Papers  /  Bridges, Not Cliffs: What Actually Helps Workers When Techno…

Bridges, Not Cliffs: What Actually Helps Workers When Technology Displaces Their Jobs

InstanceSomaSoft Research (prepared by Claude Code for the AURI project)
Published2026-11-22
SAGL-1.0 preprint open access
The licence All papers
Cite this paper
SomaSoft Research (prepared by Claude Code for the AURI project). (2026-11-22). "Bridges, Not Cliffs: What Actually Helps Workers When Technology Displaces Their Jobs". SomaSoft Research. https://somasoft.ai/papers/bridges-not-cliffs. Licensed under SAGL-1.0.

SomaSoft is a research project, not a product: no users, no deployments. Nothing here is medical, legal or financial advice. Limits and corrections

Bridges, Not Cliffs

What actually helps workers when technology displaces their jobs

Draft, 22 November 2026. Not peer reviewed. Part of the series introduced in "The Transition Agenda", following "The Power to Say No" and "Beyond the Price."


Reading conventions. [I] marks our own inference or argument. It is not sourced, and it is offered to be disagreed with. Sentences without that label report findings from the studies listed in the references, with the publisher and date given there and in the facts file that accompanies this paper. Where a study's result is mixed, contested, or comes from a single site, we say so in the same paragraph rather than in a footnote. Where we could not verify something, we leave it out.


1. The question

Every wave of automation comes with a promise that displaced workers will be retrained into better jobs. The promise is sincere. It is also usually untested at the moment it is made.

[I] This paper asks instead which kinds of help have been evaluated against a comparison group, what those evaluations found, and how durable the effects were. A programme that sounds right but has not been tested is not a bridge. It is a hope, and machine intelligence may change many jobs at once, leaving little time to learn by trial.

2. How deep the cliff is

Losses are large and long. In their study of high-tenure workers displaced in Pennsylvania between 1980 and 1986, Jacobson, LaLonde and Sullivan (1993) found long-term earnings losses averaging 25 percent per year, still present five years after separation. The losses began before workers left, and they were substantial even for those who found new jobs in similar firms.

Recessions double the damage. Using Social Security earnings records from 1974 to 2008, Davis and von Wachter (2011) estimated that men displaced in a mass layoff when national unemployment was above 8 percent lost about 2.8 years of pre-displacement earnings in present value. When unemployment was below 6 percent, the loss was about 1.4 years.

Unemployment insurance covers the gap, not the fall. In a later review, the same Pennsylvania authors noted that unemployment insurance offsets about half of earnings losses during the roughly six months of eligibility, but that most of the long-run loss comes from re-employment at lower pay, which standard unemployment insurance does not address (Jacobson, LaLonde and Sullivan, 2005).

Regions do not bounce back on their own. Autor, Dorn and Hanson (2021) followed American commuting zones exposed to rising Chinese import competition over 2000 to 2019. The shock plateaued in 2010, yet the more exposed areas still showed lower manufacturing employment, lower employment-to-population ratios and lower income per head a decade later. They estimate that about 86 percent of the manufacturing job loss showed up as a fall in overall employment, not as moves into other work. People mostly did not move away. Government transfers rose modestly, mainly through Social Security and Medicare, and replaced only a small share of lost income.

[I] Taken together, these studies describe a cliff with three features. The fall is steep for people with long tenure in one kind of work. It is steeper when many people fall at once. And places do not climb back without help. Any serious response to AI-driven change has to be judged against all three.

3. What the evidence says about the main kinds of help

3.1 Trade Adjustment Assistance: the long-running American experiment

The United States runs Trade Adjustment Assistance for workers who lose jobs to trade.

The 2012 national evaluation was discouraging in the short run. D'Amico and Schochet's synthesis for the Department of Labor (2012) compared participants under the 2002 amendments with matched unemployment-insurance claimants. Nearly 66 percent of participants received training, against 27 percent of the comparison group. Participants did much worse in the first two years, largely because they were in training rather than working, and spent on average 79.8 weeks in the programme. By the fourth year they had almost closed the gap: fourth-year earnings were between about $3,300 lower and about the same, depending on the comparison group. Over the four-year window, the estimated net benefit to society was negative $53,802 per participant. Participants who received programme-funded training did better than those who received income support alone, whose results stayed negative throughout. The authors cautioned that four years may be too short to see the full return.

A longer, quasi-experimental view is more positive. Hyman (2018) used the effectively random assignment of petitions to more and less lenient investigators, with Census data on about 300,000 workers. Workers whose petitions were approved had roughly $50,000 more cumulative earnings ten years out, from both higher pay and higher participation. But the annual gains faded to nothing over the same ten years.

[I] The two studies are not in conflict. They describe a programme that costs participants heavily at first, can pay back over a decade, and whose gains fade. That is neither the success its supporters claim nor the failure its critics claim.

3.2 Generic retraining

The broader record on publicly funded training is weaker than its political popularity suggests.

