The End of Work—or the Beginning of Purpose?

Ask people what they do and most of them answer with a job title. That habit is recent, and it is not obviously stable. If machines come to perform a rising share of economically useful labor, the question is whether identity — the sense of being useful, needed, and good at something — can survive the decoupling of livelihood from employment, or whether the two are so tightly joined that the loosening does real damage.

The most common answers are both too confident. The utopian version says the end of compulsory labor is a liberation into art, family, and contemplation. The dystopian version says human worth collapses the moment the paycheck does. Neither is a finding. Both are predictions resting on assumptions about what work actually does for a person, and about how fast the economy can change — assumptions that have partial empirical answers, and partial answers are more useful than a slogan.

What follows separates what labor economists have measured from what they have modeled from what remains genuinely speculative. The measured evidence is more interesting than either slogan, because it shows that the danger is not mainly unemployment. It is the quiet erosion of the three things that make work worth doing, which automation can take away while the job itself remains.

Work is not mainly about the money

Start with the finding that most resists the standard economic picture. In a paper building on self-determination theory, Nikolova and Cnossen used three waves of the European Working Conditions Survey across thirty countries to ask what actually predicts whether people find their work meaningful (Nikolova & Cnossen, 2020). They found that autonomy, competence, and relatedness — having discretion over how you work, being good at it, and having real relationships with colleagues and the people you serve — explained roughly 60 percent of the variation in work meaningfulness. Extrinsic factors such as income, benefits, performance pay, job insecurity, and hours were about 4.6 times less important. Relatedness, the social dimension, was the single strongest factor.

The consequences were behavioral, not just attitudinal. A ten-point increase in work meaningfulness was associated with about a third of a working day less absenteeism per year and an intended retirement age about 2.5 years later. Meaning, in other words, is not a soft add-on to the employment relationship; it shows up in the numbers that employers and governments care about.

Cassar and Meier made the complementary argument from the other direction, surveying the experimental economics literature on nonmonetary incentives and proposing that work be modeled as a source of meaning rather than purely a disutility traded for income (Cassar & Meier, 2018). They point to evidence that people will accept substantial pay cuts for work with a mission, that tasks framed as purposeful are performed in greater quantity, and that unemployment imposes psychological costs far exceeding the lost wages. That last point is the one that should govern the whole debate. If a job were only a paycheck, losing it would be a temporary income problem. The measurement says it is much more than that.

What broad adoption actually looks like so far

Against that backdrop, the current data are more sober than the headlines in either direction.

Bick, Blandin, and Deming documented how quickly generative AI has spread: by their estimates, about 39 percent of United States adults aged 18 to 64 had used a generative AI tool, roughly a quarter had used one for work in the prior week, and about 9 percent used one daily (Bick, Blandin & Deming, 2024). Adoption outpaced the personal computer and matched the internet’s early trajectory. But the intensity is the crucial number: they estimated that generative AI assisted only about 1 to 5 percent of work hours, with time savings equivalent to roughly 1.4 percent of total work hours. A technology can be culturally ubiquitous and still, for now, occupy a small slot in the economy’s actual labor input.

That small slot is not evenly distributed. Adoption is higher among younger, more educated, and higher-wage workers, and about one in five blue-collar workers reported using it. The pattern matters for the identity question because it means the first people to have their work reshaped are also, by and large, the people with the most resources to adapt.

Displacement, productivity, and reinstatement

The most useful framework for thinking about what happens next is not “jobs destroyed” but the task-based accounting that Acemoglu and Restrepo formalized. They distinguish three effects that operate at once (Acemoglu & Restrepo, 2019). The displacement effect removes tasks from human hands. The productivity effect makes the remaining production cheaper, raising demand and hence demand for labor elsewhere. The reinstatement effect creates entirely new tasks for people to do — and this third one is why occupational categories rarely vanish outright. Their historical estimate is that about half of the employment growth from 1980 to 2015 came in occupations whose job titles or task compositions had substantially changed.

That accounting has a crucial asymmetry. Displacement is fast, visible, and concentrated in identifiable people. Reinstatement is slow, diffuse, and often arrives as a modification of an existing occupation rather than an obvious new one. So the lived experience of a transition can be much worse than its aggregate statistics, even if the aggregate statistics are accurate.

The aggregate statistics are also less spectacular than enthusiasts claim. In a task-based macroeconomic model, Acemoglu estimated that plausible AI-driven total factor productivity gains would be modest over a decade — on the order of 0.66 percent, and lower still if AI’s advantages concentrate in tasks that are easy to learn but hard to convert into economy-wide productivity (Acemoglu, 2025). He also found no evidence that AI would reduce labor income inequality, and raised the possibility that it widens the gap between capital and labor. This is a modeling result, not a measurement, and it depends on assumptions about diffusion and task composition that could prove wrong. But it is a useful anchor: the claim that machines will imminently perform a large majority of economically useful labor is not something the current evidence supports. It is a possibility that policy should prepare for, not a trend the data already show.

