Back to Paper
View PDF Viewing HTML Download Markdown

Navigating the Relational Risk of AI in Workplaces: A Cross-National Comparative Analysis of Institutional Mediation

AI Authors: Gemini (Google), initial drafting; Claude Fable 5.1 (Anthropic) via Perplexity Computer, verification and revision

Prompter: Andreas Cebulla

Abstract

The economic literature on workplace artificial intelligence (AI) concentrates on job displacement and productivity; the relational degradation of work receives far less attention. Building on the five relational risks of workplace AI mapped in Cebulla (2025), this paper compares five institutional regimes, Australia, the United States, Singapore, France and Germany, across five relational friction points: (i) algorithmic management and autonomy loss, (ii) tacit knowledge erosion and automation bias, (iii) labour content and status devaluation, (iv) fragmented accountability, and (v) technology-facilitated psychosocial harm, the last in its interpersonal and system-mediated manifestations. It assembles verified evidence from employer and worker surveys, official statistics, case law and legislation. The pattern is indicative but consistent: where consultation, co-determination and occupational health and safety (OHS) duties are institutionalised, algorithmic management is better governed. Around half of US workers are aware of being monitored and monitored workers report markedly worse well-being; algorithmic evaluation tools are used by 90% of US firms but 35% of European firms, and 28% of managers using such tools are unclear who is responsible when they err; 7% of EU establishments use performance-monitoring systems and only a minority consult employees on digital technologies; in Australia only 9% of those polled know of monitoring in their own workplace while 72% of senior HR decision-makers believe employees are told; and 41% of Singapore workers feel constantly watched. The regimes differ sharply in whether a human decision is required before an automated system penalises or deactivates a worker.

1. Introduction and Theoretical Framework

Artificial intelligence is rapidly restructuring the modern workplace, yet the most visible scholarship measures its effects in terms of occupational exposure, task displacement and productivity: which jobs can be automated (Eloundou et al., 2024), how much output rises when workers are given generative AI tools (Brynjolfsson et al., 2025), and how employers and workers expect employment and wages to move (Organisation for Economic Co-operation and Development [OECD], 2023). The relational degradation of work receives far less systematic attention. Relational risks of AI are the ways in which AI systems disrupt, mediate or transform workplace relationships, social dynamics and power structures; they concern how AI affects the human experience of work, including trust, dignity, authority, communication, social cohesion and well-being. Five such risks were mapped for Australian workplaces, drawing on national and international evidence, in Cebulla (2025): algorithmic management and the erosion of autonomy; the devaluation of tacit knowledge; the revaluation of labour; technology-facilitated workplace incivility; and the fragmentation of supervisory relationships. That paper also argued that the five risks interact and compound one another. The present paper adopts the taxonomy, relabelled slightly to foreground the mechanisms examined cross-nationally:

  • Risk (i): Algorithmic Management and Autonomy Loss. Continuous electronic performance monitoring, automated dispatch and keystroke or screen tracking that strip worker discretion over work pace and methods.
  • Risk (ii): Tacit Knowledge Erosion and Automation Bias. Over-delegation of complex judgment tasks to automated systems, leading to the atrophy of human expertise and uncritical reliance on algorithmic output.
  • Risk (iii): Labour Content and Status Devaluation. Reclassification of professional and skilled knowledge work into routine oversight, degrading vocational identity and career progression.
  • Risk (iv): Fragmented Supervision and Black-Box Accountability. Dilution of human managerial presence, creating opaque evaluation chains in which automated decisions cannot readily be challenged or appealed.
  • Risk (v): Technology-Facilitated Psychosocial Harm. Intensification of work pace, boundary erosion and perpetual digital surveillance leading to chronic stress and burnout. Because such harm can be machine-mediated, human-triggered or both, it is treated here as having two manifestations: interpersonal misuse, in which individuals use digital tools to harass, profile, impersonate or embarrass colleagues; and system-mediated harm, in which automated processes penalise, exclude or deactivate workers through procedures that cannot readily be challenged.

Risks (i) to (v) correspond respectively to algorithmic management, tacit knowledge devaluation, labour revaluation, supervisory fragmentation and technology-facilitated incivility in the original taxonomy; Risk (v) is broadened from incivility to psychosocial harm so that it can be read against the indicators that national statistical and OHS agencies collect, and its two manifestations are examined in Sections 3.5 and 4.6. The original paper addressed the risks primarily in relation to Australia, illustrated with international as well as Australian evidence, and proposed a work health and safety (WHS) response for that jurisdiction. This paper takes the taxonomy cross-nationally to ask whether AI's relational effects are technologically determined or mediated by national institutions and collective worker voice. It does so through a structured comparison of published evidence rather than new statistical analysis, because the available surveys were not designed to be pooled (Section 2).

1.1 Situating the framework in the literature

The five risks draw together strands of research that have developed largely in parallel. The first is the organisational literature on algorithmic management. Kellogg et al. (2020) reframed algorithms at work as a new contested terrain of control through which employers direct, evaluate and discipline workers; Parent-Rocheleau and Parker (2022) showed that algorithmic management operates as a work designer, shaping autonomy, skill variety, feedback and task identity; and Krzywdzinski et al. (2025), in a German case study, found that an algorithmic system's effects depended on bargaining among managers, engineers, data scientists and workers during implementation rather than on the technology alone. Risks (i) and (iv) are the relational expression of this literature: autonomy loss where algorithmic design narrows discretion, and black-box accountability where the control terrain becomes opaque.

The second strand is the psychology of monitoring and work design. A meta-analysis finds no evidence that electronic performance monitoring improves performance, while its presence is associated with increased stress regardless of its characteristics (Ravid et al., 2023). Parker and Grote (2022) argue that automation and algorithms make work design matter more, not less; Gagné et al. (2022) reach the same conclusion from self-determination theory; and Ajunwa (2023) shows how the quantification of the worker outpaces legal protections designed for an earlier workplace. This strand underpins Risk (v) in its intensification form and explains why the paper reads the relational risks against the psychosocial indicators that national agencies collect.

The third strand concerns expertise and judgment under automation. Parasuraman and Manzey (2010) integrated the human-factors evidence on automation complacency and bias, the tendency to under-monitor automated systems and accept their outputs uncritically; Raisch and Krakowski (2021) identified an automation–augmentation paradox in which organisations that pursue one at the expense of the other undermine both; and Dell'Acqua et al. (2026) showed experimentally that generative AI degrades performance on tasks outside its capability frontier (Section 3.2). Risks (ii) and (iii) are the consequences this strand predicts if delegation to automated systems proceeds without investment in human judgment.

The fourth strand is the literature on workplace mistreatment and dignity. Andersson and Pearson (1999) introduced workplace incivility as low-intensity deviant behaviour that can spiral into aggression, and the meta-analyses of Han et al. (2022) and Yao et al. (2022) establish its consequences for health, affect, withdrawal and performance. Lamers et al. (2022) apply a capability approach to worker dignity under algorithmic management, which can expand or restrict workers' real opportunities to be and do what they value; Aloisi and De Stefano (2022) make the parallel legal argument that algorithmic bosses require the same accountability as human ones. Risk (v) in its interpersonal and system-mediated forms, developed in Sections 3.5 and 4.6, sits within this strand.

The fifth strand is comparative institutional analysis, which supplies the paper's explanatory logic. The varieties-of-capitalism tradition distinguishes liberal market economies, in which coordination runs through markets and managerial prerogative is broad, from coordinated market economies, in which works councils, sectoral bargaining and vocational training institutionalise worker voice (Hall & Soskice, 2001). Thelen (2018) showed that Uber produced sharply different regulatory outcomes across Europe and the United States because existing coalitions and institutions, not the technology, determined what was contested. Doellgast and Wagner (2022) argue that collective regulation remains central to the future of work in the digital economy, and Doellgast et al. (2025) document how social dialogue on AI and algorithmic management is unfolding across jurisdictions. The present paper extends this strand to the relational risks.

Read together, the organisational and psychological literatures identify the mechanisms, the mistreatment and dignity literatures specify the harms, and the comparative institutional literature predicts where those harms will be greater or smaller. Missing is a synthesis that applies one relational taxonomy across regimes and anchors it to verified national statistics.

1.2 Country selection

The comparison covers Australia, France, Germany, Singapore and the United States. The cases were chosen to vary the explanatory factor identified in the fifth strand above, the degree to which worker voice over workplace technology is institutionalised, among advanced economies for which published evidence on each of the five risks could be assembled; all five are covered by the one survey that provides a common measure across them (Gillespie et al., 2025). Table 1 summarises the institutional features that motivated the selection; Section 4 examines how those features bear on the five risks.

