Supportive HR helps gig workers — until opaque algorithmic oversight muddles the link to performance; clear, explainable oversight restores performance gains and reduces confusion.

Main Finding

Using Double/Debiased Machine Learning (DML) to estimate a moderated mediation model on survey data from 464 Indian gig workers, the paper finds a clear non‑monotonic (U‑shaped) conditional indirect effect of supportive HR practices (HRP) on worker performance via the relational psychological contract: the mediated positive effect is strong at low algorithmic control, weakest at moderate (a “murky middle”), and strengthens again at high control — but only when the algorithmic system is experienced as clear, explainable, and offering recourse. By contrast, the mediated effect of HRP on worker wellbeing is positive and stable across levels of algorithmic control. Methodologically, the paper shows DML can recover these nonlinear conditional indirect effects without imposing linear functional forms, and that standard linear specifications can be biased or misleading.

Key Points

• Theoretical framing:
  • HR practices are costly signals and job resources that form a relational psychological contract; this contract mediates HR → performance and HR → wellbeing.
  • Algorithmic control (AC) is modeled as a continuous moderator with two key dimensions: intensity (surveillance/constraint) and clarity (transparency/predictability/recourse).
  • The “murky middle” hypothesis: moderate intensity combined with low clarity hampers translation of relational contract into performance; high intensity + high clarity can restore this pathway.
• Hypotheses:
  • H1a/H1b: Relational psychological contract mediates HRP → performance and HRP → wellbeing.
  • H2a: The indirect effect on wellbeing remains positive and stable across AC.
  • H2b: The indirect effect on performance follows a U‑shape across AC (weakest at moderate levels).
  • H2b’: Recovery at high AC depends on perceived clarity/explainability.
• Empirical findings:
  • Mediation via relational contract observed for both performance and wellbeing.
  • Performance: confirms U‑shaped CNIE — weakest at moderate AC, recovery at high AC conditional on clarity.
  • Wellbeing: mediated effect remains positive and largely invariant to AC.
  • Linear parametric models can produce biased or opposite conclusions; DML recovers the nonmonotonic pattern.
• Robustness & alternatives considered:
  • Monte Carlo simulations and benchmark comparisons.
  • Tested transactional contract as robustness check (distinct mechanism).
  • Discussed alternative mechanisms (cognitive load, platform heterogeneity) and potential unobserved confounding.

Data & Methods

• Sample:
  • N = 464 gig workers in metropolitan India (food delivery, ride‑hailing, last‑mile logistics).
  • Mixed recruitment (local coordinators, intercept surveys, online forums); multiple languages with back‑translation; informed consent; confidentiality.
  • Post‑hoc power analysis: > .80 power to detect small–medium effects (f² ≈ 0.03).
• Measures (multi‑item scales, mostly 5‑point Likert):
  • HR Practices (HRP): 32 items on staffing, training, compensation, appraisal.
  • Relational Psychological Contract (PC_Rel): 9 items.
  • Transactional Psychological Contract (PC_Tran): 8 items (robustness).
  • Algorithmic Control (AC): 8 items (primarily captures intensity; clarity discussed theoretically).
  • Worker Performance: 8‑item Individual Work Performance scale.
  • Worker Wellbeing: 18‑item wellbeing scale.
  • Controls: age, gender, education, marital status, gig‑experience (months), full/part‑time, platform fixed effects.
• Estimation strategy:
  • Primary estimand: Conditional Natural Indirect Effect (CNIE) θ(w) — how the indirect effect of HRP → M → Y varies with continuous moderator W (AC).
  • Double/Debiased Machine Learning (DML): cross‑fitting and Neyman‑orthogonal scores to flexibly control high‑dimensional confounding using machine learners while delivering debiased causal estimates.
  • Complementary checks: Monte Carlo simulation to show parametric bias, benchmark parametric models for comparison, sensitivity discussion about sequential ignorability and selection on unobservables.
• Limitations noted by authors:
  • Cross‑sectional survey, potential unobserved confounding and platform‑level heterogeneity.
  • Algorithmic clarity dimension is primarily theoretical; measurement instrument largely captures intensity.
  • Outcomes are self‑reported; objective task data not used.
  • Sequential ignorability assumption required for causal claims.

Implications for AI Economics

• For platform design and firm strategy:
  • Algorithmic oversight need not uniformly undermine human‑centered HR investments; its effect is conditional. Platforms should avoid a partially defined control regime that is opaque — the “murky middle” reduces effectiveness of relational HR.
  • Strong oversight can coexist with high performance if the system is transparent, predictable, and offers meaningful recourse (explainability, interpretable metrics, human review/appeal).
  • Investing in clarity (explanations, accountability, recourse mechanisms, intelligible metrics) can protect worker wellbeing benefits from HR investments and restore performance gains even under high control.
• For labor supply, productivity, and welfare analyses:
  • Nonlinear effects matter: estimates of the impact of algorithmic management on productivity, turnover, or welfare may be misspecified if models impose linear relationships.
  • Policy interventions (e.g., transparency mandates, fairness/appeal rights) may have asymmetric returns: clarity can unlock productivity benefits of strict algorithmic regimes while protecting wellbeing.
• For empirical AI economics research:
  • Use of flexible causal methods (DML, causal forests, meta‑learners) is essential when interactions and moderators are continuous and effects may be nonmonotonic.
  • Standard linear interaction models can be misleading — both substantively (wrong sign) and quantitatively — so researchers should test for nonlinear conditional effects and consider orthogonalized, machine‑assisted estimation approaches.
• For regulation and governance:
  • Empirical support for targeting explainability and recourse (not just restricting monitoring intensity) — regulators should consider requiring meaningful transparency and appeal mechanisms in algorithmic management.
• Future research directions:
  • Longitudinal or experimental designs to address selection on unobservables and to observe dynamics of contract formation and enactment.
  • Objective performance metrics and richer measures of clarity (separate intensity vs clarity scales).
  • Platform‑level randomized interventions (e.g., different explanation/appeal treatments) to test causal mechanisms and policy levers.

Summary conclusion: The paper shows that algorithmic management’s impact on the returns to human‑centered HR is nonlinear — ambiguous, partly enforced control undermines performance pathways, but transparent and explainable algorithmic systems can restore them. Methodologically, it demonstrates DML as a practical tool for recovering such conditional indirect effects in AI and organizational economics settings.