What work is for
Nearly a century ago, Keynes (1930) predicted that technological progress would compress the workweek and force his grandchildren to confront how to live well without toil. The prediction failed on hours but posed the right question: what is work for when machines can do the task?
One answer comes from a normative tradition that predates modern labor economics. Catholic social teaching, beginning with Rerum Novarum (Leo XIII, 1891), holds that dignity belongs to the worker prior to and independently of the worker's market productivity — work is “for man,” in the formulation of Laborem Exercens (John Paul II, 1981, §6) — and its most recent documents apply the person-first logic directly to AI, calling for deployment accompanied by verifiable protections for employment, retraining, worker participation, and the quality of work (Dicastery for the Doctrine of the Faith and Dicastery for Culture and Education, 2025, §§67–70; Leo XIV, 2026, §§149–156). MacIntyre (1981) supplies a secular version of the same distinction: a practice has internal goods — judgment, craft, service, standards of excellence — and external goods — output, income, status. An AI system can increase an organization's external goods while hollowing out the internal goods that make medicine, teaching, law, or the skilled trades worth practicing. But the litmus test for a deployment is whether it preserves and extends the forms of human agency and excellence internal to the work, alongside whether it raises output.
Vocation is broader than paid employment: caregiving, civic participation, contemplation, rest, and artistic creation also constitute a flourishing life, and human dignity cannot be made conditional on holding a job (Arendt, 1958; Pieper, 1952). And meaningfulness is a property of good work rather than of work as such: oppressive, alienating, or voice-denying jobs can impede flourishing (Veltman, 2016; Yeoman, 2014). In this sense, the objective that follows, particularly for certain religious traditions, is neither employment maximization nor task preservation – it is to expand people's real capabilities to exercise agency, develop skills, contribute to goods they have reason to value, and participate in decisions that shape their working lives. A deployment that automates dangerous, degrading, or needlessly repetitive tasks advances that objective; one that strips workers of discretion, authorship, or voice undermines it, whatever it does to throughput.
The empirical record behind the normative claim and philosophical worldview about work as more than paid labor is unusually strong. Jahoda (1982) distinguished the manifest function of employment, income, from its latent functions: time structure, social contact, collective purpose, status and identity, and regular activity. Identity economics reaches a similar conclusion from a different direction: workers' self-conceptions and occupational commitments influence behavior in ways that monetary incentives cannot capture (Akerlof and Kranton, 2005). Panel studies find that the well-being penalty of unemployment remains large after accounting for lost income, and that adaptation is frequently incomplete (Winkelmann and Winkelmann, 1998; Lucas et al., 2004). Meta-analytic evidence covering hundreds of studies reaches the same conclusion for mental health (Paul and Moser, 2009); the adverse effect of involuntary unemployment is among the most replicated findings in subjective well-being research. The converse also holds: employment in jobs with the poorest psychosocial conditions may confer no mental-health advantage over unemployment and can produce worse outcomes (Butterworth et al., 2011). Work protects well-being when it supplies the latent goods, and fails to protect it when it does not.
Research on work design identifies those goods with some precision. Self-determination theory names autonomy, competence, and relatedness as basic psychological needs associated with well-being and high-quality motivation (Deci, Olafsen, and Ryan, 2017; Van den Broeck et al., 2016). Economists find that autonomy, competence, and supportive relationships explain substantially more variation in perceived meaningfulness than compensation and conventional career incentives (Cassar and Meier, 2018; Nikolova and Cnossen, 2020). A meta-analysis links meaningful work strongly to engagement, organizational commitment, job satisfaction, life satisfaction, and health (Allan et al., 2019). Meaning is a performance variable as well as a welfare variable.
How AI could change what work is for
AI is diffusing rapidly in the workplace, particularly following the emergence of large language models. Workers using AI, at least a little, rose from about one-fifth of U.S. workers in mid-2023 to nearly one-half by early 2026, according to the Gallup Workforce Panel, which asks more than 20,000 workers each quarter (Makridis, 2026a); representative worker surveys find similarly rapid generative-AI adoption (Bick, Blandin, and Deming, 2026). What that diffusion does to the character of work is an open question. The task-based literature explains how a technology can displace labor from tasks, reinstate labor into new tasks, or raise productivity within tasks; the net effect depends on the balance, and the balance is substantially a choice embedded in deployment, organizational practice, and policy rather than a property of the technology alone (Acemoglu and Restrepo, 2019; Autor, 2024). Acemoglu (2025) warns of “so-so automation” that displaces workers without large productivity gains. Autor (2024) argues that AI could instead extend expertise to workers who lack elite credentials and rebuild middle-skill work. These scenarios differ in employment and wages, and they differ just as sharply in what work would supply: the first shrinks discretion and authorship, the second extends them to more people.
The workplace evidence so far leans toward the augmentation reading where AI is deployed as an assistant. A generative AI assistant raised customer-support productivity by 14–15 percent on average, with the largest gains for novice and lower-skill workers (Brynjolfsson, Li, and Raymond, 2025). Noy and Zhang (2023) find comparable compression of performance gaps in professional writing tasks, and Dell'Acqua et al. (2026) document large but uneven gains among consultants across a “jagged frontier” of tasks. At the market level, there is also evidence in favor of augmentation. Using state-by-industry exposure built from pre-2020 occupational composition, Johnston and Makridis (2026) find employment roughly 4 percent higher where AI is more likely to augment workers and no significant employment change where it is more likely to automate tasks. Individual-level studies find modest average labor-market effects to date alongside meaningful heterogeneity (Humlum and Vestergaard, 2025), including declines in employment for early-career workers in exposed occupations (Brynjolfsson, Chandar, and Chen, 2025).