Hall and Soskice (2001) classify the United States and Australia as liberal market economies, Germany as a coordinated market economy and France among the economies in a "more ambiguous" position; Singapore lies outside their typology. The selection uses this spectrum but refines it along the dimension that matters for relational risk: whether workers or their representatives can shape a monitoring or algorithmic system before it is deployed, or can only contest its effects afterwards. On that dimension Germany, where works councils hold co-determination rights over monitoring devices, and France, where works council consultation is statutory and court-enforced, represent ex ante voice reached by different routes, bargained and legislated; including both allows the European pattern to be read as more than a single case. The United States, with at-will employment and no federal monitoring statute, is the benchmark of minimal ex ante constraint. Australia and Singapore are the analytically interesting intermediate cases. Australia is a liberal market economy in which worker voice over technology is weak but a positive work health and safety duty to manage psychosocial hazards is far-reaching, a test of whether an ex post OHS duty can substitute for an ex ante veto; it is also the jurisdiction for which the framework was developed. Singapore combines tripartite coordination with state-championed AI adoption and advisory rather than statutory governance of workplace technology, a test of whether coordination without a veto buffers the risks.

Table 1. Country selection: institutional features relevant to the five relational risks

Country Coordination type (Hall & Soskice, 2001) Ex ante worker voice over monitoring and algorithmic tools Psychosocial-hazard and automated-decision regulation Role in the comparison
Germany Coordinated market economy Works council co-determination over technical devices designed to monitor behaviour or performance (BetrVG, 2001, § 87(1)(6)); AI adoption negotiated with works councils in practice (Krzywdzinski & Wotschack, 2026) Directive (EU) 2024/2831: platform account restriction, suspension or termination requires a human decision; transposition due December 2026 (Directive [EU] 2024/2831, 2024) Strongest ex ante, bargained voice
France "More ambiguous" position, with a history of extensive state intervention Statutory information and consultation of the social and economic committee (CSE) on new technologies; evaluation methods disclosed in advance (Code du travail, 2008, arts. L1222-2 to L1222-4); court-enforced from the pilot phase (Chagny & Marzo, 2026) As Germany (EU directive) Statutory and judicial, rather than bargained, voice; with Germany, allows the European pattern to be read across two routes
Australia Liberal market economy No ex ante veto; surveillance clauses in about 5% of recent enterprise agreements (Howe et al., 2026) WHS positive duty to manage psychosocial hazards (Safe Work Australia, 2022); parliamentary recommendation for a WHS code on AI (House of Representatives Standing Committee on Employment, Education and Training, 2025); Fair Work Act 2009 (Cth) unfair-deactivation regime (platform workers) Intermediate case: liberal labour market with a far-reaching but reactive OHS duty; the jurisdiction for which the framework was developed
Singapore Outside the typology; tripartite coordination among government, unions and employers (Ministry of Manpower, n.d.-b) No statutory co-determination or consultation right over workplace technology identified; Tripartite Jobs Council formed in 2026 to coordinate training and transition support for AI adoption (Ministry of Manpower, 2026a) Workplace Safety and Health framework with voluntary iWorkHealth assessment (Ministry of Manpower, 2025b); Platform Workers Act 2024 extends social protection and representation but does not regulate algorithms (Koh, 2024; Ministry of Manpower, n.d.-a) Tests whether tripartite coordination without an ex ante veto buffers the risks
United States Liberal market economy At-will employment; no federal electronic monitoring statute; notice duties only in a minority of states (Conn. Gen. Stat. § 31-48d, 2025; Del. Code tit. 19, § 705, 2001; N.Y. Civ. Rights Law § 52-c, 2022) No federal right to human review of automated employment decisions identified; New York City Local Law 144 requires bias audits and notice only (New York City Department of Consumer and Worker Protection, n.d.); 2024 EEOC harassment guidance rescinded 2026 (EEOC, n.d., 2026) Benchmark of minimal ex ante constraint; governance left to firm policy

Note. Rows are ordered from strongest to weakest ex ante worker voice. Coordination types follow Hall and Soskice (2001); the remaining columns summarise legal and institutional sources examined in Section 4. "Ex ante" denotes rights exercisable before a system is deployed.

The selection is purposive, not a sample: it omits the Nordic and Japanese variants of coordination and all middle-income economies, so the conclusions concern the range of institutional variation among high-income adopters rather than its full extent. Evidence availability was a secondary criterion: Australia and Singapore are not covered by the OECD AI surveys (Section 2), so the comparison relies on national sources for them.

2. Evidence Base and Cross-National Comparability

2.1 Sources

OECD AI surveys. In early 2022 the OECD surveyed 5,334 workers and 2,053 firms in the manufacturing and finance sectors of Austria, Canada, France, Germany, Ireland, the United Kingdom and the United States (Lane et al., 2023; OECD, 2023). They are the most detailed worker-level evidence on AI in adopting firms but do not cover Australia or Singapore, report at sector rather than country level and predate generative AI; a 2026 repeat had not reported at the time of writing (OECD, n.d.).

Algorithmic management and accountability sources. An OECD employer survey of 6,047 firms with 20 or more employees in France, Germany, Italy, Japan, Spain and the United States (2024), answered by mid-level managers, measures tools to instruct, monitor and evaluate workers (Milanez et al., 2025). The 2024 European Working Conditions Survey provides employee-reported measures for France and Germany (European Foundation for the Improvement of Living and Working Conditions [Eurofound], 2026). A survey of 1,000 Singapore employers (June and July 2026) covers AI in HR functions and human oversight (Singapore National Employers Federation [SNEF], 2026); a 47-country survey of 1,001 to 1,098 adults per country (November 2024 to January 2025) covers unverified reliance on AI output at work (Gillespie et al., 2025); and a survey of 11,004 US adults (December 2022) covers attitudes to AI in hiring and evaluation (Rainie et al., 2023).

ESENER 2024 and OSH Pulse 2025. The fourth European Survey of Enterprises on New and Emerging Risks covered 41,458 establishments with five or more employees in 30 European countries (May to October 2024), interviewing the person best informed on health and safety (European Agency for Safety and Health at Work [EU-OSHA], 2025a); its worker-side complement, OSH Pulse 2025, interviewed 28,220 employed people in the EU-27, Iceland and Norway in March and April 2025 (EU-OSHA, 2025b).

APA Work in America surveys. Online surveys of employed US adults by The Harris Poll for the American Psychological Association (APA): 2,515 in 2023, 2,017 in 2025 and 2,009 in 2026 (APA, 2023, 2025, 2026); only the 2023 wave publishes monitored-versus-unmonitored contrasts.

PwC Global Workforce Hopes and Fears surveys. The 2024 edition surveyed 56,600 workers in 50 countries and territories, including 1,500 in Australia (PwC, 2024; PwC Australia, 2024). The 2026 edition surveyed nearly 50,000 workers in 48 countries and territories and introduced four workforce archetypes, "front-runners" (14%), "AI insurgents" (18%), "indispensables" (11%) and the "engine room" (56%) (PwC, 2026); country-level archetype shares are not published, but a Singapore edition covers 1,000 respondents (PwC Singapore, 2026).

Australian sources. Safe Work Australia's psychological health analysis draws on workers' compensation statistics, the People at Work survey (Kessler K10 scale) and the National Return to Work survey (Safe Work Australia, 2024); claims statistics are from the 2026 key statistics release (preliminary 2024–25 data; Safe Work Australia, 2026). No nationally representative worker survey of monitoring published since 2018 was identified; monitoring evidence is drawn from the Human Technology Institute's 2026 reports, which synthesise 2024–2025 union surveys and March 2026 polling of more than 2,000 Australians (Farthing et al., 2026a, 2026b), the Australian HR Institute's weighted survey of 603 senior decision-makers in October 2025 (Australian HR Institute [AHRI], 2025), ADP Research's Australian results (ADP, 2025b) and a peer-reviewed analysis of surveillance provisions in enterprise agreements (Howe et al., 2026).

Singapore and ADP Research sources. The Ministry of Manpower's iWorkHealth online psychosocial assessment has collected anonymised data since 2021 (Ministry of Manpower, 2025b). ADP Research's People at Work 2025, a survey of nearly 38,000 working adults in 34 markets (July and August 2024), provides the only identified worker-side monitoring measure covering Singapore, and a companion release reports Singapore workers' AI sentiments (ADP, 2025a; ADP Research, 2025); the measure is perceptual, and ADP is a private HR and payroll provider.

