AI and robots are hollowing out routine low‑skilled roles and exerting downward wage pressure, even as augmentative AI boosts productivity for some — a dynamic that is widening inequality and spurring unions and governments to press for stronger workplace technology governance.

Main Finding

This systematic review concludes that AI and robotics are reshaping labour markets through a task‑selective dual logic: (1) strong substitution of programmable, routine low‑skill tasks (originally concentrated in manufacturing but now expanding into clerical and customer‑facing cognitive work), and (2) conditional capability‑enhancement when firms embed AI as an assistive tool. Net employment effects therefore vary by task composition, firm adaptation, and institutional context. Across settings the dominant distributional consequence is intensified wage polarisation (higher returns for high‑skill roles; stagnation or compression for middle/low‑skill workers), rising workplace monitoring and reduced autonomy for many low‑skill employees, and growing “AI anxiety” with measurable negative wellbeing and behavioural outcomes. Institutional responses (vocational training systems, collective bargaining, regulatory initiatives for algorithmic governance) materially moderate these effects. The literature nevertheless has important gaps—especially limited developing‑country evidence, inadequate within‑occupation/task microdata, underexplored psychological mechanisms, and few rigorous evaluations of reskilling policies.

Key Points

• Substitution vs augmentation
  • AI/robots substitute rule‑based, codifiable, repetitive tasks; generative AI and algorithms have extended exposure into clerical/customer‑service roles.
  • When integrated as assistants (with training and workflow redesign) AI raises marginal productivity and can narrow performance gaps without immediate layoffs.
• Quantified displacement and wage effects (representative findings)
  • Acemoglu & Restrepo: each additional industrial robot reduces employment by ~0.18–0.34 percentage points per 1,000 workers and lowers wages ~0.25–0.5%.
  • Eloundou et al. (2023): ~80% of U.S. workers are exposed to LLM influence in at least 10% of tasks.
• Wage polarisation and firm behaviour
  • Automation increases demand and wage premiums for high‑skill roles (data analytics, algorithm development) while compressing wages and opportunities for middle/low‑skill workers.
  • Firms often prefer hiring external high‑skill talent over investing in training low‑skill incumbents, amplifying internal pay dispersion.
• Job quality and autonomy
  • Increased automation and algorithmic management are associated with greater task standardisation, intensified monitoring, lower perceived autonomy and meaningfulness—especially for low‑skill roles.
  • Outcomes are conditional: collaborative/assistive deployments and supportive institutions can mitigate negative effects.
• AI anxiety and wellbeing
  • AI anxiety (fear of job loss, skill obsolescence, loss of control) correlates with higher job insecurity, burnout, and workplace deviance in multiple studies.
• Institutional and collective responses
  • Strong vocational training systems, union bargaining for algorithmic transparency, and emerging workplace tech rights legislation buffer adverse labour effects.
• Research gaps highlighted by the review
  • Sparse evidence from developing countries, insufficient microtask/within‑occupation heterogeneity analysis, incomplete causal work on psychological transmission mechanisms, and few rigorous (RCT/quasi‑experimental) evaluations of reskilling interventions.

Data & Methods

• Type of study: systematic literature review synthesising domestic and international empirical work over the past decade (peer‑reviewed articles, firm studies, cross‑national analyses, surveys, experiments).
• Principal empirical methods in the reviewed literature:
  • Instrumental variable and regional panel studies (e.g., Acemoglu & Restrepo) identifying causal impacts of robot penetration.
  • Cross‑country productivity and labour‑share analyses (e.g., Graetz & Michaels).
  • Firm‑level panel regressions and case studies (China: Zhou & Jin; Germany: Dauth et al.).
  • Field experiments and longitudinal firm studies of generative AI deployments (Brynjolfsson et al.; Brynjolfsson, Li & Raymond).
  • Occupational task mapping and NLP analysis of job texts to measure task‑level exposure (Eloundou et al.; Henseke et al.).
  • Surveys and large‑scale public opinion data (Pew, Eurofound) and employer surveys (OECD).
  • Micro surveys, experience sampling, preregistered experiments, and panel studies probing psychological outcomes (Yam et al.; Nikolova et al.; Cheng et al.).
• Data sources used across studies: commuting‑zone/regional labour statistics, International Federation of Robotics, firm administrative panels, online vacancies/job descriptions, employee surveys, and experimental/field trial data.
• Methodological limitations noted by the review:
  • Overrepresentation of high‑income countries and manufacturing sectors.
  • Few longitudinal individual‑level datasets linking task changes to long‑term earnings and wellbeing.
  • Lack of randomized/quasi‑experimental evaluations of retraining/reskilling programs and governance interventions.

Implications for AI Economics

• Model and measurement implications
  • Models should operate at the task level (not just occupation) to capture heterogeneous exposure and substitution/augmentation dynamics.
  • Incorporate psychological costs (AI anxiety, reduced autonomy, burnout) and monitoring‑driven productivity trade‑offs into welfare and labour market models.
  • Collect and use microtask data (NLP on job ads, workflow logs) and longitudinal panels to estimate dynamic reallocation and wage trajectories.
• Policy and institutional design
  • Active labour market policies (targeted reskilling, on‑the‑job training) need rigorous causal evaluation; evidence suggests vocational systems can materially buffer displacement.
  • Encourage firm incentives for internal upskilling rather than exclusive external hiring to reduce inequality amplification.
  • Design governance/regulatory interventions to ensure algorithmic transparency, worker participation in deployment, and rights around automated monitoring and decisions.
  • Consider distributional policies (tax/transfer, wage subsidies) to address automation‑driven inequality while preserving productivity gains.
• Organizational strategy
  • Promote AI deployments framed and built as assistive tools with accompanying process redesign and employee involvement to maximize complementary effects and job quality.
• Research agenda priorities for AI economics
  • Expand research to developing and middle‑income countries to assess external validity of current findings.
  • Undertake longitudinal causal studies linking task exposure to earnings, mobility, and wellbeing.
  • Perform randomized/quasi‑experimental evaluations of reskilling programs, collective bargaining clauses on algorithmic governance, and firm‑level AI deployment strategies.
  • Interdisciplinary work combining labour economics, psychology, and organizational studies to unpack the mechanisms of AI anxiety and to design mitigation strategies.

If you want, I can (a) produce a one‑page executive summary tailored for policymakers, (b) extract and format the main quantitative estimates into a short table, or (c) propose specific empirical designs to evaluate a reskilling program or an algorithmic‑transparency bargaining clause. Which would be most useful?