Diffusion also runs through management. Managers' trust, communication, and modeling of AI use are associated with adoption and perceived productivity gains (Makridis, 2026a), and in a preregistered experiment with 2,000 U.S. and U.K. managers, information about AI's labor-displacing potential reduced intended adoption, advocacy, and staffing (Dominski et al., 2026). Whether a deployment augments or automates is partly decided in the middle of the org chart. That is good news for design: choices made by identifiable actors can be studied, codified, and changed.
The prior generation of automation offers a warning about what happens to meaning when no one designs for it. Drawing on worker surveys and industry robotization data for 20 European countries over 2005–2021, Nikolova, Cnossen, and Nikolaev (2024) find a consistent negative relationship between robotization and perceived work meaningfulness and autonomy; their instrumental-variables estimates indicate that doubling robotization is associated with a 0.9 percent decline in meaningfulness and a 1 percent decline in autonomy, with the autonomy losses concentrated among workers in repetitive, monotonous tasks. Whether generative AI repeats or reverses that pattern depends on how it changes discretion, task completeness, skill use, social interaction, monitoring, and worker voice — which is to say, on work design (Kellogg, Valentine, and Christin, 2020; Parker and Grote, 2022; Krakowski et al., 2025). Early AI-specific evidence is consistent with the design account: frequent AI use is associated with greater concern about technological displacement, while stronger workplace practices are associated with less concern (Makridis, 2026e); across 343 U.S. cities, AI job growth is associated with economic growth and subjective well-being (Makridis and Mishra, 2022); and a decade of technological change is associated with greater employee empowerment and life satisfaction, with the relationships strongest in high-trust workplaces (Makridis and Han, 2021). However, these positive associations can mask substantial heterogeneity across, for instance, lower resourced cities and/or workplaces with worse managers.
Making purpose and meaning a design criterion
Measurement infrastructure. Statistical agencies cannot govern what they cannot see, and today they see neither AI use nor its consequences for the quality of work. AI capabilities are embedded in software, cloud services, and organizational routines rather than in countable equipment, so conventional national accounts struggle to observe them (Highfill, Wasshausen, and Prunchak, 2025; Corrado, Haskel, and Jona-Lasinio, 2021). Granularity changes the answer: occupational exposure indices identify where AI could matter (Frey and Osborne, 2017; Eloundou et al., 2024), while realized-use measures identify where it does, and linking the Gallup state-by-industry use measure to LEHD and BEA data reveals stronger post-2020 output paths in high-use cells that a coarser industry-level firm-adoption measure (Bonney et al., 2024) does not detect (Makridis, Highfill, and Samuels, 2026). The blueprint for integrating AI investment and use data into U.S. national statistics (Makridis and Brynjolfsson, 2026) should be extended to intangible outcomes: purpose, autonomy, skill use, and voice measured alongside adoption and output. The Gallup Workforce Panel is a working prototype — quarterly, worker-level, linkable to administrative output data — and complementing surveys with telemetry such as token volumes and platform activity would sharpen the picture further (New, Meyjes, and Carson, 2025). The flourishing literature supplies a validated outcome structure spanning happiness, health, meaning and purpose, character, relationships, and financial stability (VanderWeele, 2017; VanderWeele et al., 2025; Makridis et al., 2025). The Bureau of Labor Statistics and Census Bureau should add realized-use and work-design modules to existing worker and firm surveys, and the BEA should treat AI-era intangibles as a measurement priority on par with the IT revolution of the 1990s.
Causal evidence on deployment design. The existing evidence establishes that outcomes differ sharply across deployments of AI (e.g., organizational structures; Makridis, 2026a); it does not yet establish which design features do the work. Candidate margins include task allocation between worker and model, complementary training, managerial practice, worker voice in system design, and the pricing of labor relative to compute. Field experiments of the kind that identified the customer-support productivity gains (Brynjolfsson, Li, and Raymond, 2025) and the human-centered configuration effects (Krakowski et al., 2025) should be extended to pre-register meaning, autonomy, and skill-development outcomes alongside productivity. The question is answerable with existing methods; what is missing is the commitment to measure the full outcome vector.
Governance that makes flourishing a design criterion. Firms decide deployment design; procurement and policy shape firms' incentives. Public and enterprise procurement contracts can require disclosure of how a system changes task discretion, monitoring, and worker voice — an auditable work-design statement analogous to a security review. Distribution belongs in the same conversation: employee ownership and profit sharing (Makridis, 2026c; Makridis, 2025), compensation for the data that makes systems valuable (Lanier and Weyl, 2018; Makridis and Ammons, 2025), and conventional instruments such as wage floors and collective bargaining are the channels through which productivity gains reach the workers whose effort and information produce them. The trade-off should be named plainly: making meaning a design criterion constrains some deployments that would maximize short-run output. The well-being and performance evidence summarized above is the reason to accept that constraint — meaning is associated with the engagement, commitment, and health on which sustained performance depends (Allan et al., 2019), and the robotization record shows what happens to meaning when no one is watching it (Nikolova, Cnossen, and Nikolaev, 2024).
The near-term AI labor story is likely one of concentrated dislocation with diffuse gains — real productivity improvements distributed unevenly across sectors, places, firms, and career stages, with pain clustered in particular occupations and communities. The China trade shock demonstrated that aggregate gains can coexist with durable, geographically concentrated losses that reshape politics for a generation (Autor, Dorn, and Hanson, 2013; Autor et al., 2020). Whether AI produces a comparable rupture depends on choices that have not yet been made: what we measure, what evidence we demand of deployments, and whether the quality of work — its capacity to supply purpose, agency, and connection — counts among the criteria of success. The tools to make it count already exist. The decision to use them is the open question.