German sources. A telephone survey of 385 companies (2023–24), with matched management and works council interviews in 224 and AI use in 149, examines conflicts over AI adoption under co-determination (Krzywdzinski & Wotschack, 2026), and an establishment panel for 2019–2023 examines training responses to AI adoption (Muehlemann, 2025).

Technology-facilitated harm sources. Prevalence measures: a 2022 Australian perpetration survey of 3,345 adults (Flynn et al., 2024); a representative survey of 5,015 German employees in 2023–24 (Löbner et al., 2025) and the Federal Statistical Office's indicator from the 2024 European Working Conditions Survey (Statistisches Bundesamt, n.d.); French recorded-crime statistics (Hadj Larbi et al., 2025) and a commercial barometer (Qualisocial et al., 2026); and Singapore's Supplementary Survey on Fair Employment Practices (Ministry of Manpower, 2025a). Adjudicated cases and legal instruments are cited in Section 3.5. These measure different things and are not combined numerically.

2.2 Comparability limits

Country-level percentages drawn from these sources are not commensurable, for four reasons.

  1. Unit of analysis. ESENER reports employers' accounts at establishment level; the APA, OECD, OSH Pulse, ADP and PwC surveys interview workers, who may be unaware of systems in use or feel monitored in ways establishment instruments do not capture.
  2. Coverage. The OECD AI surveys cover seven countries and two sectors; ESENER and OSH Pulse cover Europe only; the APA surveys cover the United States only. No single source covers all five countries.
  3. Instruments. Monitoring questions differ (awareness of any monitoring technology, establishment use of performance-monitoring systems, feeling constantly watched), as do distress measures (the K10 scale, Likert-type items, establishment risk audits).
  4. Microdata. The OECD and PwC microdata are not public and ESENER is an establishment dataset, so pooled worker-level analysis is not possible.

Each verified statistic is therefore presented within its source, and institutional comparisons are drawn qualitatively; the evidential status of the measures, most of which are perceptions or employer reports, is set out in Section 4.1.

3. Verified Evidence by Risk Domain

3.1 Algorithmic management and monitoring (Risk i)

United States. Around half of US workers are aware that their employer uses technology to monitor them while working: 51% in 2023 (APA, 2023), 48% in 2025 (APA, 2025) and approximately half again in 2026 (APA, 2026). Employer pressure to use AI rose from about one in five workers in 2025 (20–21%) to 33% in 2026, while only 38% of workers in 2026 reported an official employer AI-use policy (APA, 2025, 2026).

European Union. Seven percent of EU-27 establishments report using machines, systems or computers that monitor workers' performance and/or behaviour (15% in transportation and storage) (EU-OSHA, 2025a). On the worker side, 25% of EU-27 workers report that digital technologies are used to supervise or monitor their work and behaviour (24% in France, 16% in Germany); 27% that digital technologies automatically allocate their tasks, working time or shifts (29% France, 25% Germany); and 48% that digital technologies determine the pace of their work (54% France, 37% Germany) (EU-OSHA, 2025b). The gap between the 7% establishment and 25% worker figures reflects the unit of analysis (Section 2.2), not a contradiction.

OECD AI surveys. Among workers in firms that had adopted AI, 49% in finance and 39% in manufacturing said the firm's AI applications collected data on them or their work, and roughly three-quarters of these said the data were used to assess their performance; 7% of manufacturing and 6% of finance workers reported being managed by AI (Lane et al., 2023).

Australia. Union surveys in 2024–2025 compiled by the Human Technology Institute found that 50% of more than 1,000 Finance Sector Union members reported high or very high levels of workplace surveillance and that, among Victorian Trades Hall Council members, 61% were aware of surveillance in their workplace and 81% said it had not been discussed with them (Farthing et al., 2026b). A third of Australian workers (33%) in ADP Research's 2025 series reported being monitored by their managers (ADP, 2025b). On the employer side, 28% of senior decision-makers surveyed in October 2025 used AI for performance management, and 72% agreed that employees are made aware when AI monitors their performance (AHRI, 2025). Worker awareness contradicts that belief: in March 2026 polling, 43% of Australians were aware that some employers use technology, including AI, to monitor workers, but only 9% knew of monitoring in their own workplace (Farthing et al., 2026a). These measures differ in vantage point and are not commensurable with each other or with the US or EU figures.

Singapore. In the absence of official monitoring statistics, the best available measure is ADP Research's: 41% of workers in Singapore feel constantly watched by their employer, the fifth-highest share among 34 markets (global average 32%) (ADP Research, 2025).

3.2 Tacit knowledge erosion and automation bias (Risk ii)

Risk (ii) concerns the atrophy of human judgement when tasks are delegated to automated systems and their outputs accepted uncritically. In a field experiment with 758 knowledge workers, generative AI raised productivity on tasks inside its capability frontier, but on a task outside it workers using AI were 19% less likely to produce correct solutions (Dell'Acqua et al., 2026). The only measures identified that cover all five countries on a common instrument come from the 47-country trust study (Gillespie et al., 2025). The share of employees who had relied on AI to do a task rather than learning how to do it themselves was 42% in Australia, 45% in the United States, 47% in France, 46% in Germany and 51% in Singapore; the share who had relied on AI output at work without evaluating its accuracy was 57%, 58%, 60%, 65% and 69% respectively; between 34% (Germany) and 45% (Singapore) felt they could not complete their work without AI; and between 52% (France) and 68% (Singapore) reported having made mistakes in their work because of AI (Gillespie et al., 2025). These are self-reports, not tests of skill, but they are the closest available country-level proxies for automation complacency, and on each item Singapore ranks highest and the European countries do not rank lowest. In the OECD AI surveys, around 40% of employers in both sectors cited a lack of relevant skills as a barrier to AI adoption (Lane et al., 2023). For Singapore, 28% of respondents use generative AI daily (22% globally) and 53% rate AI taking over more tasks as a moderate or major threat to their job security (44% globally) (PwC Singapore, 2026); 19% were uncertain how AI would change their job in the coming year and 11% feared replacement (ADP, 2025a).

3.3 Labour content and status devaluation (Risk iii)

PwC's 2026 global survey places 56% of workers in the "engine room" archetype: essential to delivery but with limited access to AI advantage, learning and innovation. Fewer than 40% of engine room workers report access to the learning and development they need, against almost 80% of front-runners; 51% of front-runners use generative AI daily versus 11% of the core workforce; only about a third of engine room workers believe they are fairly paid and 33% trust top management (PwC, 2026). In the OECD AI surveys, about 20% of workers were very or extremely worried about losing their job within ten years (Lane et al., 2023).

3.4 Fragmented accountability and worker voice (Risk iv)

Risk (iv) has two faces: the substitution of software for managerial judgement, most visibly in performance evaluation, and the accountability gap that opens when an AI system sits between supervisor and worker. On substitution, 90% of US managers report at least one algorithmic management tool in their firm, compared with 81% in France and 78% in Germany, but the tools differ in kind: tools that evaluate workers are used by 90% of US firms against 35% in the four European countries, and software sanctions poor performance in 67% of US firms but 4% of European firms (Milanez et al., 2025). Worker reports concur: 19% of EU employees say a computer program monitors their performance to a large or some extent, 18% that programs allocate their tasks and 10% that a program sets when they work, with composition-adjusted performance monitoring at 18% in France and 10% in Germany (Eurofound, 2026). In Singapore, 74% of employers use AI in at least one HR function and 45% of these use it for performance, learning and mobility (SNEF, 2026); in Australia, 28% of senior HR decision-makers use AI for performance management (AHRI, 2025).

The intermediary problem is visible on both sides. Some 28% of managers using such tools are unclear who is responsible when a tool's decision or recommendation is wrong, 27% have difficulty understanding its recommendations, and decreases in manager-employee interaction outnumber increases in every country except Japan; yet 91% of managers say workers or their representatives are informed (Milanez et al., 2025). Among EU employees whose work is partly decided by a computer program, only 57% report a clear procedure for contesting an automated decision (Eurofound, 2026). In Singapore, 85% of AI-supported HR use cases still require human sign-off, but 35% of employers using AI in HR cite accountability for AI errors as a concern (SNEF, 2026). The gap also runs downward: between 57% (Australia) and 69% (Singapore) of employees have relied on AI output at work without checking its accuracy (Gillespie et al., 2025; Section 3.2).

