Can the same algorithm help a delivery worker perform better and make that worker feel more threatened? New research suggests the answer may be yes. A survey of 389 food-delivery workers in China found that algorithmic management was linked to both positive and negative interpretations of work, with different associations with wellbeing.
A manager that is also a source of stress
For a food-delivery courier, the app is more than a map. It may allocate orders, suggest routes, evaluate performance and influence pay. Those functions can help workers organise a difficult day, but they can also create uncertainty and pressure. A new peer-reviewed study of 389 delivery workers in China finds evidence of both experiences occurring within the same system. Algorithmic management was associated with stronger perceptions of opportunity and stronger perceptions of threat. These two interpretations, in turn, were linked to workplace wellbeing in opposite directions. The research offers a more precise picture than the familiar claim that automated management is simply good or bad for workers.
What the researchers investigated
The study, published on 15 October 2026 in Frontiers in Psychology, examined the relationship between algorithmic management, workers’ appraisals of their jobs and wellbeing. Researchers Zha Xiaofang, Zhang Yue, Zhang Dandan, Feng Xinying and Suziana Mat Yasin drew on stress-appraisal theory. Their central question was not merely whether workers encounter algorithms, but how workers interpret the demands those systems create. A challenge appraisal means seeing an aspect of work as a potentially manageable opportunity for improvement or achievement. A threat appraisal means seeing it as a possible loss, restriction or danger to important goals. These are not necessarily mutually exclusive reactions.
The headline finding: two opposing pathways
The statistical model identified a positive association between algorithmic management and challenge appraisal, with a standardised path coefficient of 0.500. The association with threat appraisal was also positive, at 0.324. Both were reported as statistically significant at p below 0.001. Challenge appraisal was positively associated with workplace wellbeing, with a coefficient of 0.451. Threat appraisal was negatively associated with wellbeing, with a coefficient of minus 0.202. These figures describe associations in the fitted model, not percentages of workers affected and not the probability that a specific courier will become happier or more stressed.
Why the distinction matters
Consider two workers receiving automated performance feedback. One may regard the information as a useful signal for improving delivery times and earnings. Another may see the same feedback as evidence that an opaque system can penalise them without explanation. A single worker could also feel both reactions on different days. The study’s contribution is to treat these appraisals as separate pathways rather than forcing every experience onto a single positive-to-negative scale. This matters for platform operators because improving the usefulness of an algorithm does not automatically eliminate the anxieties it may generate.
How the survey was conducted
Researchers collected questionnaires in two waves, separated by approximately three weeks. At the first wave, participants reported demographic information, perceived algorithmic management and psychological resilience. The second wave measured challenge appraisal, threat appraisal and workplace wellbeing. Initially, 657 people entered the survey. After screening, 571 usable first-wave responses remained. Of these, 496 people completed the second questionnaire, and the final matched analytical sample comprised 389 respondents. The authors reported a 68.1% retention rate from the valid first-wave sample. The time separation was intended to reduce some common-method bias, although it does not turn the study into a controlled experiment.
Who the 389 workers were
The final sample included 153 delivery workers in Shanghai, 115 in Nanjing and 121 in Suzhou. Men accounted for 86.9% of participants. About 43.2% were aged 18 to 25, and 82.0% reported working full time. The largest monthly-income group, accounting for 37.5%, reported earning 5,001 to 6,000 Chinese yuan. These characteristics help define the population behind the findings. The sample is informative about surveyed delivery workers in three Chinese cities, but it should not be treated as a representative estimate for all gig workers worldwide.
What algorithmic management included
The researchers measured algorithmic management using a 14-item scale covering practices such as assessment and feedback, training and guidance, allocation and pricing, and tracking and monitoring. These categories capture an important distinction: algorithms can provide informational support while also exercising control. A courier may value accurate allocation of orders yet object to intensive tracking or unclear penalties. The study examined these dimensions both within a broader management construct and through additional analyses that separated their relationships with worker appraisals. This allowed the researchers to ask whether different functions of the technology were associated with different psychological experiences.
Feedback and pricing looked different from tracking
In supplementary regression analyses, assessment and feedback were positively associated with challenge appraisal, with a reported coefficient of 0.366. Allocation and pricing were also positively associated with challenge appraisal, at 0.136. By contrast, tracking and monitoring were positively associated with threat appraisal, at 0.144. The authors reported these relationships as statistically significant, while other dimension-specific relationships were not. This pattern does not establish that monitoring causes distress. It does suggest that the way a platform deploys its technology may matter more than whether it uses algorithms at all.
The role of resilience
Psychological resilience was examined as a potential moderator: could workers’ ability to cope with difficulty alter how they interpreted algorithmic management? The researchers found a positive interaction between algorithmic management and resilience for challenge appraisal, with a coefficient of 0.133. Workers with higher resilience showed a stronger positive challenge-related association. But the interaction for threat appraisal, at 0.036, was not statistically significant. In other words, resilience did not provide convincing evidence of a comparable buffer against the threat-related pathway. This is a notable distinction for organisations tempted to rely on resilience training instead of addressing structural working conditions.
