The double-edged sword effects of perceived algorithmic control on platform workers’ service performance

The study examines how perceived algorithmic control used by online labor platforms influences platform workers’ service performance, an outcome critical to platform survival and customer retention yet underexplored empirically. In the expanding gig economy, platforms employ algorithmic systems for guidance, tracking evaluation, and behavioral constraints. Drawing on the transactional theory of stress, the authors posit that workers appraise algorithmic control as either a challenge (leading to problem-focused coping like job crafting) or a threat (leading to emotion-focused coping like withdrawal). Integrating regulatory focus theory, they further argue that promotion focus amplifies the positive challenge path, while prevention focus amplifies the negative threat path. The study proposes and tests a moderated multiple mediation model linking perceived algorithmic control to service performance via job crafting and withdrawal behavior, moderated by promotion and prevention focus.

The paper reviews algorithmic control as a core mechanism in gig platforms aligning worker behavior with organizational goals via standardized guidance, tracking evaluation, and behavioral constraints. It defines perceived algorithmic control as workers’ comprehensive perception of real-time algorithmic guidance, evaluation, and constraints. Literature on outcomes of algorithmic control has largely emphasized continuance intention, workaround use, technostress, and well-being, with scant attention to service performance. Job crafting is presented as proactive, volitional changes to job design or social environment that enhance person–job fit and performance; algorithmic guidance and feedback can frame control as a challenge, stimulating job crafting. Withdrawal behavior encompasses counterproductive reductions in time/effort (e.g., absences, undeserved breaks) and is linked to hindrance appraisals, reduced efficacy, and lower performance; surveillance, fairness concerns, and reduced autonomy under algorithmic control can spur withdrawal. Regulatory focus theory distinguishes promotion (growth/eagerness) and prevention (security/vigilance) foci, shaping appraisals and behaviors. Hypotheses: H1 job crafting mediates a positive effect of perceived algorithmic control on service performance; H2 withdrawal mediates a negative effect; H3a/b promotion focus strengthens the PAC→job crafting link and weakens PAC→withdrawal; H4a/b prevention focus weakens PAC→job crafting and strengthens PAC→withdrawal; H5a/b promotion focus strengthens the positive and weakens the negative indirect effects on service performance; H6a/b prevention focus weakens the positive and strengthens the negative indirect effects.

Design: Cross-sectional survey of platform workers in China (online ride-hailing drivers and delivery riders). Data collection: Online questionnaire via Credamo from May–August 2022. Sampling: 391 invited; 286 valid responses after excluding 105 for failed attention checks, patterned or random responses (response rate 73.15%). Demographics: 51.4% male; ages 18–30 (50.7%), 31–40 (37.4%), ≥41 (11.9%); education high school or below (39.5%) vs college or above (60.5%); daily service time <4 h (21.0%), 5–8 h (40.9%), >9 h (38.1%); full-time (54.2%) vs part-time (45.8%); industry: ride-hailing (49.3%), delivery (39.2%), others (12%). Measures (5-point Likert 1–5): Perceived algorithmic control (11 items; standardized guidance, tracking evaluation, behavioral constraints; Pei et al., 2021); Job crafting (4 items; Vogel et al., 2016); Withdrawal behavior (4 items; Lehman & Simpson, 1992); Promotion and prevention focus (9 items each; Koopmann et al., 2016); Service performance measured as self-reported customer rating (platform 1–5 scoring). Controls: gender, age, education, average daily service time, occupation type, industry type. Reliability/validity: Cronbach’s alphas—PAC 0.923; JC 0.919; WB 0.840; Promotion 0.975; Prevention 0.967. CFA (Mplus 8.3): six-factor model fit χ²/df=1.81, CFI=0.939, TLI=0.935, RMSEA=0.053, SRMR=0.049; alternative models showed poorer fit. CMV checks: Harman’s single-factor first factor 30.8% (<40%); unmeasured latent method factor model did not materially improve CFI over theoretical model. Analysis: SPSS hierarchical regressions to test direct and interaction effects; bootstrapped mediation (5,000 resamples) and moderated mediation (Edwards & Lambert; Preacher et al.) estimating indirect effects at ±1 SD of moderators.