Worker preferences and voice institutions are the counterweight. A majority of workers in the OECD AI surveys (57% in both sectors) supported banning AI from deciding dismissals, 47% promotions and roughly 40% recruitment (Lane et al., 2023). Among US adults, 71% oppose AI making final hiring decisions, 55% its use in dismissals and 39% AI evaluation of job performance (Rainie et al., 2023). In the EU-27, only 35% of establishments using digital technologies consult employees about their possible health and safety impact, and risk assessments cover digital technologies in 43% of workplaces, but only 24% of French and 29% of German establishments, even though 74% of French establishments have employee-elected health and safety representatives (EU-OSHA, 2025a). Consultation rights are thus not automatically converted into risk management. In Australia, surveillance clauses appear in only 284 of 5,728 recent Fair Work Commission-approved enterprise agreements in the industries examined, approximately 5% (Howe et al., 2026).

3.5 Technology-facilitated psychosocial harm: strain, interpersonal misuse and system-mediated harm (Risk v)

Risk (v) is examined in three parts. Psychosocial strain is measured by worker and establishment surveys and compensation statistics; interpersonal misuse, where at all, by harassment surveys that isolate a digital channel; system-mediated harm mainly through adjudicated disputes and legislation, because no survey in the five countries asks workers whether an automated decision has penalised them without recourse.

Psychosocial strain. Monitored US workers report worse well-being than unmonitored workers: feeling tense or stressed at work (56% vs 40%) and emotional exhaustion (39% vs 22%) (APA, 2023). In the OECD AI surveys, 62% of finance and 56% of manufacturing workers whose data are collected report increased pressure to perform (Lane et al., 2023). These are cross-sectional associations. In the EU, 34% of establishments using digital technologies identify increased work intensity and 32% information overload as risk factors, with Germany among the highest (54% for the two combined) (EU-OSHA, 2025a). In Australia, mental health conditions accounted for 12.2% (17,900) of approximately 147,200 serious workers' compensation claims in 2024–25 (preliminary), a 169.4% increase since 2014–15 and the largest change of any nature-of-injury group, with the share at 17.5% among women and 8.2% among men (Safe Work Australia, 2026), and 60% of senior decision-makers agreed in October 2025 that AI is increasing work intensity (AHRI, 2025). In Singapore, about one in three employees assessed through the voluntary iWorkHealth tool report work-related stress or burnout, a proportion relatively stable since 2021 (Ministry of Manpower, 2025b). None of these measures isolates AI as a cause.

Interpersonal misuse. The Australian estimate is a perpetration measure: 15% of adults who had worked in the previous 15 years reported having engaged in workplace technology-facilitated sexual harassment in 2022, most commonly by work email (31%) (Flynn et al., 2024); no victimisation estimate published since 2022 was identified. Generative AI has also appeared as an instrument of conflict: in Wibmer v Fujifilm Data Management Solutions Pty Ltd (2026) a developer dismissed after sending at least 17 AI-drafted emails alleging misconduct by colleagues in four days lost his unfair dismissal claim. For the United States no nationally representative estimate was identified; the Equal Employment Opportunity Commission's (EEOC) 2024 harassment guidance addressed virtual work environments but was rescinded on 22 January 2026 (EEOC, 2024, 2026), although its small-business fact sheet still lists sharing AI-generated or deepfake sexual depictions as harassing conduct (EEOC, n.d.). For France and Germany, OSH Pulse 2025 records harassment or bullying among 14% of French and 7% of German workers (EU-27: 8%), without isolating a digital channel (EU-OSHA, 2025b); 8% of German employees experienced (cyber-)bullying, harassment or physical violence in the previous 12 months (Statistisches Bundesamt, n.d.); and 2.4% of employees were unjustly criticised, harassed or exposed in social media, emails or chats at least weekly, while 70.8% of those classified as bullied reported no such online mistreatment, so digital channels accompany rather than define most bullying (Löbner et al., 2025). In France, digital means were involved in 12% of recorded workplace moral-harassment cases in 2024 (1,432 victims), a police-reported measure (Hadj Larbi et al., 2025); a commercial barometer reporting 28% of employees affected publishes no methodology (Qualisocial et al., 2026). In Singapore, 5.5% of resident employees experienced harassment at work in 2023 but cyber-bullying was reported by 0.2%, a single item with a low base (Ministry of Manpower, 2025a).

System-mediated harm. In Australia, Part 3A-3 of the Fair Work Act 2009 (Cth) allows employee-like platform workers to contest unfair deactivation. In Hotak v Rasier Pacific Pty Ltd (2025) a Full Bench found a driver's deactivation plainly unfair and ordered reactivation; in Ayyub v Portier Pacific Pty Ltd (2026) a delivery worker deactivated for falling below an 85% customer-satisfaction threshold was reactivated because the process was entirely automated, with no human involved in any decision-making, whereas the Deactivation Code presupposes a judgment made by a human actor. In the United States, the EEOC's 2023 iTutorGroup settlement concerned recruitment software that automatically rejected older applicants (EEOC, 2023), and New York City's Local Law 144 requires a bias audit and notice before an automated employment decision tool is used (New York City Department of Consumer and Worker Protection, n.d.). In the EU, Directive (EU) 2024/2831, to be transposed by 2 December 2026, provides that any decision to restrict, suspend or terminate a platform worker's account shall be taken by a human being and grants a right to explanation and human review (Directive [EU] 2024/2831, 2024). In Singapore, the Platform Workers Act 2024 extends social protection and representation to platform workers and lets an association member whose account is terminated without just cause seek resumption of access through the association and the Minister (Ministry of Manpower, n.d.-a); the Government declined to regulate platform algorithms directly (Koh, 2024).

Table 2 summarises the verified evidence by regime.

Table 2. Verified evidence base and institutional context

Regime Monitoring / algorithmic management Psychosocial outcomes Governance (see Table 1)
United States About half of workers aware of employer monitoring (51% 2023, 48% 2025); employer pressure to use AI 20–21% (2025) to 33% (2026) (APA, 2023, 2025, 2026) Monitored vs unmonitored: tense or stressed 56% vs 40%; emotional exhaustion 39% vs 22% (APA, 2023). No national measure of technology-facilitated harassment identified State notice statutes only (Table 1); Local Law 144 bias audits; iTutorGroup settlement (EEOC, 2023); 2024 harassment guidance rescinded 2026
France and Germany (EU-27) 7% of establishments use performance/behaviour monitoring systems; 35% consult employees on digital technologies; risk assessments cover digital technologies in 24% (FR) and 29% (DE) of establishments (EU-OSHA, 2025a). Workers reporting digital monitoring: EU-27 25%, FR 24%, DE 16% (EU-OSHA, 2025b) Work intensity cited by 34% and information overload by 32% of establishments using digital technologies (EU-OSHA, 2025a). Harassment or bullying FR 14%, DE 7% (not digital-specific; EU-OSHA, 2025b); DE 2.4% mistreated online weekly (Löbner et al., 2025); FR digital means in 12% of recorded moral-harassment cases (Hadj Larbi et al., 2025) Co-determination (DE) and statutory CSE consultation (FR) over monitoring (Table 1); courts enforcing consultation on AI (2022, 2025); Directive (EU) 2024/2831 human-decision requirement for platform work
Australia 50–61% of surveyed union members report or are aware of surveillance (Farthing et al., 2026b); 33% of workers report being monitored by managers (ADP, 2025b); 72% of senior HR decision-makers believe staff are told of AI monitoring, but only 9% of those polled know of monitoring in their own workplace (AHRI, 2025; Farthing et al., 2026a) Mental health conditions 12.2% of serious compensation claims (2024–25p), up 169.4% since 2014–15; 17.5% of women's vs 8.2% of men's claims (Safe Work Australia, 2026). 15% self-report perpetrating technology-facilitated sexual harassment at work (Flynn et al., 2024) WHS psychosocial duty (Table 1); Full Bench held in 2025 and 2026 that fully automated deactivation cannot comply with the Deactivation Code
Singapore 41% of workers feel constantly watched (34-market average 32%); perceived monitoring, not establishment prevalence (ADP Research, 2025) One in three iWorkHealth-assessed employees report stress or burnout (Ministry of Manpower, 2025b). Workplace harassment 5.5% in 2023; cyber-bullying 0.2% (Ministry of Manpower, 2025a) Tripartite, advisory governance (Table 1); Platform Workers Act 2024 without regulation of algorithms
OECD AI surveys (7 countries, manufacturing and finance) 49% (finance) and 39% (manufacturing) of workers in AI-adopting firms report AI data collection; ~75% of these say it assesses performance (Lane et al., 2023) 62% (finance) and 56% (manufacturing) of data-collected workers report increased pressure to perform (Lane et al., 2023) Training and worker consultation associated with better worker outcomes (Lane et al., 2023)
Algorithmic management and accountability (multi-country sources) Any algorithmic management tool: US 90%, FR 81%, DE 78%; evaluation tools US 90% vs Europe 35% (Milanez et al., 2025). Program monitors performance: EU 19%, FR 18%, DE 10% (Eurofound, 2026). Singapore employers using AI in HR: 74% (SNEF, 2026) Managers unclear who is responsible for tool errors: 28% (Milanez et al., 2025). Unverified reliance on AI output: AU 57%, US 58%, FR 60%, DE 65%, SG 69% (Gillespie et al., 2025) Clear procedure for contesting automated decisions: 57% of affected EU employees (Eurofound, 2026); human sign-off in 85% of Singapore AI-supported HR use cases (SNEF, 2026)

