Machine learning is reshaping work: a systematic review finds AI powering algorithmic management, platform coordination and safety monitoring, but raises urgent transparency and fairness issues under the EU's 2024 AI rules.

Main Finding

This systematic review (Scopus, 2014–2024) finds rapid and growing application of machine learning (ML) and AI across multiple areas of employment research. From an initial 802 hits the authors retained 25 papers and produce a taxonomy showing ML’s roles in (1) organizational performance (OP), (2) occupational safety & health (OSH), (3) personnel selection and development, and (4) algorithmic management (AM). ML is already used for predictive modeling, real-time safety monitoring, “virtual recruiter” interfaces, AI coaching, and causal uplift analysis of safety policies. The review emphasizes that while ML can improve productivity and safety, important gaps remain on fairness, transparency, and empirical field evidence — amplified by the new EU AI regulatory framework (2024) that raises explainability and compliance requirements.

Key Points

• Scope and selection
  • Database: Scopus; time window 2014–2024.
  • Search combination: terms linking employment/AM/OSH/aesthetic labor with ML/AI/NN/SVM/deep learning/generative AI.
  • PRISMA screening: 802 → 47 (title/abstract) → 18 (full text) + 7 (secondary citations) = 25 retained.
• Domains covered (majority to minority): Organizational performance (largest share), OSH, personnel selection, employee satisfaction, algorithmic management, personnel development, unemployment prediction.
• Predominant research methods in the reviewed corpus:
  • Supervised ML model development and application (majority).
  • PLS-SEM survey-based impact studies (many papers discuss applicability and perceptions).
  • Conceptual frameworks, literature reviews, some field experiments and case studies.
• ML techniques observed:
  • Supervised: logistic/regression variants, Naive Bayes, KNN, SVM, decision trees / random forests, gradient boosting, BART, multilayer perceptrons / deep NN, distance-weighted discrimination.
  • Unsupervised: hierarchical clustering, autoencoders, unsupervised DL.
  • Others: causal uplift modeling, genetic algorithms, digital twins.
• Representative applications:
  • Predictive employability models (students, applicants), virtual recruiters extracting nonverbal signals, ML-driven OSH monitoring (smart helmets, BLE + digital twins, real-time image recognition), AI coaching/chatbots, uplift modeling for safety policy impact, unemployment detection via smart-meter data.
• Governance & ethics:
  • Strong emphasis on explainability (XAI), fairness, transparency — particularly in light of the EU AI regulation adopted in 2024.
• Gaps and limitations:
  • Few real-world field experiments and limited access to sensitive workplace data.
  • Many studies focus on tools/outcomes rather than worker impacts (autonomy, bias, labor relations).

Data & Methods

• Search and selection:
  • Source: Scopus; years 2014–2024.
  • Keywords combined employment/AM/OSH concepts with ML/AI model terms.
  • PRISMA-like three-phase screening: title/abstract triage, full-text review, secondary citation retrieval.
• Empirical methods in the 25 studies:
  • ML model building and evaluation (classification, regression, clustering, causal uplift).
  • PLS-SEM and survey-based causal/associational analysis (many papers examine perceived impacts of AI adoption).
  • Systematic literature reviews and conceptual frameworks (taxonomies, readiness for regulation).
  • Field experiments/case studies: fewer, but include deployments such as smart helmets and cold-storage monitoring.
• Data characteristics reported across reviewed papers:
  • Heterogeneous: administrative HR data, sensor/image streams for safety, survey responses, smart-meter data, recruitment datasets.
  • Data sensitivity is a recurrent constraint limiting access and replication.

Implications for AI Economics

• Labor demand and skill composition
  • ML adoption supports task automation and enhanced monitoring (AM), shifting demand toward digital/analytical skills and increasing demand for reskilling investments. Economists should quantify occupation- and task-level exposure to these ML systems.
• Productivity vs. employment trade-offs
  • Evidence suggests ML can raise organizational productivity and improve safety (potentially reducing firm risk), but distributional effects on employment and wages are unclear. Firm-level causal estimates (productivity, hiring, separations) are needed.
• Measurement and policy targeting
  • ML tools improve forecasting of unemployment, employability, and safety incidents, enabling targeted labor-market interventions. Economists can use these predictive tools for more granular policy design if fairness constraints are handled.
• Regulation and compliance costs
  • The EU AI Act (2024) and emphasis on XAI/fairness will raise compliance costs (auditability, explainable architectures, documentation). This may (a) increase adoption costs, especially for high-risk employment applications, and (b) favor investments in explainable models or governance layers. Model choice and implementation costs should be included in adoption/benefit calculations.
• Market structure and platform labor
  • Algorithmic management practices can lower monitoring and transaction costs but may strengthen platform monopsony power, alter bargaining positions, and create privacy/externality issues. Empirical work should estimate impacts on wages, turnover, and market power.
• Safety and externalities
  • ML-based OSH improvements (real-time monitoring, predictive risk assessment) can reduce accident externalities and firm default risk (uplift causal findings). Valuing these safety gains is important for cost–benefit assessments of workplace AI.
• Research & data needs for economists
  • Greater access to granular firm and platform data, more field experiments, and pre-registered causal studies are needed to move from descriptive or predictive claims to credible causal inference about AI’s labor-market effects.
  • Interdisciplinary research combining ML, labor economics, industrial relations, and regulation will better capture welfare and distributional outcomes.
• Practical suggestions for economic analysis
  • Include regulatory compliance and explainability constraints in diffusion models of AI adoption.
  • Use ML-derived forecasts to design and evaluate active labor-market policies, but explicitly test for and correct classifier biases.
  • Examine heterogeneity by firm size, industry, and worker skill to identify winners/losers and potential policy compensations (training, income support).
  • Study how AM affects bargaining, contract design, and platform pricing/fees to infer broader market effects.

Overall, the review documents a maturing but uneven research landscape: powerful predictive and monitoring ML tools are emerging in employment contexts, yet robust causal evidence on macro and distributional labor-market effects — and on how regulation will shape adoption — remains limited. Economists have opportunities to quantify these effects, assess policy trade-offs, and incorporate the costs of explainability and fairness into analyses of AI-driven labor transformations.