Every major AI workforce transition report — from the OECD, the McKinsey Global Institute, the White House Council of Economic Advisers, the World Economic Forum — includes a section on retraining and upskilling. These sections follow a consistent template: identify displaced workers, enroll them in relevant training, connect them with new employment opportunities, measure outcomes. The template is reasonable. The programs built on it are not working at the scale required, and understanding exactly why requires looking past the design documents and into the actual experiences of workers who try to use them.

The gap between the policy paper version of retraining and the field reality is not primarily a funding problem, though funding is always mentioned first. The actual problems are structural — they are baked into the assumptions about who displaced workers are, what time horizons they can operate on, and what kinds of learning environments they can access. Until those assumptions are revised, more money in the existing framework produces modestly better outcomes at significantly higher cost.

Here is what the field data for 2025-2026 shows about who is actually experiencing AI-driven displacement and who is actually completing retraining programs. The displaced workers are disproportionately in their 30s and 40s (not fresh graduates who have the fewest transition costs), living in mid-tier cities (not major metros with dense alternative employment and robust community college systems), and supporting dependents (which means they cannot absorb an 18-month income reduction for training). The workers completing government-funded retraining programs, by contrast, skew younger, are more likely to be in urban areas with good transit access to training centers, and are more likely to be single or in dual-income households.

This is not a coincidence. It’s a selection effect built into program design. Retraining programs that require physical attendance during business hours, that pay stipends of $400-600 per month (well below what a mid-career worker with a mortgage and two children needs to sustain a household), and that assume participants can treat skill acquisition as a full-time job for a year are not programs designed for the workers most affected by AI displacement. They are programs designed for workers who resemble recent graduates.

The Community College AI Workforce Initiative, launched in 2024 with bipartisan support and $2.1 billion in funding, is the most instructive case study. The program offers accelerated certificates in AI-adjacent skills — data annotation, AI model auditing, digital business operations, and AI-assisted manufacturing maintenance. The curriculum design is genuinely good. Industry advisory panels shaped it; major employers endorsed it; the certificates carry real credential value. Enrollment in the first year exceeded projections by 40%.

Completion rates: 34%.

That 34% figure is not a failure of the curriculum. Exit surveys show high satisfaction among participants who completed the program. The attrition happens between weeks four and twelve, when the combination of coursework demand, financial strain, and childcare complexity becomes unmanageable. The single most commonly cited reason for dropout, in the 2025 program evaluation: “I couldn’t afford to keep going.”

Policy papers respond to this data point by recommending higher stipends. Higher stipends are better than lower stipends. They don’t solve the problem, because the problem isn’t primarily financial — it’s temporal. A 38-year-old data entry supervisor in Dayton who has been doing one job for 14 years and is now asked to rebuild her professional identity in 12-18 months is being asked to do something that does not match how middle-aged humans actually relearn. The research on adult learning is fairly clear: skill acquisition in middle adulthood is slower, requires more repetition, and is much more sensitive to stress and cognitive load than skill acquisition in early adulthood. Retraining programs designed around the learning curve of a 22-year-old produce predictable attrition when applied to a 38-year-old who is simultaneously managing a financial crisis.

Germany’s Kurzarbeit model — short-time work schemes that maintain employment relationships during economic transitions — has attracted renewed attention in this context, because it solves the temporal problem by not asking workers to leave their employment entirely before retraining. Workers stay attached to their employer, work reduced hours, receive government-subsidized income support for the remainder, and use the freed time for training. The employer maintains institutional knowledge and has incentive to participate in skill development. The worker maintains income, identity, and organizational context.

The German Kurzarbeit AI Transition program (KfW AI-Qualifizierung), extended to cover AI-driven role changes in 2024, has shown significantly better retention than comparable US retraining programs — completion rates in the 60-70% range, wage outcomes 12-18 months post-training that are roughly 15% above the control group. The critical difference is not that German workers are more disciplined. It’s that the program architecture doesn’t force workers to choose between subsistence and skill development.

The US political economy of this solution is complicated. Kurzarbeit-style programs require employer participation, which requires employer buy-in, which requires that retraining be in employers’ interests. In a tight labor market, it often is. In a labor market where AI substitution is the reason workers need retraining in the first place, the employer whose workflow has been automated has limited incentive to fund the retraining of workers their new AI system was supposed to replace. The alignment problem is real.

There’s a second structural failure that policy papers consistently underweight: geography. The retraining infrastructure — community colleges with relevant AI curriculum, workforce investment boards with employer connections, training facilities with reliable internet and updated equipment — is concentrated in major metros and college towns. Displaced workers in Peoria or Bakersfield or Clarksburg, West Virginia are not facing a retraining program problem in the way that a policy paper describes it. They are facing a local infrastructure problem. The programs that exist for them involve either commuting significant distances or enrolling in online programs, which have completion rates lower than in-person programs for exactly the population most affected by displacement.

The online program completion data is worth stating plainly. For workers over 35 with dependents and full-time caregiving responsibilities, online-only retraining completion rates in 2025-2026 averaged 18-22%. That’s not an argument against online learning as a delivery mechanism — it’s an argument against treating online delivery as a solution to geographic access problems without also providing the support infrastructure (childcare, reliable broadband, coaching, peer community) that makes online learning viable.

What actually works, based on the limited field evidence available? Small-cohort, employer-connected programs with income replacement above 80% of prior wages, flexible scheduling, and active case management. The Manufacturing Institute’s AI Transition program in the Midwest, which works with a consortium of employers who have committed to rehiring graduates, shows 71% completion and 83% employment at or above prior wages six months post-completion. The program is expensive. It costs roughly $14,000 per participant, versus $4,200 for the standard community college certificate approach.

The expense is not waste. It’s the cost of actually working. The question is whether policymakers are willing to pay the real cost of effective retraining rather than the discounted cost of programs that look effective in a policy paper and dissolve on contact with reality.

The policy paper version of retraining is a compelling narrative about human adaptability and institutional support. The field version is a story about childcare logistics, mortgage payments, commute distances, and the cognitive limits of adult learning under financial stress. Both stories are true. Only one of them produces useful policy.

Get the best articles from Think Different

One email a week with practical Mac workflows, useful comparisons and honest tech recommendations.