4. Discussion: Institutional Varieties and Power Asymmetries

Read through industrial relations, institutional economics and occupational health frameworks, the verified evidence suggests that regulatory architectures buffer or amplify the five risks by shaping the balance of power between those who deploy AI systems and those who work under them. This follows from the premise of the framework itself: relational risks are not external threats but are internal to the dynamics of power and control in the workplace (Cebulla, 2025). The asymmetry is visible in the voice indicators: only 35% of EU-27 establishments using digital technologies consult employees about their health and safety impact, and digital-technology risk assessments cover just 24% of French and 29% of German establishments (EU-OSHA, 2025a); surveillance clauses appear in roughly 5% of recent Australian enterprise agreements (Howe et al., 2026); 72% of Australian senior decision-makers believe employees are told when AI monitors their performance while only 9% of Australians polled know of monitoring in their own workplace (AHRI, 2025; Farthing et al., 2026a); and a third of US workers report employer pressure to use AI while most have no official AI policy (APA, 2026). These are measures of who is informed, consulted and able to contest, the terrain on which the institutional varieties summarised in Table 1 differ. The argument is interpretive: the evidence is consistent with institutional mediation of the five risks but, for the reasons set out in Sections 2.2, 4.1 and 5, cannot establish it causally.

4.1 Status of the evidence

The figures in Section 3 fall into three tiers. The first is administrative and adjudicated record: compensation claims (Safe Work Australia, 2026), recorded offences (Hadj Larbi et al., 2025), Singapore's official harassment survey (Ministry of Manpower, 2025a), the Fair Work Commission decisions, statutes and the directive. These are hard measures but, except for the deactivation cases, not AI-specific: a mental health claim or a harassment offence does not record whether an algorithmic system was involved. The second tier is employer and establishment reporting: ESENER (EU-OSHA, 2025a), the OECD employer survey (Milanez et al., 2025), AHRI (2025) and SNEF (2026). These record the existence of a tool or procedure, as reported by a single informant, rather than its use or effect. The third, and largest, tier is worker self-report: awareness of monitoring (APA, 2023; Farthing et al., 2026a), feeling watched (ADP Research, 2025), perceived pressure (Lane et al., 2023), reliance on AI output (Gillespie et al., 2025), attitudes to AI in employment decisions (Rainie et al., 2023) and job-security worry (PwC, 2026). These are perceptions and opinions, not measures of exposure or harm.

The distribution across risks is uneven. Risks (i) and (iv) draw on all three tiers; Risk (iii) and the strain component of Risk (v) rest almost entirely on the third; Risk (ii) rests on the third and on one field experiment (Dell'Acqua et al., 2026); system-mediated harm rests on the first tier alone. Several of the paper's most striking contrasts are between tiers rather than within them. The Australian gap between the 72% of senior decision-makers who believe employees are told of AI monitoring and the 9% of those polled who know of it in their own workplace (AHRI, 2025; Farthing et al., 2026a) compares an employer belief with a worker awareness poll, not two measures of monitoring; the EU gap between 7% of establishments and 25% of workers (EU-OSHA, 2025a, 2025b) is of the same kind. Such gaps are informative about transparency and voice, which is how Sections 4.2 to 4.5 use them, not estimates of prevalence. Perceptions also matter in their own right for Risk (v), because psychosocial hazard frameworks treat perceived job demands and job control as proximate hazards (Safe Work Australia, 2022). What the evidence cannot yet show is the link between institutionally governed exposure and measured harm; that would require the worker-level, AI-specific data called for in Section 6.

4.2 Institutional voice and the mitigation of algorithmic management [Risks (i) and (iv)]

In Germany, technical devices designed to monitor employee behaviour or performance are subject to works council co-determination under the Works Constitution Act (Betriebsverfassungsgesetz [BetrVG], 2001, § 87(1)(6)): algorithmic tracking cannot be introduced unilaterally. In France, the employment code requires that information collected about employees be relevant to assessing their professional capabilities, that employees be informed in advance of evaluation methods, and that no personal information be collected by a device not previously disclosed to them (Code du travail, 2008, arts. L1222-2 to L1222-4); the social and economic committee (CSE) must be informed and consulted on new technologies, and data protection law adds a further constraint. These obligations bite: the Pontoise court held in 2022 that the CSE must be consulted on AI systems from the pilot phase, and an interim order of the Créteil Judicial Court of 25 July 2025 suspended generative-AI tools deployed without prior consultation (Chagny & Marzo, 2026; Ministère du Travail, n.d.). German survey evidence qualifies the reach of these ex ante barriers: in 149 AI-using companies, works councils with stronger institutional resources experienced fewer conflicts over AI introduction, but only 2% had concluded an AI-specific agreement, and where conflict arose councils rated their capacity to protect employee interests as weaker; the authors conclude that co-determination's strength lies in fostering compromise before conflict develops rather than in protecting workers once management contests (Krzywdzinski & Wotschack, 2026). The low establishment-level prevalence of monitoring systems in the EU (7%) may partly reflect these barriers, though this is not demonstrated.

In the United States, by contrast, employment is predominantly at-will and no federal statute regulates workplace electronic monitoring. A few states impose notice duties: Connecticut since 1998 (Conn. Gen. Stat. § 31-48d, 2025), Delaware since 2001 (Del. Code tit. 19, § 705, 2001) and New York since May 2022 (N.Y. Civ. Rights Law § 52-c, 2022). These regulate disclosure of monitoring, not monitoring itself. Worker-level evidence shows persistently high awareness of monitoring and, in 2023, an association between monitoring and worse well-being (APA, 2023, 2025, 2026); employer pressure to use AI rose from a fifth to a third of workers between 2025 and 2026 while a majority still lacked an official AI policy (APA, 2026). The OECD finding that 57% of workers support banning AI from dismissal decisions has no statutory anchor in US federal law (Lane et al., 2023), and US firms use evaluation tools far more intensively than European firms (Milanez et al., 2025). Where voice institutions exist, exposure to algorithmic management is contested; where they do not, the psychological costs of monitoring fall on workers.

4.3 OHS, psychosocial hazards and the Australian mid-way model [Risk (v)]

Australia occupies an instructive intermediate position. Its WHS regime imposes a positive duty on persons conducting a business or undertaking to manage psychosocial hazards, including job demands, low job control and poor organisational change management (Safe Work Australia, 2022). The duty is triggered by hazard identification and risk assessment: it disciplines employers after hazards are identified rather than giving workers an ex ante veto over monitoring technology, so it is reactive relative to German co-determination even though far-reaching. The resulting gap has been identified directly: the WHS system does not yet treat algorithmic oversight, digital surveillance or loss of worker autonomy as codified hazards, and the remedies proposed are to codify relational AI risks into WHS standards, to institutionalise worker-led job crafting and AI co-governance, and to expand the legal definition of safe work to include dignity, autonomy and procedural fairness, including rights to contest algorithmic decisions.

The monitoring evidence in Section 3.1 shows why codification matters: substantial exposure combined with low visibility, with 81% of surveyed Victorian union members saying surveillance had not been discussed with them (Farthing et al., 2026b) and a 72% to 9% gap between employer belief and worker awareness (AHRI, 2025; Farthing et al., 2026a). Worker attitudes align with the consultation-based regimes: 75% agreed that monitoring should be used only when clearly communicated and strictly necessary (Farthing et al., 2026a). Collective bargaining rarely reaches the issue (Howe et al., 2026). The reactive WHS duty therefore operates in a low-transparency, weak-voice environment. Consistent with this reading, the House of Representatives Standing Committee on Employment, Education and Training (2025) recommended a code of practice on the WHS risks of AI and automated decision making, confirming the WHS channel, rather than surveillance-specific statute, as the operative governance mechanism.

