Algorithmic controls are a double-edged sword for food-delivery riders: real-time tracking and punitive constraints heighten pressure, worsen mental health and raise risky riding, while standardized guidance reduces pressure and improves safety — effects depend on riders’ perceived autonomy.

Main Finding

Perceived algorithmic control has a “double‑edged sword” effect on food delivery riders: algorithmic tracking evaluation and behavioral constraint increase work pressure, which harms riders’ mental health and increases risky riding behavior; in contrast, algorithmic standardized guidance reduces work pressure, which improves mental health and reduces risky riding. Perceived autonomy moderates these pathways — it buffers the harms of tracking evaluation, strengthens the benefits of standardized guidance, but unexpectedly amplifies the harms of behavioral constraint.

Key Points

• Algorithmic control is decomposed into three perceptual dimensions:
  • Tracking evaluation (continuous monitoring and assessment),
  • Behavioral constraint (limits/penalties on how tasks are done),
  • Standardized guidance (algorithmic recommendations/instructions).
• Work pressure (consistent with the Job Demands–Resources model) mediates the link between perceived algorithmic control and two outcomes:
  • Mental health (negative outcome),
  • Risky riding behavior (safety-related behavior).
• Mediation results:
  • Tracking evaluation → ↑ work pressure → ↓ mental health and ↑ risky riding.
  • Behavioral constraint → ↑ work pressure → ↓ mental health and ↑ risky riding.
  • Standardized guidance → ↓ work pressure → ↑ mental health and ↓ risky riding.
• Moderation by perceived autonomy:
  • Reduces the negative indirect effects of tracking evaluation.
  • Strengthens the positive indirect effects of standardized guidance.
  • Paradoxically amplifies the negative indirect effects of behavioral constraint.
• Sample and analysis: 466 Chinese food delivery riders; structural equation modeling (SEM) with bootstrapped mediation/moderation tests.

Data & Methods

• Sample: 466 food delivery riders in China (survey data).
• Constructs measured (from abstract): perceived algorithmic control (three sub-dimensions), work pressure, perceived autonomy, mental health, and risky riding behavior.
• Analytical approach:
  • Structural equation modeling to estimate relationships among latent constructs.
  • Bootstrapping to test indirect (mediated) effects and moderation of pathways.
• Theoretical framing: Job Demands–Resources (JD‑R) model — algorithmic features as demands/resources that alter work pressure and downstream outcomes.

Implications for AI Economics

• Platform design trade-offs: Algorithms that increase monitoring and constrain behavior can raise short‑term efficiency but impose psychological costs and externalities (higher risky driving → public safety risk). Standardized guidance can deliver efficiency gains while lowering pressure and reducing negative externalities.
• Worker welfare and labor supply: Increased work pressure and reduced mental health may lower retention, reduce effective labor supply, and impose health-related costs (absenteeism, turnover). Platforms’ profit-maximizing algorithmic policies should internalize these longer-run costs.
• Externalities and social costs: Algorithmic control can create negative externalities (road accidents, medical costs) that markets may not price; regulations or platform policies should address these.
• Incentive alignment and mechanism design: Designers must balance monitoring/constraint with autonomy and supportive guidance. Optimal algorithmic governance should consider heterogenous worker preferences for autonomy and the non‑linear effects found (e.g., autonomy can worsen some constraints).
• Policy and regulation: Findings support policies promoting transparency, worker control options, limits on intrusive monitoring, mandated rest/breaks, and safety‑oriented algorithmic incentives.
• Directions for economic research:
  • Quantify welfare trade-offs (productivity vs mental/health costs) and incorporate psychological costs into labor supply models.
  • Model dynamic interactions between algorithmic control, worker behavior, and platform pricing/assignment mechanisms.
  • Evaluate heterogeneous responses across worker types and across gig sectors; pursue causal and longitudinal designs to assess long‑term impacts.
  • Consider externalities in social welfare analyses and optimal regulation of algorithmic management.

Summary: The study shows that not all algorithmic control is equal — guidance can be beneficial while monitoring and constraints can be harmful through increased work pressure. For AI economics, this highlights important trade‑offs in algorithmic governance, worker welfare impacts, and socially relevant externalities that platforms and policymakers must consider.