The Last Time We Retrained the World on Purpose

The strongest warning comes from Stanford's Digital Economy Lab. Using 3.5 million to 5 million monthly ADP payroll records through September 2025, the researchers found a 6% raw employment decline after late 2022 for workers aged 22 to 25 in the most AI-exposed occupations. Their regression-adjusted estimate was a 16% relative decline compared with the least exposed group. More experienced workers in the same occupations were stable or growing.

That is unusually strong evidence because it observes payroll, not intentions. It is still observational. The authors control for firm shocks, exclude technology firms and remote-capable work in robustness checks, and find the decline concentrated where AI is more likely to automate than augment tasks. They do not claim to have run the impossible experiment of observing the same economy without generative AI.

Two other datasets add shape without proving causation. SignalFire's 2025 talent report, built from a platform tracking 650 million public professional profiles, found new graduates were 7% of Big Tech hires and less than 6% of startup hires. PwC's 2026 Jobs Barometer analysed 2.4 million US entry-level postings and found the most AI-exposed roles were seven times more likely to ask for traditionally senior skills such as judgement and leadership.

PwC also found that these seniorised entry-level roles grew 35% from 2019 while other entry-level roles shrank 10%. That complicates the clean collapse story. The first rung may be changing shape rather than disappearing everywhere. The danger is that the surviving rung asks people to arrive with capabilities they once learned on the job.

Counterevidence matters. A Danish study linking adoption surveys to administrative records found no detectable near-term effect of chatbots on earnings or recorded hours across 11 exposed occupations, despite reported productivity gains. The honest conclusion is narrower than either the doom or boom narrative: early-career pressure is visible, uneven, and not yet a settled economy-wide causal fact.

The career clock assumes that education happens first and work follows. The skill clock now moves with products and tasks. The institution clock moves through budgets, accreditation, procurement, and political approval. A transition becomes painful when the second clock outruns the other two.

PwC's 2026 analysis covers more than one billion job advertisements across 27 countries and territories. It finds that the skills requested in the most AI-exposed jobs are changing more than twice as fast as in the least exposed roles. Jobs requiring specific AI skills grew 69% against 9% for the overall job market, and carried an average wage premium of 62%.

Those are job-ad and wage associations, not proof that a course in AI produces a 62% raise. They show that the destination can be valuable while the route into it is unstable. The market is rewarding scarce combinations of technical and human judgement, but its vacancy language can change faster than a degree or national curriculum.

The World Economic Forum's Future of Jobs Report turns employer expectations into a useful scale model. Out of every 100 workers, surveyed employers expect 59 to need training by 2030. They expect to upskill 29 in their current roles, reskill and redeploy 19, and leave 11 without access to the training they need. It is a forecast from employers representing 14.1 million workers, not an observed global outcome.

The transition conversation usually begins with tuition. The data says time is the harder constraint. OECD's 2025 adult-learning report, drawing on the 2023 Survey of Adult Skills, finds that roughly 40% of adults participate in learning each year across participating countries. Only 8% enter formal learning. Non-formal job-related learning reaches 37%, but 42% of those activities last one day or less.

Among adults who encountered a barrier, 48% named work or family time as the main reason they could not participate. Cost accounted for 13%. Suitability, prerequisites, timing, or location together accounted for 14%. A free course does not solve childcare, rent, or the foregone wages of six months away from work.

Participation is also regressive. The people who already have education, income, and employer support are more likely to receive more learning. Training offered through a company naturally reaches current employees, not the displaced worker outside its walls or the graduate who never entered.

This is the sharpest lesson from Singapore's model. The SkillsFuture Level-Up Programme gives citizens aged 40 and above a S$4,000 midcareer credit and, for eligible full-time training, an allowance equal to 50% of average income over the previous 12 months. The allowance has a S$300 monthly floor, a S$3,000 monthly cap, and a 24-month lifetime limit. The innovation is not another course catalogue. It is acknowledging that time has a price.

Training can work and still fail at the last mile. Employers may teach one language in public and hire in another. Harvard Business School and Burning Glass Institute used job postings and 65 million US career histories to study 11,300 roles before and after employers removed degree requirements. Their skills-based hiring study found that only 3.6% of roles dropped a requirement during the study window. Within the affected roles, the share of non-degree hires rose by about 3.5 percentage points on average.

Across the market, the authors estimate a net 0.14% increase in non-degree hiring: about 97,000 incremental opportunities among 77 million annual hires, or fewer than one in 700. Among firms that did remove requirements, only 37% produced real and sustained hiring change.

The problem was not that people lacked certificates. Applicant tracking, manager habits, interview design, referral networks, and career ladders still treated a degree as a proxy for trust. Removing two words from a job description did not replace the machinery behind them.

Evaluated career-pathway programs show the same need for modest claims. A US Department of Labor meta-analysis of 46 evaluations found large gains in educational progress and industry-specific employment, but only small gains in general employment and short-term earnings, with no meaningful medium- or long-term earnings gain on average.

The average hides programs that combine more of the transition system. Year Up's randomized evaluation found average earnings roughly 30% higher for treatment-group members across six years. By contrast, the national WIA randomized study found benefits from intensive career services but no conclusive incremental effect from simply gaining access to WIA-funded training over 30 months. The contrast points toward integrated sector training, coaching, paid experience, and employer demand rather than a voucher alone.