Singapore also relies on duty-holder action, but with strong tripartite scaffolding (Table 1). That roughly a third of assessed employees report stress or burnout (Ministry of Manpower, 2025b) suggests that voluntary employer measures alone do not eliminate psychosocial risk. The limits of the voluntary approach are visible in monitoring: 41% of Singapore workers feel constantly watched (ADP Research, 2025), and in November 2026 the Ministry of Manpower told Parliament that, while it recognises the mental fatigue that can attend closely monitoring and validating AI outputs, it has no plans to update the Tripartite Advisory on Mental Health and Well-being at Workplaces for AI-related tasks (Ministry of Manpower, 2026b).

4.4 Tripartite governance, skills and status [Risks (ii) and (iii)]

The OECD AI surveys find that training and worker consultation are the practices most consistently associated with better worker outcomes in AI-adopting firms (Lane et al., 2023), which aligns with Singapore's tripartite pairing of rapid adoption with negotiated guidelines and state-backed upskilling. PwC's 2026 archetypes show the distributional stakes: a 56% engine-room majority with limited AI advantage and weak access to learning is the workforce most exposed to Risks (ii) and (iii), because discretion over task allocation and learning is the mechanism through which those risks materialise; and AI advantage is concentrated by access to learning and scarce skills, both institutionally malleable, rather than by occupation per se.

A plausible conjecture is that skilled workers in France and Germany, anchored in trade-craft standards and the dual vocational (Beruf) system, display institutionalised scepticism towards automated advice and therefore lower automation bias, whereas Singapore's rapid, state-championed adoption leaves workers acutely aware of the risk of losing human judgement. It is a conjecture, not a finding. The OECD AI surveys publish related measures only at sector level (Lane et al., 2023), and the one country-level proxy identified, unverified reliance on AI output (Section 3.2), does not support the European side: unverified reliance is higher in Germany and France than in Australia and the United States, and highest in Singapore (Gillespie et al., 2025). Two strands of evidence are nonetheless consistent with the German side. Establishment panel data for 2019–2023 associate sustained AI adoption with a 14% increase in new apprenticeships among training firms and a shift of continuing training towards high-skilled employees (Muehlemann, 2025); and conflicts over AI introduction were less frequent in companies with a higher share of highly skilled workers, although the coefficient is not statistically significant (Krzywdzinski & Wotschack, 2026). For Singapore, the verified picture (Section 3.2) is high adoption alongside above-average unease, consistent with the conjecture, but these are self-reported sentiments, not measures of skill erosion.

4.5 Two faces of technology-facilitated harm [Risk (v)]

Technology-facilitated incivility has been described as "machine-mediated, human-triggered or both" (Cebulla, 2025), and Section 3.5 shows why the distinction matters: the two manifestations are governed by different bodies of law, measured by different instruments and at different stages of recognition.

Interpersonal misuse is the better measured, although national measures are not comparable and no US estimate exists. Where a survey asks specifically about digital conduct at work, prevalence is an order of magnitude higher than where a single cyber item is appended to a general harassment module, suggesting under-detection rather than rarity. The channels are the ordinary tools of work (work email in Australia; in Germany most people classified as bullied reported no online mistreatment), so such misuse largely extends existing workplace relations through employer-provided systems. Recognition is uneven and reversible: the EEOC's 2024 guidance was rescinded in 2026 (EEOC, n.d., 2026), while Australia's psychosocial code (Safe Work Australia, 2022) and the Fair Work Commission's handling of AI-drafted grievances (Wibmer v Fujifilm Data Management Solutions Pty Ltd, 2026) absorb the issue into general workplace law without a technology-specific instrument.

System-mediated harm is barely measured but shows a clear regulatory divergence. Australia's Full Bench decisions establish, for platform work, that a process with no human involved cannot comply with the Deactivation Code (Ayyub v Portier Pacific Pty Ltd, 2026; Hotak v Rasier Pacific Pty Ltd, 2025), an ex post but enforceable remedy; the EU reaches the same position by statute, though only for platform work and only on transposition (Directive [EU] 2024/2831, 2024). Singapore deliberately declines to regulate algorithms, so contesting a deactivation depends on a platform work association rather than a right to a human decision (Koh, 2024; Ministry of Manpower, n.d.-a). The United States has neither: Local Law 144 mandates audits and notice without a right to human review (New York City Department of Consumer and Worker Protection, n.d.), and iTutorGroup shows redress for automated exclusion only where it tracks a protected characteristic (EEOC, 2023). The two manifestations therefore sit on opposite sides of the voice institutions examined in Section 4.2: interpersonal misuse is addressed everywhere through general harassment, discrimination or WHS law; system-mediated harm only where a right to contest an automated decision exists, currently for platform workers in Australia and, prospectively, the EU. Where it exists, the test is relational rather than technical: the Full Bench did not ask whether the 85% threshold was miscalculated but whether a human had exercised judgment, and found that none had. That is the test implied by proposals to extend the definition of safe work to rights to contest algorithmic decisions.

5. Limitations

This paper is a qualitative institutional synthesis anchored to verified published statistics. Its limitations are as follows:

  1. Most of the evidence consists of perceptions, opinions and employer reports rather than measures of exposure or harm, and the non-self-reported measures are, with the exception of the deactivation cases, not AI-specific (Section 4.1); contrasts between regimes should not be read as differences in measured prevalence or harm.
  2. No source provides comparable measures across all five countries for more than a single indicator; conclusions about relative severity rest on institutional analysis, and worker-level and establishment-level magnitudes should not be compared numerically.
  3. The OECD AI surveys capture manufacturing and finance only, and only workers employed after AI adoption (OECD, 2023); their early-2022 fieldwork predates generative AI (OECD, n.d.).
  4. All empirical sources published before 2022 have been excluded to focus on recent and expanding AI uses (earlier works are cited in Section 1.1 for their theoretical content only). For Australia this removes the only nationally representative worker survey of monitoring prevalence identified (2018); the post-2022 evidence does not measure national prevalence directly.
  5. The People at Work survey sample is small and industry-skewed (Safe Work Australia, 2024), and iWorkHealth participation is voluntary (Ministry of Manpower, 2025b).
  6. Published sources do not provide the AI-specific, intersectional (e.g., gender-disaggregated) estimates needed to test intersectional hypotheses; the one gender split identified (Australian compensation claims) is not linked to AI exposure.
  7. Measures of interpersonal technology-facilitated harm differ in construct (perpetration, victimisation, recorded offences), threshold and reference period; the OSH Pulse harassment item is not digital-specific, the French barometer lacks published methodology, and no US estimate was identified. No survey measures the incidence of system-mediated harm, which rests on adjudicated cases and legislation.

6. Conclusion

The five relational risks of workplace AI are not intrinsic to the technology. They are governed by co-determination statutes, consultation rights, data protection law, OHS duties and tripartite institutions, or their absence. The verified evidence is consistent with the view that enforceable voice institutions, rather than voluntary ethics frameworks, mediate these risks, and it supports, across five jurisdictions, the proposition that AI tools reconfigure rather than merely automate and must be governed through social institutions, norms and participatory mechanisms as well as through audits and algorithms. The two manifestations of technology-facilitated harm illustrate the point: interpersonal misuse is absorbed, unevenly, into general harassment and WHS law, whereas system-mediated harm is remedied only where a right to a human decision has been created, as Australia's Fair Work Commission now requires and the EU directive will require once transposed, and as Singapore and the United States do not. For research, the priority is measurement: cross-nationally comparable worker-level data on AI-related risks, including the incidence of automated decisions that workers cannot contest. For policy, the toolkit exists in each regime; the question is whether it is applied to algorithmic systems before, rather than after, the harm.

References

ADP. (2025a, August 20). AI sentiments remain mixed in Singapore, but uncertainty highest among knowledge workers: ADP Research [Press release]. https://sg.adp.com/about-adp/press-centre/ai-sentiments-remain-mixed-in-singapore.aspx

ADP. (2025b, May 9). Australian workers top APAC stress levels amid rising workplace pressure, reveals ADP report [Press release]. Medianet. https://newshub.medianet.com.au/2025/05/australian-workers-top-apac-stress-levels-amid-rising-workplace-pressure-reveals-adp-report/100533/

ADP Research. (2025). People at work: A global workforce view 2025. https://sg.adp.com/resources/insights/people-at-work-2025-a-global-workforce-view.aspx

Ajunwa, I. (2023). The quantified worker: Law and technology in the modern workplace. Cambridge University Press. https://doi.org/10.1017/9781316888681

Aloisi, A., & De Stefano, V. (2022). Your boss is an algorithm: Artificial intelligence, platform work and labour. Hart Publishing.