The national WIA experiment was inconclusive on training. The Workforce Investment Act evaluation by Fortson, Rotz and colleagues for the Department of Labor (2017) randomly assigned job seekers to different levels of service. It found that individualised staff assistance raised earnings by about $7,100, or 20 percent, over 30 months. On training itself the evidence was inconclusive, and Mathematica's summary said it suggests training was not effective in that window.

Effects take time. Card, Kluve and Weber's meta-analysis of more than 200 evaluations (2018) found average effects of active labour market programmes close to zero in the short run, becoming more positive two to three years after the programme ends, with larger long-run gains for programmes that build human capital.

What is studied matters. For displaced workers in Washington State, Jacobson, LaLonde and Sullivan (2005) found that one academic year of more quantitative community college courses raised later earnings by about 14 percent for men and about 29 percent for women. They found little evidence of gains from less quantitative courses, and the estimated social return to them was negative. They also warned that people who choose to retrain are those who expect to gain most, so a bigger subsidy would draw in people with smaller gains.

[I] "Retraining" is not one thing. Training chosen without regard to demand, taken by people who must stop earning to attend, often fails. Training in specific, quantitative, demanded skills often works, but slowly.

3.3 Sectoral programmes: the strongest randomised evidence

A small group of programmes works with employers in one sector, screens applicants, trains them for specific occupations, and stays with them into the job. These have the strongest randomised evidence in the American workforce field.

[I] Three cautions belong next to these results. First, the gains vary greatly by provider: WorkAdvance had one clear success and three weaker sites. Second, most of these trials served low-income adults and young people entering the labour market, not experienced workers displaced from long careers, so the transfer to AI-driven displacement is an untested inference. Third, all of them depend on employers who actually hire graduates. That is the part an employer deploying AI can supply.

3.4 Wage insurance

Wage insurance pays a share of the difference between a worker's old wage and a lower new one, for a limited time, if they take a new job.

One strong study found it paid for itself. Hyman, Kovak and Leive (2024) studied the American version, Reemployment Trade Adjustment Assistance, which pays up to half the wage gap for up to two years to workers aged 50 or older, without requiring training. Comparing workers just above and just below age 50, with a sample of 76,500 workers from about 1,000 firms, they found eligibility raised employment by 8 to 17 percentage points in the two years after layoff, fading to zero after four. Earnings rose by more than $18,000, or 26 percent, over four years, excluding the subsidy itself. Most of the gain came from shorter spells without work. Taxes on higher earnings and lower unemployment payments fully offset the cost.

[I] This is the best-evidenced idea in this paper per dollar spent, and also the narrowest. It was tested on older trade-displaced workers. The authors themselves note an open risk: at scale, employers might lower wage offers knowing the subsidy exists.

3.5 Short-time work: Germany in 2009 and 2020

Germany's Kurzarbeit lets employers cut hours instead of jobs, with the state replacing part of the lost pay.

Scale. The annual average number of short-time workers in 2009 was about 1.14 million. In April 2020 it was about 6.0 million, roughly 18 percent of the eligible workforce (Federal Agency for Civic Education, 2020; Kagerl, 2024).

Effect. An evaluation by Kagerl at the federal employment agency's research institute (2024) found clear positive employment effects in 2020, most of which faded quickly. Its back-of-the-envelope estimate is that up to half a million jobs were saved, while 8.75 million people, about 3.7 million full-time equivalents, were supported. The author reads the gap as evidence of large windfall effects. Reviewing 2009, the same paper notes that model-based work found short-time work saved jobs but reduced allocative efficiency, and that firm-level evidence for Germany was not clear-cut.

[I] Short-time work suits a temporary fall in demand, not a permanent change in what work is needed. Paying people to stay in a job that will not come back delays the cliff. For AI-driven change it is useful only as a holding pattern while a firm genuinely redesigns roles.

3.6 Danish flexicurity

Denmark combines easy hiring and firing with generous income support and compulsory active labour market programmes. Kreiner and Svarer (2022) argue that the first two parts are not enough on their own, that the activation requirements are what make the model affordable, and that Denmark spends more on active labour market programmes than any other OECD country. Andersen (2021) reports that in a normal year such as 2007, 30 percent of unemployment spells lasted under four weeks.

[I] The model shows that a labour market can be flexible and secure at once. It does not show that the package transplants to countries with different institutions, and much of the evidence on the whole system is descriptive.

3.7 Portable benefits

Where health cover, pensions and leave are attached to a single employer, losing the job means losing them. Utah's SB 233, in force from 3 May 2023, lets companies pay into benefit plans for independent contractors without reclassifying them. In Pennsylvania, DoorDash ran a six-month pilot in 2024 contributing 4 percent of participating couriers' earnings to benefit accounts, and about 4,400 signed up.

[I] We found no independent evaluation of either. The only reported results on the Pennsylvania pilot come from the company. Portable benefits are a sensible principle without outcome evidence yet.