Work-based learning is promising because it produces evidence employers already understand: work completed under real constraints. Yet the Department of Labor's evidence synthesis still finds mixed results across program types and populations. Paid experience is a stronger bridge than a certificate alone, not a magic spell.

The Servicemen's Readjustment Act was signed on June 22, 1944, before the war ended. Veterans Affairs records report 7.8 million education and training beneficiaries under the Second World War program: about 2.2 million in colleges or universities and 5.6 million in other training. The same VA history records $14.5 billion in education and training payments.

The bill did more than pay tuition. It offered a living allowance, unemployment support, and loan guarantees, while colleges expanded and the postwar economy absorbed graduates. Research finds substantial educational effects, but the clean claim that one law built the middle class mistakes a system for a single lever.

Access was never equal. Segregated colleges had limited capacity, local administration enabled discrimination, and Black veterans were often excluded from education and housing benefits in the South. Recent NBER work on the GI Bill and racial inequality reinforces why a universal entitlement on paper can widen inequality when delivery institutions are not universal in practice.

Four pieces made the transition mechanism credible: an entitlement people could plan around, living money, institutions able to absorb demand, and an economy with jobs for the newly trained. Remove any one of them and the bridge ends halfway across.

No country has already solved the AI transition. What exists is a library of partial mechanisms, measured with very different standards.

Singapore directly prices the time constraint. Germany's dual system puts paid productive work inside training. Denmark historically spends heavily on active labour-market policy and combines flexibility with income protection, yet OECD reviews also find that some programs have weak or negative effects. France's CPF demonstrates administrative scale, but course starts are easier to find than sustained earnings. India has built enormous skilling and AI infrastructure, while official audits continue to expose gaps in verification and placement data.

Time horizon changes the verdict. A meta-analysis of 207 active labour-market program studies found that effects often improve two or more years after participation. Classroom and occupational training generally look better with time, while public-sector job creation performs poorly. A dashboard that stops at course completion or six-month placement can miss both delayed gains and delayed failure.

The pattern is not that governments fail and markets work, or the reverse. Programs become legible when they report the whole chain: who entered, who finished, who was hired into relevant work, what they earned after 12 and 24 months, and which groups were excluded.

A serious transition policy would not promise to predict every job AI will create or remove. It would insure movement while the evidence catches up.

Governments would fund time as well as tuition, publish provider-level completion and earnings outcomes, and attach workforce conditions to large AI subsidies. The objective would not be more course starts. It would be more people crossing into durable work without losing housing, healthcare, or family stability on the way.

Employers would rebuild paid entry pathways. If routine junior tasks are automated, companies need a new place where judgement can be learned under supervision. Apprenticeships, fellowships, rotations, and subsidised first jobs are not charitable extras. They are the mechanism by which an experienced workforce continues to exist five years from now.

Founders, researchers, and investors would connect tools to live demand while remaining honest about scope. A product tracking 2,500 selected companies can surface questions and examples. It cannot declare what the whole labour market is doing. Broader claims need broader evidence, and the underlying sources should remain inspectable.

Capital can finance assessment, skill maps, matching, and work-based delivery. It cannot replace social insurance for a parent or displaced worker who needs income during the crossing. The investable layer and the public layer need each other.

We do not yet have a clean causal estimate of generative AI's employment effect across the economy. Adoption is recent, exposure measures differ, interest rates and post-pandemic hiring matter, and employment adjusts more slowly than software usage.

Employer forecasts are not outcomes. Job advertisements are not jobs filled. Professional profiles miss people outside their platforms. Payroll studies see employment but not every task inside it. Program enrolments do not equal completion, and completion does not equal earnings.

The evidence also cuts in both directions. PwC finds headcount growing faster at companies most able to use AI. The Danish payroll study finds no near-term earnings or hours shock. Stanford finds a concentrated early-career decline in exposed occupations. These results can coexist because they measure different populations, periods, and mechanisms.

That uncertainty is not a reason to wait. It is a reason to build reversible programs, publish outcomes quickly, and refuse grand claims. The GI Bill was legislated before the war ended, but its lesson is not that planners predicted everything correctly. It is that they treated transition capacity as infrastructure.

The bill for this transition is still unwritten. The clock, as established, is not the slow one.

This report synthesises two bounded Parallel ultra2x research runs completed on July 12 and July 24, 2026, followed by direct review of the primary sources linked below. Evidence is classified by what it observes: payroll and administrative records, job advertisements, professional profiles, employer forecasts, or evaluated programs. These categories are never merged into one causal estimate.

Statistics were retained only when their population, period, method, and limitation could be stated. Forecasts are labelled as forecasts. Associations are not presented as experimental effects. Program inputs such as budgets, credits, and course starts are kept separate from outcomes such as completion, placement, retention, and earnings.

The landed. tracker covers roughly 2,500 selected, mostly AI-relevant companies. It is a narrow product and research universe, not a representative sample of the global labour market. No landed. count or ratio carries a broad market claim in this report.

Evidence cut: July 24, 2026. Monetary figures are nominal values reported by their sources unless stated otherwise.