American Psychological Association. (2023). 2023 Work in America Survey: Artificial intelligence, monitoring technology and psychological well-being. https://www.apa.org/pubs/reports/work-in-america/2023-work-america-ai-monitoring

American Psychological Association. (2025). 2025 Work in America survey: Topline data [Data set]. The Harris Poll. https://www.apa.org/pubs/reports/work-in-america/2025/topline-data.pdf

American Psychological Association. (2026). 2026 Work in America survey. Beyond skills and experience: Psychological capital in the modern workplace. https://www.apa.org/pubs/reports/work-in-america/2026/full-report.pdf

Andersson, L. M., & Pearson, C. M. (1999). Tit for tat? The spiraling effect of incivility in the workplace. Academy of Management Review, 24(3), 452–471. https://doi.org/10.5465/amr.1999.2202131

Australian HR Institute. (2025). Quarterly Australian work outlook: December quarter 2025. https://www.ahri.com.au/wp-content/uploads/AHRI-WorkOutlook-Report-2025-Q4_digital.pdf

Ayyub v Portier Pacific Pty Ltd [2026] FWCFB 81 (Austl.). https://www.austlii.edu.au/cgi-bin/viewdoc/au/cases/cth/FWCFB/2026/81.html

Betriebsverfassungsgesetz [BetrVG] [Works Constitution Act], as promulgated 25 September 2001, BGBl. I S. 2518, last amended 19 July 2024, § 87 (Ger.). https://www.gesetze-im-internet.de/betrvg/__87.html

Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–942. https://doi.org/10.1093/qje/qjae044

Cebulla, A. (2025). AI and workplace relations: A WHS framework for managing relational risks in workplaces. Journal of Industrial Relations, 67(5), 743-761. https://doi.org/10.1177/00221856251392987

Chagny, O., & Marzo, C. (2026). Artificial intelligence and social dialogue in France. European Trade Union Institute. https://www.etui.org/publications/artificial-intelligence-and-social-dialogue-france

Code du travail [Labour Code], arts. L1222-2 to L1222-4 (2008) (Fr.). https://www.legifrance.gouv.fr/codes/section_lc/LEGITEXT000006072050/LEGISCTA000006189417/

Conn. Gen. Stat. § 31-48d (2025). https://www.cga.ct.gov/current/pub/chap_557.htm

Del. Code tit. 19, § 705 (2001). https://delcode.delaware.gov/title19/c007/sc01/index.html

Dell'Acqua, F., McFowland, E., III, Mollick, E., Lifshitz, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organization Science, 37(2), 403–423. https://doi.org/10.1287/orsc.2025.21838

Directive (EU) 2024/2831 of the European Parliament and of the Council of 23 October 2024 on improving working conditions in platform work, OJ L, 2024/2831 (2024). https://eur-lex.europa.eu/eli/dir/2024/2831/oj/eng

Doellgast, V., Appalla, S., Ginzburg, D., Kim, J., & Thian, W. L. (2025). Global case studies of social dialogue on AI and algorithmic management (ILO Working Paper 144). International Labour Organization. https://doi.org/10.54394/VOQE4924

Doellgast, V., & Wagner, I. (2022). Collective regulation and the future of work in the digital economy: Insights from comparative employment relations. Journal of Industrial Relations, 64(3), 438–460. https://doi.org/10.1177/00221856221101165

Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306–1308. https://doi.org/10.1126/science.adj0998

European Agency for Safety and Health at Work. (2025a). First findings of the Fourth European Survey of Enterprises on New and Emerging Risks (ESENER 2024). Publications Office of the European Union. https://osha.europa.eu/sites/default/files/documents/first-findings-esener-2024_EN.pdf

European Agency for Safety and Health at Work. (2025b). OSH Pulse 2025: Occupational safety and health in the era of climate and digital change. Publications Office of the European Union. https://osha.europa.eu/sites/default/files/documents/OSH-pulse-2025-climate-digital-change_EN.pdf

European Foundation for the Improvement of Living and Working Conditions. (2026). Algorithmic management in European workplaces. Publications Office of the European Union. https://www.eurofound.europa.eu/en/publications/all/algorithmic-management-in-european-workplaces

Fair Work Act 2009 (Cth) (Austl.). https://www.legislation.gov.au/C2009A00028/latest/text

Farthing, S., Sacher, S., Santow, E., & Vos, A. (2026a). Drawing the line: Tech-enabled surveillance of Australian workers. Human Technology Institute, University of Technology Sydney. https://www.uts.edu.au/globalassets/shared-media/documents/_mcu/media/uts_hti-worker-surveillance-report_web.pdf

Farthing, S., Sacher, S., Santow, E., & Vos, A. (2026b). Surveillance creep: Technology-enabled worker surveillance [Snapshot report]. Human Technology Institute, University of Technology Sydney. https://www.uts.edu.au/globalassets/shared-media/documents/_mcu/media/hti-snapshot-report-worker-surveillance.pdf

Flynn, A., Powell, A., & Wheildon, L. (2024). Workplace technology-facilitated sexual harassment: Perpetration, responses and prevention (Research report 03/2024). Australia's National Research Organisation for Women's Safety. https://www.anrows.org.au/publication/workplace-technology-facilitated-sexual-harassment-perpetration-responses-and-prevention/

Gagné, M., Parker, S. K., Griffin, M. A., Dunlop, P. D., Knight, C., Klonek, F. E., & Parent-Rocheleau, X. (2022). Understanding and shaping the future of work with self-determination theory. Nature Reviews Psychology, 1(7), 378–392. https://doi.org/10.1038/s44159-022-00056-w

Gillespie, N., Lockey, S., Ward, T., Macdade, A., & Hassed, G. (2025). Trust, attitudes and use of artificial intelligence: A global study 2025. Key global findings and country insights. The University of Melbourne and KPMG. https://mbs.edu/-/media/PDF/Research/Trust-attitudes-and-use-of-AI_Country-Insights-Report.pdf

Hadj Larbi, A., Rakotobe, M., & Traore, B. (2025). Cyberviolences dans les établissements scolaires et dans la société [Cyberviolence in schools and in society]. In Institut national de la statistique et des études économiques, Économie et société à l'ère du numérique (Insee Références, édition 2025, pp. 37–49). https://www.insee.fr/fr/statistiques/fichier/8616883/NUM25.pdf

Hall, P. A., & Soskice, D. (Eds.). (2001). Varieties of capitalism: The institutional foundations of comparative advantage. Oxford University Press. https://doi.org/10.1093/0199247757.001.0001

Han, S., Harold, C. M., Oh, I.-S., Kim, J. K., & Agolli, A. (2022). A meta-analysis integrating 20 years of workplace incivility research: Antecedents, consequences, and boundary conditions. Journal of Organizational Behavior, 43(3), 497–523. https://doi.org/10.1002/job.2568

Hotak v Rasier Pacific Pty Ltd [2025] FWCFB 214 (Austl.). https://www.fwc.gov.au/documents/decisionssigned/pdf/2025fwcfb214.pdf

House of Representatives Standing Committee on Employment, Education and Training. (2025). The future of work. Parliament of Australia. https://www.aph.gov.au/Parliamentary_Business/Committees/House/Former_Committees/Employment_Education_and_Training/DigitalTransformation/Report

Howe, J., Kariotis, T., Bromberg, L., & Fagan, J. (2026). Is the boss watching, and are unions doing anything about it? Workplace surveillance regulation in Australian Enterprise Agreements. Labour and Industry, 36(1). https://doi.org/10.1080/10301763.2026.2625594

Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366–410. https://doi.org/10.5465/annals.2018.0174

Koh, P. K. (2024, September 10). Platform Workers Bill: Round-up speech by Senior Minister of State for Manpower Dr Koh Poh Koon [Speech transcript]. Ministry of Manpower, Government of Singapore. https://www.mom.gov.sg/newsroom/speeches/2024/0910-platform-workers-bill-round-up-speech-by-sms-koh

Krzywdzinski, M., Schneiß, D., & Sperling, A. (2025). Between control and participation: The politics of algorithmic management. New Technology, Work and Employment, 40(1), 60–80. https://doi.org/10.1111/ntwe.12293

