Work, Wellbeing, and Choice: Empirical Lessons for AI Futures
Stephanie C. Y. Chan, Adam Bales, Katherine L. Hermann, Iason Gabriel
cs.CY
2026-09-10
DeepMind reviews unemployed people, retirees and lottery winners: 85%-89% of winners keep working, and cash transfers like UBI cannot replace work's latent benefits.
Once AI automates a large share of paid work, will people still live well? One camp worries about a "meaningfulness gap": machines take the tasks humans enjoy. The other hopes AI clears out boring jobs so people can spend time on better things. In a recent U.S. poll, 67% said people need jobs for purpose and dignity.
Those debates have mostly stayed in philosophy and slogan-grade policy. Stephanie Chan, Adam Bales, Katherine Hermann, and Iason Gabriel at Google DeepMind pull in populations that already live without ordinary employment: the unemployed, retirees, lottery winners, Gulf-state nationals in public jobs, and financially dependent spouses, across OECD countries, China, India, and the Gulf. The questions are concrete. What does paid work actually give people? How do non-workers fare? What does that imply if AI rewrites employment?
This is a cross-disciplinary review, not a new experiment. The authors searched OpenAlex, Google Scholar, PsycINFO, EconLit, and PubMed, then chased citations both ways. They favored meta-analyses, longitudinal panels, natural experiments (plant closures, lottery wins, policy discontinuities), and RCTs. They say up front they did not use a single systematic-review protocol; selection is judgment-laden and coverage is not exhaustive.
The five groups are built as contrasts. Unemployment maps to involuntary displacement. Retirement maps to a mostly voluntary exit. Lottery wins map to sudden wealth with the option to keep working. Gulf public jobs map to state-provisioned employment. Dependent spouses map to material needs met by someone else. The review tries to isolate three mediators: whether the exit was chosen, whether work's latent benefits have substitutes, and how norms and safety nets press on people.
Wellbeing is measured in pieces. Affective wellbeing is how the day feels. Evaluative wellbeing is a judgment of life as a whole. Eudaimonic wellbeing covers meaning, autonomy, and self-realization. Mental, physical, and cognitive health sit alongside. Job loss hits the later ones hardest; day-to-day mood often holds up.
Unemployment is the worst case. Longitudinal work and plant-closure natural experiments show job loss causally cuts life satisfaction and mental health, and the hit is larger than lost wages. After Danish plant closures, mortality was 79% higher in the year of displacement and still 11% higher 20 years later. Denmark has a thick welfare state; weaker nets would likely look worse. Psychological scarring follows people: a German study using updated methods still found negative effects at least five years after re-employment. Daily affect can be propped up by extra leisure (a time-composition effect). Evaluative wellbeing still falls.
Lottery winners settle the "if money were no object" question. Across U.S. and Swedish studies, 85%-89% keep working after a win. Average prizes ran from under $400,000 in Sweden to $7.2 million in one U.S. study (2026 dollars). Among Americans who stayed on the job, average winnings were still $5.2 million. What they buy is flexibility: Swedish winners take unpaid leave in week-long stretches or cut hours, and winners spend about 11% of unearned income on leisure over a lifetime.
Retirement is the other pole. In large U.S. panels, more than 90% of retirees are satisfied with life. Adults 60 and over score about 7-8 on a 10-point life-satisfaction scale, against a U.S. average of 7.7. Wang's longitudinal tracking finds about 70% hold steady through the transition, about 25% dip then recover, and about 5% improve linearly, often after leaving a stressful job. No subgroup showed a lasting, permanent drop. The samples required surviving about eight years into retirement, so survivor bias is in play. The 20%-30% who felt forced into retirement do not get these gains; their path looks more like unemployment.
Substitutes do not need to look like a full-time job. Formal volunteering shows gains in life satisfaction, purpose, cognition, and physical health at one hour a week, with thinning returns past two hours. Hobbies and light caregiving help too. Intensive grandparenting raises stress, more for grandfathers than grandmothers, in both the U.S. and China. In 2001, Gulf states employed more than 60% of nationals in the public sector (90% in the U.A.E.). Public workers report higher happiness than private-sector peers, but after the oil-price crash, high reservation wages and human-capital gaps made a shift to private jobs painful. Participants in the U.S. Civilian Conservation Corps lived about one year longer and earned 4.6% more over a lifetime.
Choice is a variable of its own. Voluntary retirees have better physical and mental health than involuntary ones, even after controlling for health, income, job traits, and life events. Among mothers, the worst-off are those who stay home but want to work, worse than those who work for money but would rather be home. In a European study, depression fell 23% when unemployed people crossed their country's early retirement age, with no meaningful change in income or routine. After plant closures, the rise in depression among American workers was almost four times the rise among Europeans. After "gray" divorces, women's standard of living drops 45% on average, men's 21%.
| Group | Key number | Read-through for AI |
| Unemployed / plant closures | Denmark: mortality +79% in year 1, +11% at 20 years | Involuntary exit hurts the body; welfare does not cancel it |
| Lottery winners | 85%-89% keep working | Sudden wealth does not empty the office |
| Retirees | >90% satisfied; 20%-30% forced | Voluntary exit is what holds |
| "Retiring from unemployment" | Depression -23%, income unchanged | A change in the social label eases the same days |
| Volunteering | Benefits from 1 hour/week | Latent benefits need not be a full-time job |
For people building AI products or writing transition policy, the paper nails a popular slogan: cash does not replace what a job does. UBI and universal basic capital swap in the manifest benefit (income). They do not restore time structure, status, collective purpose, or cognitive exercise, and they can lock people into long-term dependence on a state or a platform. Gulf public jobs and dependent spouses are two mirrors: life can look fine until conditions change and there is no exit.
Four design rules follow. Give people options during the transition; do not dump them in a single forced exit. Fund a portfolio of substitutes (volunteering, civic roles, state jobs), and for older adults one or two hours a week already moves the needle. Assume the moral "norm to work" will linger in the near term, so uneven displacement will keep stigmatizing those without jobs. Move safety nets fast: scarring starts the day the job ends.
This is a policy review, not a new model. It will not tell you which occupation a model eats next year. It will tell you whether people can live well after that happens.
The authors put the biggest hole in section 8: almost all of this evidence comes from contemporary post-industrial societies, where the duty to work, the available substitutes, and the safety nets are products of an industrial social contract. A future society could look different. If automation arrives soon and hits unevenly, these findings travel better. If a whole society moves into a post-work status together, norms themselves may rewrite, and the extrapolation thins out.
The review is not a systematic review. Search and inclusion reflect author judgment and do not claim to be exhaustive. The FIRE (financial independence, retire early) movement is almost absent. High structural unemployment and novel welfare experiments in poorer countries are flagged, not mined. UBI trials were left out on purpose: participants usually fail the "not working or not needing to work" inclusion rule, and those studies already dominate AI policy talk. Retirement tracking has survivor bias. High Gulf wellbeing may be inflated by wage premiums and light workloads, so it is a weak general proof that guaranteed jobs make people happy. The mortality numbers come from Danish plant closures; jumping from that to white-collar displacement by models takes several extra steps.
None of the five groups is a clean analogue for "AI does most of the work." A lottery win is private luck, not a society-wide jump in productivity. Retirement happens late in life, not at thirty when a model takes the job. The authors use the groups as contrasts. The numbers should not be copied straight into a 2030 scenario.