• Perceived algorithmic control (PAC) positively related to job crafting (JC) and to withdrawal behavior (WB): JC β=0.285, p<0.001 (Model 1); WB β=0.337, p<0.001 (Model 4).
• JC positively predicted service performance (SP): β=0.431, p<0.001 (Model 8). WB negatively predicted SP: β=-0.477 (Model 9; text notes p<0.001).
• Mediation (bootstrapping, 5,000 resamples): PAC→JC→SP positive indirect effect Estimate=0.306, p<0.001, 95% CI [0.180, 0.441] (H1 supported). PAC→WB→SP negative indirect effect Estimate=-0.097, p=0.050, 95% CI [-0.219, -0.019] (H2 supported). Total effect of PAC on SP significant: Estimate=0.210, p<0.001, 95% CI [0.037, 0.379]; positive indirect effect > negative indirect effect.
• Moderation by promotion focus (PM): PAC×PM on JC positive and significant (γ=0.479, p<0.001; Model 2) (H3a). PAC×PM on WB negative and significant (γ=-0.161, p<0.001; Model 5) (H3b).
• Moderation by prevention focus (PF): PAC×PF on JC negative and significant (γ=-0.342, p<0.001; Model 3) (H4a). PAC×PF on WB positive and significant (γ=0.345, p<0.001; Model 6) (H4b).
• Moderated mediation (indirect effects at ±1 SD): Promotion focus—PAC→JC→SP: High PM 0.443 [0.243, 0.621]; Low PM 0.169 [0.065, 0.317]; difference 0.274 [0.091, 0.452] (H5a). PAC→WB→SP: High PM -0.043 [-0.191, 0.051] (ns); Low PM -0.151 [-0.273, -0.063]; difference 0.108 [0.017, 0.237] (H5b).
• Prevention focus—PAC→JC→SP: High PF 0.174 [-0.019, 0.327] (ns); Low PF 0.439 [0.277, 0.668]; difference -0.266 [-0.562, -0.065] (H6a). PAC→WB→SP: High PF -0.262 [-0.459, -0.116]; Low PF 0.069 [-0.059, 0.211] (ns); difference -0.331 [-0.615, -0.135] (H6b).
• Measurement quality: Cronbach’s alphas ≥0.84; CFA fit acceptable (CFI=0.939; RMSEA=0.053); CMV not a major concern.

Findings support a double-edged effect of perceived algorithmic control on platform workers’ service performance. When perceived as a challenge, algorithmic control’s standardized guidance, feedback, and procedural fairness cues stimulate problem-focused coping via job crafting, enhancing service performance. When perceived as a threat, surveillance, perceived unfairness, and reduced autonomy elicit emotion-focused coping via withdrawal behavior, undermining service performance. Regulatory focus shapes these appraisals and responses: promotion-focused workers experience more positive emotions, treat algorithmic control as instrumental for gains, craft their jobs more, and withdraw less; prevention-focused workers experience more negative emotions and vigilance, attenuating job crafting and intensifying withdrawal. Overall, the positive path via job crafting is stronger than the negative path via withdrawal, indicating that as algorithmic technologies mature, their capacity to enhance performance through guided improvement may outweigh adverse effects, especially among promotion-focused workers.

This study contributes by (1) empirically linking perceived algorithmic control to service performance, extending outcomes research beyond continuance intentions and well-being; (2) integrating transactional stress theory to reveal parallel positive (job crafting) and negative (withdrawal) mechanisms, uniting bright and dark sides of algorithmic management; and (3) identifying regulatory focus as a boundary condition, with promotion focus strengthening positive and weakening negative indirect effects, and prevention focus doing the opposite. Practically, platforms should enhance workers’ understanding of algorithmic intent (training, communication, support) and design algorithms with human-centered considerations to bolster perceived fairness and autonomy. They can assess and cultivate promotion focus through selection and training to encourage adaptive responses and job crafting. Future research should examine additional mechanisms and multilevel moderators (e.g., platform care, leadership styles), use longitudinal or experimental designs to strengthen causal inference, obtain objective performance metrics or matched data to reduce self-report bias, and employ mixed methods for richer insights.

• Cross-sectional design limits causal inference; longitudinal or experimental studies are needed.
• Service performance measured via self-reported customer ratings; all variables self-reported, raising potential social desirability and common method bias (though CMV checks suggest limited impact). Future work should use matched or objective platform data.
• Mediators limited to job crafting and withdrawal behavior; other mechanisms (e.g., trust, perceived fairness, autonomy, technostress) warrant examination.
• Moderators focused on individual traits (promotion/prevention focus); organizational and team-level moderators (platform care, leadership style) should be explored.
• Methodologically, only quantitative survey methods were used; future mixed-methods (case studies plus surveys) could provide deeper validation.