Artificial Intelligence and work: a critical review of recent research from the social sciences

• Here, work means formal economic activity producing goods and services for profit or wages, excluding social-reproductive activities from the study’s scope.The authors acknowledge that this is a restrictive definition; social-reproductive activities may also involve AI.Source: a. Definitions and scope · p. 3
• AI’s effects are considered across formal work processes, from coordinating industrial processes and managing workers to evaluating information and serving customers.Source: a. Definitions and scope · p. 3
• Technological development and deployment are shaped by social context, not explained by technical advances alone.The passages connect this perspective to work in philosophy and sociology of technology.Source: b. A critical approach · p. 4
• Companies invest in AI in hopes of reducing labour and transaction costs and increasing output through production rationalization.The cited discussion associates labour-cost reduction with Bessen et al. 2018 and transaction-cost reduction with Gurkaynak, 2019; Lobel 2018.Source: c. The capitalist imperative · p. 5
• Nationalistic pressures link AI development to geopolitical competition, including an alliance of corporate and military interests.The discussion cites US military investment and other powers’ AI research investment (Berman 1992; Morgan et al., 2020; Barton, Woetzel et al. 2017; Allen 2019; Savage 2020).Source: c. The capitalist imperative · pp. 5–6

2. Technological Unemployment.

• Predictions depend on assumptions about tasks and skills, yet “skill” and “routine” are slippery concepts whose meaning shifts across perspectives and work settings.The review notes that social-science discussions often take concepts such as skill for granted, despite their empirical complexity.Source: §2.1 · p. 7
• AI may automate routine activities within professional jobs and high-skill work, while some nonroutine manual tasks resist codification and human labor may complement automated processes (Acemoglu and Restrepo 2018; Gries and Naudé 2018; Delfanti and Frey 2021).Human work may also remain in automated processes because people perform some tasks better or more cheaply than machines.Source: §2.1 · pp. 8–9

2.2.1 Pessimistic scenarios

• The estimates predict that 47% of all jobs in the U.S. are at risk of technological replacement, while 54% of jobs in the EU and USA were under threat (Bowles 2017).Source: §2.2.1 · pp. 9–10

2.2.1. Optimistic scenarios

• Some empirical studies report no significant employment effects from natural language processing or from AI at the U.S. occupation and industry levels in the last decade (Squicciarini and Staccioli 2022; Acemoglu, Autor, Hazell, and Restrepo 2020).Source: §2.2.1 · p. 11
• Arntz, Gregory and Zierahn (2016) estimated much lower automation risk—around 9% across the OECD, 12% in Austria, and 6% in South Korea—after changing how tasks were linked to occupations.Source: §2.2.1 · p. 11
• Economic arguments against universal technological unemployment include productivity-driven labor demand, increased demand for capital, deeper automation without labor substitution, and new tasks created by AI automation (Acemoglu and Restrepo 2017).New-task creation is identified as a key component of labor economists’ arguments against universal technological unemployment.Source: §2.2.1 · p. 12
• In optimistic accounts, automating some production steps can raise the value of remaining human tasks, complementing rather than replacing workers.This complementarity may also produce a multiplier effect in related service occupations and manufacturing industries (Goos, M. et al. 2015).Source: §2.2.1 · pp. 12–13

2.2.3 Critical assessments

• Predictive models may fit past employment trends yet fail to capture future realities, with idealizing assumptions and existing job profiles obscuring new tasks and work’s contextual complexity.Case studies also risk overgeneralization across occupations and national settings, where economic, educational, policy, firm-strategy, and geographic factors differ.Source: a. Methodological doubts · pp. 14–15
• AI deployment also depends on public policy, firm strategy, geography, computational hardware, energy, resource availability, and environmental constraints.Geopolitical and resource limits may hamper deployment, while environmental impacts could constrain it if sustainability becomes a serious economic parameter.Source: a. Methodological doubts · p. 15
• Capitalist profit-seeking may favor increasing output over shedding jobs, while weak labor power and cheap labor can bound automation; the absence of mass unemployment can coincide with precariousness and exploitation.Automation may also shift workers toward low-wage, precarious service and care occupations rather than produce massive job churn.Source: b. Marxist critiques · pp. 16–17

Algorithmic management

• Algorithmic management applies AI to traditional managerial tasks, including hiring, work-process optimization, worker evaluation, scheduling, monitoring, and incentives.Business-oriented accounts also emphasize improved workflows, lower costs, and predictive power.Source: §3 · pp. 17–18

Surveillance and control

• Worker monitoring supplies data for AI and may support security, fraud prevention, or health and safety, but it also increases information asymmetry and managerial power.The review notes that monitoring can be used for apparently benign or benevolent purposes as well as control.Source: §3 · p. 18
• Surveillance can invade privacy at every stage of the data cycle, increase performance pressure and harm well-being, while algorithmic control can diminish worker choice and creativity.The passages connect privacy infringements to managerial leverage and describe reduced respite as a concern in warehouse and platform work.Source: §3 · p. 18

Bias, transparency, and accountability

• Algorithmic management can reproduce discrimination, obscure decision processes and require workers’ blind trust, while distancing organizations from accountability and workers’ avenues for recourse.The cited concerns include algorithmic or managerial bias, opaque decisions, coercive direction, and reduced worker control over remedies.Source: §3 · p. 19

Forms of platform work

• Platform work includes capital platform work, crowdwork, and app-work, which differ in the relationship between workers, platforms, and the physical location of work.Crowdwork is performed via the platform, while app-work provides services in the physical world.Source: §4 · p. 20

Four thematic families

• The review identifies four thematic families in platform-work research, including contrasting views of platforms as enabling exchange or extending control, alongside debates about precarity and institutional context.The review also notes that platform technology may be used for control or for worker protection, while powerful platforms can shape their institutional environments.Source: §4 · pp. 21–22

Hidden labor

• Human labor fills gaps in AI systems and can remain hidden behind claims of automation, a phenomenon Gray and Suri call “ghost work.”The passages identify micro-task crowdwork as one form of this labor, including work in the AI industry.Source: §4 · p. 22

Worker experiences

• Platform workers have varied conditions and experiences, so broad generalizations about platform work are unreliable.The passages describe differences in job quality, health effects, flexibility, autonomy, and control across types of work.Source: §4 · pp. 22–23

Work and social protection

• AI-related unemployment, wealth polarization, or widespread platform employment could place further strain on social protection systems organized around full-time, well-paid jobs.The review notes that crowdwork and app-work are not yet widespread employment models.Source: §5 · p. 23

Limits of policy proposals

• The review describes detailed ideal-scenario policy proposals as unrealistic under current ideological and geopolitical conditions, particularly in countries leading AI innovation.It presents this gap between proposals and political-economic reality as a reason to examine the forces shaping AI development.Source: §5 · p. 24

Politics of AI work

• Political conflicts over AI include disputes about gig-worker legal status, worker privacy, and regulatory frameworks, with powerful companies actively contesting regulatory efforts.The review describes interventions including lobbying, legal challenges, and mobilization of customers.Source: §5 · p. 24
• The review questions whether states will enforce “good AI” principles amid high-stakes geopolitical competition and concludes that ethical considerations may have little sway over some AI extensions.Control of semiconductor supply chains is one example of the geopolitical stakes discussed.Source: §5 · p. 25

Conclusion

• AI does not itself determine good or bad outcomes for work; the world in which it is developed and deployed matters.The review identifies economic and political parameters as shaping those outcomes.Source: §5 · p. 25