Turnover intention reflects more than dissatisfaction or workload; it can express a judgement about the legitimacy of organisational control. The structural logic of control begins with disciplinary power. Foucault (1977) argued that visibility produces self-regulation, because individuals who assume continuous observation adjust their own behaviour. Artificial intelligence monitoring intensifies this logic by making observation datafied and continuous rather than spatial and episodic (Leicht-Deobald et al., 2019; Mettler, 2023). We supply the missing bridge in three steps. Observation produces self-regulation; self-regulation experienced as imposed rather than chosen threatens autonomy; and a sustained autonomy threat motivates withdrawal. Disciplinary power therefore explains the structural condition, but the route to turnover runs through the employee’s appraisal of that condition.

Autonomy is central to this route, and self-determination theory justifies its centrality rather than asserting it. Deci and Ryan (2000) demonstrate that autonomy is a basic psychological need whose frustration predicts disengagement and withdrawal. Artificial intelligence monitoring frustrates autonomy in a specific way that distinguishes it from strict EPM. Under observable EPM, an employee can infer the standard and adapt to it. Under opaque AI monitoring, the standard is unknowable, so adaptation becomes impossible, and control becomes unattainable. The threat is not merely ‘being watched’ but ‘being judged by a standard one cannot see’. This unpredictability converts monitoring into a sustained autonomy frustration (Gagné et al., 2022; McAnally & Hagger, 2024).

Privacy theory contributes something autonomy theory cannot. Autonomy concerns control over one’s actions; privacy concerns control over information about oneself. Artificial intelligence’s inferential capacity attacks the second specifically, because it produces information that the employee never disclosed and may not recognise as theirs. Nissenbaum (2004) names this violation through contextual integrity: A privacy harm occurs when information flows exceed the norms appropriate to a context. Artificial intelligence monitoring routinely exceeds those norms by inferring states from behavioural traces. Privacy theory therefore identifies a distinct injury – boundary overreach in the informational domain – that the autonomy concept does not contain (Carter, 2024; Mettler, 2023).

The integration rests on one ordering principle: Cognitive appraisal precedes motivation, and motivation precedes behaviour. Artificial intelligence monitoring establishes the structural condition. Boundary appraisal registers the condition as intrusion when informational norms are violated. Reactance converts intrusion into a motivational state directed at restoring threatened freedom (Brehm, 1966). The AI-specific features sharpen this sequence through a reasoned pairing rather than an enumeration. Opacity and inference act together: Opacity hides the standard while inference widens it, so employees face evaluation that is both expanding and unreadable. Automation and permanence also act together: Automation removes the human discretion that once allowed appeal, while permanence removes the temporal limit that once let evaluations expire. The first pair drives the intrusion appraisal; the second pair removes the conditions under which an employee might restore control internally. Together they convert monitoring from bounded evaluation into unbounded governance (Boyd & Andalibi, 2023; Mettler, 2023).

Artificial intelligence monitoring should raise PPI for reasons specific to AI rather than to monitoring in general. Observable EPM collects what employees know they are producing. Artificial intelligence monitoring infers what they have not disclosed and conceals the basis of that inference. Employees therefore appraise the collection as exceeding both what they consented to and what the task requires, which aligns with contemporary definitions of perceived information privacy violation and workplace privacy invasion (Bhave et al., 2020; Li et al., 2025). The opacity of the system compounds the appraisal, because employees cannot verify the limits of access and must assume the broadest plausible scope. Intrusion is thus a predictable response to inference under opacity, not to surveillance as such (Mettler, 2023):

Artificial intelligence monitoring should also raise psychological reactance through a route that is distinct from the privacy route, which resolves the apparent redundancy between a direct and a mediated path. The privacy route operates through informational appraisal; the direct route operates through behavioural standardisation. When automated systems narrow the range of acceptable behaviour and tighten evaluative feedback, they restrict the employee’s perceived freedom to choose how to work, independent of any informational concern (Feng et al., 2019; Sessions et al., 2023). The two routes are conceptually separable: One threatens control over information, the other threatens control over action. We retain the direct path because the standardisation mechanism does not require the privacy appraisal to operate:

Perceived privacy intrusion should activate reactance because boundary violation is a freedom threat. Privacy-boundary perspectives and communication privacy management theory hold that violations of informational norms provoke defensive arousal and a sense of turbulence when rules about access and use are exceeded (Meng, 2024; Petronio, 2015). When employees judge that monitoring exceeds legitimate authority, they experience the violation as a constraint on personal sovereignty, and reactance follows as the motivational defence of the threatened boundary (Yost et al., 2019). Intrusion is therefore the cognitive trigger that converts exposure into resistance:

Perceived privacy intrusion should also relate directly to TI, although weaklier, through the erosion of trust. Invasive digital surveillance and privacy violations are linked to reduced organisational trust, higher stress and lower satisfaction, which in turn increase withdrawal and intentions to leave (Kalischko & Riedl, 2024; Li et al., 2025). When monitoring is unavoidable and tied to livelihood, the privacy trade-off appears coercive rather than voluntary, so intrusion influences strategic decisions about leaving rather than discomfort alone:

Psychological reactance should exert the most proximal influence on TI, and reactance theory must be extended to specify why. Reactance theory describes the motivational state but stops short of naming the behaviour it produces in employment settings. We extend it by specifying the behaviour: When employees perceive no internal channel through which to restore autonomy, exit becomes the available restoration strategy. Sustained reactance under opaque governance therefore converts into withdrawal cognition, because the organisation offers no other route to recover the threatened freedom (Feng et al., 2019; Sessions et al., 2023). This specification is the theoretical advance the hypothesis requires:

The model’s central claim is a serial cascade, and we foreground it rather than testing every possible path. The ordering follows appraisal logic: Cognition precedes motivation, and motivation precedes behaviour. Artificial intelligence monitoring expands informational and evaluative reach; this expansion is appraised as intrusion; intrusion activates reactance; reactance drives the intention to leave. Evidence from related domains shows that privacy-related appraisals and trust judgements often precede and channel effects on downstream outcomes such as commitment, satisfaction and TI, and that sequential mediation models fit such processes better than parallel ones (Kalischko & Riedl, 2024; Zhao et al., 2022). We treat the serial path as the primary explanation and the direct effects as residual mechanisms that the cascade does not fully absorb. This priority answers why serial mediation is theoretically superior to parallel mediation: Parallel mediation treats intrusion and reactance as interchangeable routes, whereas the serial form respects their causal order, in which an appraisal must occur before the motivation it provokes. Specifying this order is what allows the model to explain the process rather than merely to estimate every coefficient:

We employed a cross-sectional survey design. The target population comprised Generation Z employees in Indonesian organisations whose work is subject to automated monitoring. We define automated monitoring as the use of software systems that track work activity, evaluate performance from that activity, and retain the resulting data without continuous human mediation. This definition includes AI-driven analytics at the high-opacity pole and key-performance-indicator dashboards at the low-opacity pole, and it excludes purely manual supervision. We adopted this operational boundary to remove the conflation of AI monitoring with non-automated oversight that an undefined sample would permit.

We recruited participants through purposive sampling with a screening filter. Eligibility required three conditions: Birth within the Generation Z range, current employment in an Indonesian organisation and direct exposure to an automated monitoring system as defined above. A screening question at the start of the questionnaire excluded respondents who did not meet all three conditions. We distributed the questionnaire online, informed participants that responses were anonymous and used only for research and stated that no answer was correct or incorrect. We retained 390 usable responses after removing incomplete cases, patterned responding and responses that failed attention checks (see Table 1 for the demographic distribution).

All constructs were measured using established reflective scales and rated on a five-point Likert scale ranging from 1 (strongly disagree) to 5 (strongly agree). The original English items were translated into Indonesian and adjusted to fit the study context without changing their core conceptual meaning. The adaptation involved three steps. Firstly, the wording was simplified so that respondents could understand each statement easily. Secondly, context-specific terms were adjusted to match the research setting and the language normally used by Indonesian respondents. Thirdly, the translated items were reviewed to ensure that each item still represented the same construct as in the original scale. No items were deleted or added during this process. The final Indonesian item wording is reported in Appendix 1 Table 1-A1.