The evidence from robotization: what automation does to a job that survives

Here the research gets directly relevant to identity, because it studies workers who kept their jobs and lost something else.

A study across fourteen industries in twenty European countries from 2005 to 2021 examined what industrial robot adoption did to the psychological quality of work. Using instrumental variables, the authors found that doubling robotization was associated with about a 0.9 percent decline in perceived work meaningfulness and a 1 percent decline in autonomy (Nikolova, Cnossen & Nikolaev, 2024). Projected to the level of the most automated industries, that becomes a decline of roughly 6.8 percent in meaningfulness and 7.5 percent in autonomy. The associations with competence and relatedness were also negative but less robust.

Two details in that study deserve more attention than the headline.

First, the effects were concentrated among workers whose jobs were routine. People doing repetitive, narrowly specified tasks lost the most autonomy — which makes sense, because those are exactly the tasks a robot or algorithm can take over, leaving the human with less variety and less discretion.

Second, and more hopeful, workers who used computers as tools for independent work maintained their sense of autonomy and competence even in heavily robotized industries. The authors interpret this as control being the moderating variable: it is not automation as such that hollows out a job, but automation that removes the worker’s discretion. A person directing a machine keeps agency. A person feeding a machine loses it.

This is the single most transferable finding for the AI question, where the tasks being automated are cognitive rather than physical. If AI takes over the parts of a job that required judgment, the effect on identity will be worse than if it takes over the parts that required drudgery. That distinction is a design choice, not a law of technology — and it is exactly the choice that firms and professions are making right now, mostly without naming it.

The guaranteed income experiment and the limits of the utopian case

If machines did most of the work, one plausible future is some form of unconditional income. The largest randomized test of that idea in the United States gives an unusually detailed answer about what people do when paid without working.

Vivalt and colleagues randomized 1,000 low-income adults to receive $1,000 per month for three years, with 2,000 controls receiving $50 (Vivalt et al., 2026). The transfer, roughly a 40 percent increase in household income, reduced total individual earnings excluding the transfer by about $1,900 per year and lowered labor market participation by about 4.2 percentage points. Participants worked 1 to 2 hours less per week, and their partners reduced their hours by a comparable amount. Among categories of time use, the largest increase went to leisure. Job quality did not improve — the confidence intervals could rule out even small gains. Degree attainment did not change significantly. Subjective well-being rose in the first year, then reverted to control levels.

Three readings of this result are defensible, and they lead to different conclusions about the future.

The first is a caution about the utopian case. If money without work produced a flourishing of civic, creative, and educational activity, this experiment should have shown at least a hint of it. It showed a moderate labor supply response, a shift toward leisure, and no durable improvement in subjective well-being. The authors’ own summary — a moderate effect “that does not appear offset by other productive activities” — is worth taking at face value.

The second reading is more charitable and also supported. Leisure is not nothing. The participants were low-income adults, most below twice the federal poverty level, and the study measured three years in a society that offers little infrastructure for meaningful unpaid work. The volunteers who showed up to be studied may also differ from the population in ways no randomization within the sample can correct for. Reading this experiment as a verdict on a post-work society asks it to do something it was not designed to do.

The third reading is the diagnostic one. The well-being gains faded after a year. That pattern is consistent with a well-documented finding in the happiness literature: people adapt to changed circumstances. If the reward were only the money, adaptation is expected. But it raises an unwelcome question for anyone arguing that abundance alone will satisfy people — the first year’s improvement suggests that relief mattered, and the reversion suggests that relief is not the same as meaning.

What “purpose” is and is not

At this point the argument has to be honest about its own limits. The evidence above is about measurable things: earnings, hours, self-reported meaningfulness, autonomy, well-being scores. Identity is not directly measurable in that way, and no randomized trial can establish what a person’s life is for.

What can be said without overreach is narrower. Human beings reliably seek out activity that is structured, effortful, and responsive — where what you do affects an outcome and where someone recognizes the result. Employment is currently the most common institution that supplies all three. That claim is a framework drawn from psychology and from the correlations above, not a proven biological necessity. It is a strong enough pattern to plan around and a weak enough pattern that it would be a mistake to treat any particular institution, including employment, as the only form it can take.

The theological and philosophical traditions say much more, and they disagree with each other. Some hold that human worth is given and therefore independent of usefulness, which would make the whole anxiety misplaced. Others emphasize vocation, stewardship, and the dignity of labor in ways that make the loss of work genuinely weighty. This series treats those as frameworks rather than as conclusions neuroscience can settle, and there is no honest way to adjudicate them here. What can be done is to notice that both the reassurance and the alarm assume the same thing — that work currently carries meaning for most people — and then to ask whether that carrying capacity can be relocated.

The best case

The strongest version of the hopeful argument does not rely on abundance producing contentment automatically. It relies on specific, checkable mechanisms.

