric and defends its own different choice with an explicit, numbered list of reasons rather than ignoring the disagreement.
“Despite this, AI-exposure—perhaps the dominant empirical proxy for AI labor market impacts—remains an inappropriate tool for our analysis on three grounds”
Immediately re-asserts that a strong-looking model result is not causal, right after reporting the model's performance numbers, to preempt readers over-reading the finding.
“We emphasize again that this exercise is predictive rather than causal and do not claim that any feature identified as important by the model necessarily has a causal effect on retraining outcomes.”
Fully formalizes its custom composite metric with named notation and equations in an appendix, making every step of an otherwise opaque score auditable by another researcher.
“The final individual index is a weighted linear combination of the three components.”
Where else this leads
Same territory
- Research acceleration: The view inside OpenAI OpenAI
labor economics
Published alongside it
The nearest publications in time, across all three labs.
- Where the goblins came from OpenAI
2026-04-29 - HeadVis Anthropic
2026-05-15 - Model Spec Midtraining: Improving How Alignment Training Generalizes Anthropic
2026-05-15 - Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations Anthropic
2026-05-15
What this page was built from
Read in full, including References and Appendices A-E with all data tables (manifest text_grade: full); the source .txt carries typical PDF-extraction artifacts (repeated page headers, split-across-line words like 'computerizationdriven' and 'byHyman'), which are preserved verbatim in quotes rather than silently corrected.