We measured automated monitoring intensity (AMI) with six items grounded in the EPM literature (Stanton, 2000; Stanton & Julian, 2002; Stanton & Weiss, 2000). The items capture the perceived scope, frequency and evaluative use of automated tracking, such as ‘Automated systems are used to monitor my work activities’ and ‘Monitoring data are used in evaluating my performance’. We treat these items as indicators of perceived automated monitoring intensity rather than of verified AI features. This reading is deliberate: Employees respond to the perceived footprint of automated oversight, which is the theoretically relevant exposure. We record the corresponding boundary in the Limitations section, where we note that the items do not certify the presence of specific AI capabilities.

The PPI is measured with three items derived from established privacy scales (Malhotra et al., 2004; Smith et al., 1996) and reworded towards perceived overreach rather than general concern. The adapted items reference the judgement that monitoring exceeds what the work requires, such as ‘Monitoring practices go beyond what is necessary for work purposes’ and ‘I feel uneasy about the way monitoring data are stored or accessed’. The rewording shifts the target from diffuse concern about organisational practices to a contextual appraisal of boundary violation, which is the construct the model requires. We acknowledge the source scales’ origin in privacy concern and treat the distinction as a measurement boundary discussed in the Limitations.

The PRS is comprised of four items adapted from the Hong reactance scale (Hong & Faedda, 1996) and from state operationalisation of reactance (Dillard & Shen, 2005). We anchored each item to the monitoring context, which converts trait stems into situational appraisals, as in ‘Monitoring makes me feel my freedom to choose how I work is restricted’. These anchoring targets state reactance triggered by the present monitoring condition rather than a dispositional tendency. We note in the Limitations that the items draw on a trait instrument and that the contextual anchoring, while standard, does not fully resolve the trait–state distinction.

The TI comes with five items from the validated Turnover Intention Scale (Bothma & Roodt, 2013), which capture cognitive withdrawal, search intention and anticipated exit, such as ‘I often think about leaving this organisation’. The scale records the proximal cognitive state preceding actual turnover behaviour.

Ethical clearance to conduct this study was obtained from the Alauddin State Islamic University Research Ethics Code Enforcement Commission (Ref. No. 112/EC/KEPK-UINAM/2025).

This study shows that AI monitoring influences TI indirectly, through PPI and psychological reactance, with reactance as the proximal driver. Artificial intelligence monitoring therefore operates as a governance mechanism whose perceived legitimacy determines whether oversight stabilises performance or erodes retention. The contribution is the specified ordering of appraisal, motivation, and behaviour, which converts an unresolved debate about monitoring’s effects into a testable account of its mechanism.

Several limitations bound these conclusions and follow directly from the measurement choices. Firstly, the AI monitoring items derive from EPM instruments and capture perceived automated monitoring intensity rather than verified AI capabilities; future work should develop items that index opacity, inference, automation and permanence directly. Secondly, the Perceived Privacy Intrusion scale derives from privacy-concern measures and showed modest internal consistency, so its results warrant replication with a purpose-built intrusion scale. Thirdly, the reactance items adapt a trait instrument to a state appraisal through contextual anchoring, which does not fully resolve the trait–state distinction. Fourthly, the cross-sectional design limits causal inference, and self-report may inflate associations. Fifthly, the sample is restricted to Generation Z employees in Indonesia, and industry and contract heterogeneity were not modelled, which limits generalisability.

Future research should address these limits through longitudinal or multi-source designs, behavioural indicators such as system logs and actual turnover and AI-specific measurement. It should also test moderators that the legitimacy account predicts, including transparency, algorithmic fairness, participative implementation and autonomy-supportive leadership. From a managerial perspective, organisations should govern AI monitoring as a socio-technical arrangement, emphasising data proportionality, transparency, employee voice and autonomy-supportive practice, so that oversight does not convert into the reactance that drives talented employees to leave.