Indirect associations were also tested
Using 5,000 bootstrap resamples, the study estimated a positive indirect association between algorithmic management and wellbeing through challenge appraisal of 0.226, with a 95% confidence interval from 0.149 to 0.327. The corresponding indirect association through threat appraisal was negative, at minus 0.065, with an interval from minus 0.116 to minus 0.029. Both intervals excluded zero. These results support the statistical plausibility of two opposing pathways within the proposed model. They do not prove a causal chain from app design to appraisal to wellbeing, because participants were not randomly assigned different management systems.
A useful detail in the robustness checks
The authors examined an alternative model that added a direct path between algorithmic management and wellbeing. This model fitted the data better than the original full-mediation specification, and the added direct association was statistically significant, with a coefficient of 0.191. That finding matters because it cautions against treating challenge and threat appraisals as the complete explanation. Other experiences, including earnings stability, work intensity, autonomy and relationships with customers or supervisors, could contribute to workers’ overall wellbeing. The study is better understood as identifying relevant psychological pathways than as offering a final map of every influence.
Why a two-wave design is helpful but limited
Separating predictor and outcome questions across three weeks is stronger than collecting every response at precisely the same moment. It can reduce some immediate consistency pressures in self-reporting. Nevertheless, participants reported their own perceptions at both waves, and there was no experimental manipulation of platform rules. A worker who already feels positive about their job may rate the algorithm more favourably, while unmeasured working conditions could influence both responses. The study’s mediation language refers to statistical modelling rather than definitive proof of a sequence of causes. This distinction is essential when translating findings into policy or product recommendations.
What platforms could learn
One practical implication is to evaluate automated management as a collection of design decisions. Clear feedback, predictable task allocation and understandable pricing may be experienced as resources. Constant surveillance, unexplained changes and difficult-to-contest decisions may be experienced as threats. The findings suggest that platforms should assess both dimensions independently. A rise in operational efficiency or reported usefulness does not necessarily mean workers feel secure. User research with couriers, transparent appeal channels and meaningful explanations of decisions are plausible responses, although the study did not experimentally test whether those interventions improve wellbeing.
What regulators and employers should notice
The gig economy often blurs the boundary between technology and management. An automated instruction can have practical consequences similar to a supervisor’s decision even when no human manager is visible. For policymakers, the relevant questions may therefore include transparency, contestability, fair evaluation and the conditions under which workers can challenge a decision. The Chinese findings cannot be directly converted into evidence about South African labour law or local delivery platforms. They can, however, help frame questions worth asking wherever workers depend on apps to allocate tasks and evaluate performance.
The South African connection
Research Today has previously reported on the legal uncertainties described by Johannesburg platform workers. That reporting addressed employment protection and insecurity rather than testing the psychological model used in this new Chinese study. Read together, the articles highlight complementary concerns: whether workers have adequate rights, and how the systems governing their daily work are experienced. Readers interested in the local context can explore our report on Johannesburg gig workers and legal gaps. The two studies should not be combined as if they measured the same population or outcomes.
Limits of the evidence
The research relied on self-reported survey responses, so common-method bias and inaccurate recall remain possible. The sample came from three cities and was predominantly young and male. Attrition between survey waves may have introduced selection bias. The measures captured perceptions of algorithmic management rather than an independent technical audit of the platforms’ actual systems. The analysis also cannot establish whether app design changes would produce the same relationships over months or years. Finally, the observed coefficients depend on the chosen scales and statistical model; they should not be read as universal effects across occupations.
What should be studied next
Future work could compare delivery platforms with different levels of transparency and monitoring, follow workers for longer periods, and examine whether changing specific features alters their experiences. Combining surveys with administrative data on earnings, shifts, order allocation and complaints could reduce reliance on perceptions alone. Randomised or natural-experiment designs would offer stronger evidence about causality where ethically and practically possible. Research in different countries could also test whether employment protections and economic conditions alter the balance between challenge and threat. Those questions are particularly relevant as algorithmic systems spread beyond delivery work into warehouses, customer service and professional workplaces.
The bottom line
The study of 389 Chinese food-delivery workers found that algorithmic management was associated with both opportunity-oriented and threat-oriented appraisals, which in turn had opposite relationships with wellbeing. Resilience strengthened the positive challenge pathway but did not significantly weaken the threat pathway. The findings suggest that making workers more adaptable is not a substitute for examining how automated systems are designed and governed. They also demonstrate why claims that algorithms either liberate or exploit workers are often too simple. Better management may depend on making the useful aspects clearer while reducing unnecessary uncertainty and control.
Source Information
Study title: Algorithmic management and gig workers’ workplace wellbeing: the moderating role of psychological resilience
Authors: Zha Xiaofang, Zhang Yue, Zhang Dandan, Feng Xinying, and Suziana Mat Yasin
Journal: Frontiers in Psychology
Publication date: 15 October 2026
Study design: Two-wave observational survey, three weeks apart, analysed using structural equation modelling, regression and bootstrap mediation
Sample or scope: 389 matched food-delivery workers in Shanghai, Nanjing and Suzhou, China
Key finding: Algorithmic management was associated with both challenge and threat appraisal; resilience strengthened the positive challenge pathway but did not significantly buffer the threat pathway