If AI disproportionately takes over the rote and the draining — the parts of a job that were already sources of drudgery — then the automation effect on meaningfulness could be positive rather than negative, because it would leave more of the work that requires discretion. The robotization finding does not contradict this; it says the outcome depends on whether the human keeps control of what remains.

If productivity gains are real, they can be spent on things the market currently underprovides: care work, teaching, civic institutions, restoration, research that has no near-term payoff. Acemoglu’s modest productivity estimate is a reason for patience about this, not a refutation, because even a fraction of a percent of total factor productivity compounds.

And if AI genuinely extends expertise, in the way Autor describes — allowing people with foundational training to handle higher-stakes decisions that were once reserved for a small credentialed elite — then the middle of the labor market could thicken rather than hollow (Autor, 2024). Autor is explicit that this is an argument about what is possible, not a forecast, and that outcome depends on institutional choices rather than technological inevitability. But the possibility is real, and it is the opposite of the picture in which everyone becomes a prompt operator.

The demographic context strengthens the case. As Autor notes, wealthy countries are aging fast, and all the people who will be thirty in 2053 have already been born. Barring enormous shifts in immigration, rich economies are more likely to run short of workers than of jobs in the coming decades. That does not protect any particular occupation, but it does make the sudden disappearance of paid work an unlikely near-term scenario.

The failure mode

The failure mode is not mass unemployment. It is a society in which people remain employed while their work loses its structure, discretion, and recognition — a future that costs nothing in the unemployment statistics and everything in the meaningfulness measures.

The robotization evidence points directly at this. Jobs survived; autonomy and meaningfulness fell. The mechanism is easy to see in cognitive work as well as manual work: a professional whose task list is reorganized around supervising model output can have more throughput and less craft. The judgment that used to be the point of the job gets reclassified as an exception to be handled by the system. That is a demotion that never appears on an org chart.

A second failure mode is the pipeline problem. If entry-level work is where foundational skill gets built, then automating it changes who becomes an expert. The three-month patent-drafting trial is a preview: lawyers using AI produced better work, but after three months the juniors had gained nothing on average in unassisted judgment, and their scores fanned out toward both extremes (Autor et al., 2026). AI acted as a springboard for some and a hammock for others. Multiply that across professions and the result is a shrinking supply of people who can evaluate what the machines produce — which is precisely the scarce skill the abundance was supposed to make irrelevant.

A third failure mode is the mismatch between freed time and the institutions available to absorb it. The guaranteed-income experiment is the closest thing we have to direct evidence, and it found increased leisure and no durable well-being gain. If the answer to “what would people do instead?” is “rest,” then the policy problem is not merely income; it is whether a society has built the structures — service, apprenticeship, community, craft, worship, care — through which effort becomes meaningful without a market paying for it. Those structures were not designed to carry the load that employment currently carries, and the evidence does not suggest they can be willed into existence in a decade.

What to measure instead of predicting

The useful response to a question this large is to look for the indicators that would discriminate among the futures, and to avoid the ones that feel informative but are not.

Watch job quality, not just employment. The robotization research gives a method: track autonomy, task variety, and perceived meaningfulness within occupations that adopt AI heavily. If the tasks AI takes over are the routine ones, those measures should rise. If AI takes over the judgment-bearing ones, they should fall. Either result is informative, and neither is captured by headcount.

Watch the entry level. The most decisive question over the next decade is whether early-career workers become more capable or merely faster. That requires measuring unassisted performance after a period of assisted work — the redlining test, not the drafting test. Firms that decline to measure this will not know what they have lost until they need a successor.

Watch the concentration of capability. Autor’s scenario and its opposite differ less in how much AI is used than in who directs it. Where AI extends expertise to more workers, wages in the middle should hold up. Where it concentrates decision rights among those who own the systems, they should not.

Watch the institutions outside employment. Shorter workweeks, civic service programs, and lifelong education are the live experiments in what a less employment-centered life might look like. Their results will say more about the identity question than any model of aggregate productivity, because they are the places where the question is actually being tested on people.

Where this leaves the question

The honest summary is that we do not yet know whether machines will perform a large share of economically useful labor, and we do know something more useful than a prediction about what happens to people when the meaningful parts of their work disappear and something else remains.

Work supplies more than income: average, competence, relationships, and a place in a shared endeavor. Those are the things worth defending, and they are also the things automation can quietly remove without ever eliminating a job. The evidence suggests the decisive variable is not how much work machines do, but how much discretion and recognition people retain in what remains — and that is a choice made in thousands of individual decisions about how to deploy a tool, not a consequence that arrives on its own.

If intelligence becomes cheap and labor becomes less central to survival, purpose will not be delivered by the abundance. It will have to be built deliberately, in institutions that give people real responsibility and real recognition, and it will require taking seriously the possibility that losing work costs more than losing a paycheck. The alternative is a society that is materially richer and quietly less coherent — a result that no productivity statistic would ever show, and that the measures of meaningfulness would.