Algorithmic management in Chinese food-delivery platforms creates a cascade of interlinked health burdens: time pressure and physical exhaustion are most pervasive, but punitive algorithms and weak protections provoke the harshest reactions; findings come from analysis of 10,103 social-media posts and 32 rider interviews.

Main Finding

Algorithmic management in online food delivery (OFD) produces a tightly interconnected, cascading system of occupational health burdens for riders. From a large-scale text corpus (10,103 social-media comments) supplemented by 32 semi‑structured interviews, the authors identify five core burden dimensions — physical exhaustion, social devaluation, disciplinary distress, injury vulnerability, and health‑protection deficit — with physical exhaustion acting as the central hub in a transmission network. Crucially, the most structurally central burden (physical exhaustion) is not the one that provokes the strongest negative sentiment (punitive mechanisms and protection deficits do).

Key Points

• Five empirically derived health-burden dimensions:
  • Physical exhaustion (central hub)
  • Social devaluation (stigma, loss of dignity)
  • Disciplinary distress (algorithmic punishments, rating systems)
  • Injury vulnerability (traffic risk, accidents)
  • Health‑protection deficit (lack of insurance/compensation)
• Cascading structure: time/operational pressures → punitive mechanisms → elevated injury risk → gaps in social/health protection.
• Sentiment gap: frequency/structural prominence ≠ emotional intensity. Riders normalize time pressure (widely discussed but less emotively charged), while punitive and protection deficits elicit strongest negative responses.
• Methodological contribution: a mixed-methods pipeline combining LDA topic modeling, topic co‑occurrence network analysis (centrality metrics), sentiment intensity aggregation, plus interviews and expert validation.

Data & Methods

• Primary corpus: 10,103 valid rider comments scraped from a major Chinese social media platform (unstructured text).
• Qualitative validation: 32 semi‑structured interviews with OFD riders; expert review to label and validate topics.
• Computational methods:
  • LDA topic modeling to extract latent themes (model selection validated using perplexity and coherence).
  • Sentiment intensity analysis to compute topic-level emotional weights (feature-based aggregation of document-level sentiment weighted by topic mixture scores).
  • Topic co‑occurrence network construction; network centrality metrics (e.g., degree, betweenness) used to identify hubs and likely transmission pathways between burden dimensions.
• Framing theory: extended Job Demands–Resources (JD‑R) perspective to emphasize structural transmission and cascading effects among demands/resources.

Implications for AI Economics

• Efficiency–externality tradeoff: Algorithmic optimization that raises throughput can impose negative externalities (health, safety, and social protection costs) on workers. These externalities are structurally interlinked and may be undercounted in platform profit calculations.
• Welfare measurement: Simple metrics (income, hours, accidents) miss networked interactions and subjective distress. Economists should incorporate both structural centrality (how a shock propagates) and sentiment intensity (welfare salience) when valuing worker well‑being.
• Labor supply and productivity: Persistent cascade effects (fatigue → accidents → medical absence) can reduce long-run labor supply, raise turnover, and generate hidden costs (healthcare, enforcement, reputational). Short‑term algorithmic gains may be offset by these longer‑term losses.
• Insurance and risk markets: The identified protection deficit implies demand for portable occupational‑injury insurance and novel risk‑pooling mechanisms. Platforms’ classification of workers (independent contractors) creates regulatory arbitrage that shifts risk to markets and public budgets.
• Platform design and incentives: Algorithmic design (delivery time windows, rating rules, punishment thresholds) directly shapes labor supply decisions and risky behaviors. Economists and designers should consider incentive structures that internalize safety and health costs (e.g., incorporating safety-adjusted bonuses, slack in time windows, forgiving rating mechanics).
• Regulation and policy: Findings support policies requiring algorithmic transparency, enforceable minimum scheduling/contract standards, and mandated or subsidized portable insurance. Regulatory changes will alter platform cost structures and competitive dynamics — potentially leading to platform redesigns or increased prices for consumers if costs are internalized.
• Research directions for AI economics:
  • Quantify monetary value of cascading health burdens (healthcare costs, lost earnings, accident externalities).
  • Integrate text‑mined structural measures into empirical models of worker utility, turnover, and labor supply.
  • Simulate counterfactual algorithm designs (e.g., relaxed time windows, lower penalty rates) to estimate welfare and productivity tradeoffs.
  • Evaluate how different regulatory regimes (worker classification, mandatory insurance) affect platform incentives and equilibrium outcomes.
• Methodological takeaway: Large-scale text mining + network analysis is a promising toolkit for capturing multi‑dimensional, systemic labor harms created by algorithmic management and for informing economically grounded policy evaluation.

If you’d like, I can: - Extract the exact topic labels and top words from the LDA output (if you provide the vocabulary/topic lists), or - Draft an outline for an economic model that internalizes cascading health externalities from algorithmic scheduling.