4. What the newest evidence says about AI and jobs

No broad displacement yet. A Yale Budget Lab analysis (Gimbel, Kinder, Kendall and Lee, 2025) found the occupational mix changing more quickly than in the past, but not by a large amount, and the change predates the widespread introduction of AI. Measures of AI exposure showed no sign of being related to changes in employment or unemployment. In Denmark, Humlum and Vestergaard (2025) linked adoption surveys to administrative records and found precise null effects on earnings and recorded hours, ruling out effects larger than 2 percent two years on.

But a gap is opening at the bottom of the ladder. Using payroll data from ADP, Brynjolfsson, Chandar and Chen (2025) reported a 13 percent relative decline in employment for workers aged 22 to 25 in the most AI-exposed occupations, after controlling for firm-level shocks. In their August 2026 update, employment for this group was about 19 percent below where it would have been had it kept pace with less-exposed peers, up from 15 percent a year earlier. The authors call these descriptive patterns, not causal estimates. The gaps shrink when education is accounted for, some differences began before generative AI was widespread, and the gaps are larger in the ADP sample than in national surveys.

Official numbers. The Census Bureau reports that 17 to 20 percent of American firms used AI between December 2025 and May 2026, and 37 percent of firms with 250 or more employees. The New York Federal Reserve reports unemployment among recent college graduates aged 22 to 27 at 5.2 percent in the third quarter of 2026, with underemployment at 42.5 percent.

[I] The picture is early and partial. Where adjustment shows, it is through fewer hires of young people rather than layoffs of experienced ones. Most programmes in section 3 were built for people who lose a job, not people who never get the first one; the companion paper "When the Ladder Moves" takes up that problem.

5. A staged playbook

[I] Everything in this section is our argument, built from the evidence above.

Stage 1: The employer deploying AI, before roles change

  1. Announce early and plan roles before the tools go live. [I] Displacement losses begin before separation. Workers who know a role is changing can start a bridge while still earning.
  2. Fund sectoral training with a hiring commitment, not a voucher. [I] The programmes with the strongest evidence pair training with employers who hire. An employer can commit a number of interviews or hires for graduates of a programme it co-designs.
  3. Pay people while they train. [I] The Trade Adjustment Assistance and Project QUEST results both show early earnings losses while people train. Training during paid hours removes the main reason people drop out.
  4. Prefer quantitative and certifiable skills. [I] The community college evidence favours these strongly.
  5. Use reduced hours as a holding pattern only while redesign is real. [I] Set an end date and a plan, because short-time work saves jobs mostly when the shock is temporary.
  6. Top up lost wages for redeployed staff. [I] An internal version of wage insurance for workers moved to lower-paid roles follows the best-evidenced idea in this paper.
  7. Protect the entry rung. [I] If AI reduces junior hiring, keep a fixed share of entry roles and apprenticeships, and report the number.

Stage 2: The city or region

  1. Build an intermediary that employers trust. [I] Project QUEST and Per Scholas work because they sit between employers and job seekers. A region can fund one per major sector.
  2. Run the first cohort as a trial. [I] Random assignment among oversubscribed applicants is fair and cheap, and it is how the programmes above earned their evidence.
  3. Measure earnings for at least five years. [I] Short follow-up got both Trade Adjustment Assistance and generic training wrong.
  4. Act early in places with fewer college graduates and narrow industries. [I] Those were the places where the China shock did most lasting harm.

Stage 3: National policy

  1. Make wage insurance general, and evaluate it. [I] Extend it beyond trade and beyond age 50 in stages, with a rigorous design, because the only strong evidence is from one group.
  2. Keep short-time work ready, with experience rating. [I] Make firms that use it heavily pay back part of the cost, as the German evaluation suggests, to reduce windfalls.
  3. Fund sectoral programmes on outcomes, not enrolment. [I] Pay for earnings gains measured against a comparison group, and stop funding providers that do not produce them.
  4. Decouple core benefits from the employer. [I] Test portable benefits with independent evaluation before treating them as a solution.
  5. Publish AI-exposure employment statistics by age. [I] The early-career gap was found in private payroll data. A public agency should track it.

6. What this project is doing, and not doing

[I] This project runs no workforce programme and has tested none of the interventions here. What we can offer is the discipline we apply to our own system: a claim is not a result until it has a comparison group. We would welcome partners running sectoral programmes or wage insurance pilots who want help designing an evaluation.

7. Limits

8. Conclusion

[I] The evidence does not support the comfortable story that displaced workers can simply be retrained. It supports a narrower one. Losses from job loss are deep and long. Generic training rarely closes them. Specific training tied to employers who hire can close much of the gap, slowly. Paying people to take a new job sooner can pay for itself. Holding people in jobs works only when the shock is temporary. The employer deploying AI has the most leverage and the earliest warning, so the bridge should be paid for before the jobs change.

References