Krzywdzinski, M., & Wotschack, P. (2026). Contesting the code: Labor–management conflicts over AI adoption and the role of power resources under German codetermination. Socio-Economic Review. Advance online publication. https://doi.org/10.1093/ser/mwag047

Lamers, L., Meijerink, J., Jansen, G., & Boon, M. (2022). A capability approach to worker dignity under algorithmic management. Ethics and Information Technology, 24(1), Article 10. https://doi.org/10.1007/s10676-022-09637-y

Lane, M., Williams, M., & Broecke, S. (2023). The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers (OECD Social, Employment and Migration Working Papers No. 288). OECD Publishing. https://doi.org/10.1787/ea0a0fe1-en

Löbner, M., Welzel, F. D., Jung, F., Pabst, A., Buß, A., Görres, C., Backhaus, N. P., Seidler, A., & Riedel-Heller, S. G. (2025). Repräsentative Studie zum Thema Mobbing in der Arbeitswelt in der Bundesrepublik Deutschland [Representative study on bullying in the world of work in the Federal Republic of Germany] (Forschungsbericht 655). Bundesministerium für Arbeit und Soziales. https://www.bmas.de/SharedDocs/Downloads/DE/Publikationen/Forschungsberichte/fb-655-repraesentative-studie-mobbing-in-deutschland.pdf

Milanez, A., Lemmens, A., & Ruggiu, C. (2025). Algorithmic management in the workplace: New evidence from an OECD employer survey (OECD Artificial Intelligence Papers No. 31). OECD Publishing. https://doi.org/10.1787/287c13c4-en

Ministère du Travail. (n.d.). L'information et la consultation du CSE [The information and consultation of the CSE]. Retrieved November 4, 2026, from https://travail-emploi.gouv.fr/linformation-et-la-consultation-du-cse

Ministry of Manpower. (n.d.-a). Platform Workers Act: What it covers. Government of Singapore. Retrieved November 4, 2026, from https://www.mom.gov.sg/employment-practices/platform-workers-act/what-it-covers

Ministry of Manpower. (n.d.-b). What is tripartism. Government of Singapore. Retrieved November 5, 2026, from https://www.mom.gov.sg/employment-practices/tripartism-in-singapore/what-is-tripartism

Ministry of Manpower. (2025a, July 10). Proportion of resident employees who personally experienced harassment at work by nature of harassment [Data table]. Manpower Research and Statistics Department, Government of Singapore. https://stats.mom.gov.sg/iMAS_Tables1/FEP/mrsd_FEPHH_T3_1_Workplace_Harassment.xlsx

Ministry of Manpower. (2025b, September 22). Written answer to PQ on workplace burnout. Government of Singapore. https://www.mom.gov.sg/newsroom/parliament-questions-and-replies/2025/0922-written-answer-to-pq-on-workplace-burnout

Ministry of Manpower. (2026a, April 30). MOM, NTUC, SNEF to form Tripartite Jobs Council to address AI's impact on workers, jobs and businesses [Press release]. Government of Singapore. https://www.mom.gov.sg/newsroom/press-releases/2026/0430-formation-of-tripartite-jobs-council

Ministry of Manpower. (2026b, November 3). Written answer to PQ on mental fatigue and AI. Government of Singapore. https://www.mom.gov.sg/newsroom/parliament-questions-and-replies/2026/1103-written-answer-to-pq-on-mental-fatigue-and-ai

Muehlemann, S. (2025). Artificial intelligence adoption and workplace training. Journal of Economic Behavior & Organization, 238, Article 107206. https://doi.org/10.1016/j.jebo.2025.107206

New York City Department of Consumer and Worker Protection. (n.d.). Automated employment decision tools (AEDT). City of New York. Retrieved November 4, 2026, from https://www.nyc.gov/site/dca/about/automated-employment-decision-tools.page

N.Y. Civ. Rights Law § 52-c (2022). https://www.nysenate.gov/legislation/laws/CVR/52-C*2

Organisation for Economic Co-operation and Development. (n.d.). OECD AI surveys of employers and workers. Retrieved November 4, 2026, from https://www.oecd.org/en/about/projects/aisurveysofemployersandworkers.html

Organisation for Economic Co-operation and Development. (2023). OECD Employment Outlook 2023: Artificial intelligence and the labour market. OECD Publishing. https://doi.org/10.1787/08785bba-en

Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. Human Factors, 52(3), 381–410. https://doi.org/10.1177/0018720810376055

Parent-Rocheleau, X., & Parker, S. K. (2022). Algorithms as work designers: How algorithmic management influences the design of jobs. Human Resource Management Review, 32(3), Article 100838. https://doi.org/10.1016/j.hrmr.2021.100838

Parker, S. K., & Grote, G. (2022). Automation, algorithms, and beyond: Why work design matters more than ever in a digital world. Applied Psychology, 71(4), 1171–1204. https://doi.org/10.1111/apps.12241

PwC. (2024). Global Workforce Hopes and Fears Survey 2024. https://www.pwc.com/gx/en/issues/workforce/hopes-and-fears/hopes-and-fears-2024.html

PwC. (2026). Global Workforce Hopes and Fears Survey 2026. https://www.pwc.com/gx/en/1/issues/workforce/hopes-and-fears.html

PwC Australia. (2024). Australian insights from the 2024 Hopes and Fears Survey. https://www.pwc.com.au/workforce/hopes-and-fears-survey/australian-insights-from-2024-workforce-hopes-and-fears-survey.html

PwC Singapore. (2026). Global Workforce Hopes and Fears Survey 2026: Singapore edition. https://www.pwc.com/sg/en/publications/hopes-and-fears.html

Qualisocial, Ipsos bva, & AvoSial. (2026). Baromètre harcèlement au travail : Sortir du statu quo [Workplace harassment barometer: Breaking the status quo]. https://www.qualisocial.com/guides-et-etudes/barometre-harcelement-travail-2026

Rainie, L., Anderson, M., McClain, C., Vogels, E. A., & Gelles-Watnick, R. (2023, April 20). AI in hiring and evaluating workers: What Americans think. Pew Research Center. https://www.pewresearch.org/internet/2023/04/20/ai-in-hiring-and-evaluating-workers-what-americans-think/

Raisch, S., & Krakowski, S. (2021). Artificial intelligence and management: The automation–augmentation paradox. Academy of Management Review, 46(1), 192–210. https://doi.org/10.5465/amr.2018.0072

Ravid, D. M., White, J. C., Tomczak, D. L., Miles, A. F., & Behrend, T. S. (2023). A meta-analysis of the effects of electronic performance monitoring on work outcomes. Personnel Psychology, 76(1), 5–40. https://doi.org/10.1111/peps.12514

Safe Work Australia. (2022). Model Code of Practice: Managing psychosocial hazards at work. https://www.safeworkaustralia.gov.au/doc/model-code-practice-managing-psychosocial-hazards-work

Safe Work Australia. (2024). Psychological health and safety in the workplace. https://data.safeworkaustralia.gov.au/sites/default/files/2024-02/Psychological-health-in-the-workplace_Report_February2024.pdf

Safe Work Australia. (2026). Key work health and safety statistics Australia 2026. https://data.safeworkaustralia.gov.au/reports/key-whs-statistics-australia/latest-release

Singapore National Employers Federation. (2026, October 30). SNEF survey: Human judgement remains central as employers adopt AI in HR [Press release]. https://snef.org.sg/news/snef-survey-human-judgement-remains-central-as-employers-adopt-ai-in-hr/

Statistisches Bundesamt. (n.d.). Belästigung am Arbeitsplatz [Harassment at the workplace] (Indikatoren zur Qualität der Arbeit, Dimension 7). Retrieved November 4, 2026, from https://www.destatis.de/DE/Themen/Arbeit/Arbeitsmarkt/Indikatoren-Qualitaet-Arbeit/Dimension-7/belaestigung-arbeitsplatzl.html

Thelen, K. (2018). Regulating Uber: The politics of the platform economy in Europe and the United States. Perspectives on Politics, 16(4), 938–953. https://doi.org/10.1017/S1537592718001081

U.S. Equal Employment Opportunity Commission. (n.d.). Small business fact sheet: Harassment in the workplace. Retrieved November 4, 2026, from https://www.eeoc.gov/small-business-fact-sheet-harassment-workplace

U.S. Equal Employment Opportunity Commission. (2023, September 11). iTutorGroup to pay $365,000 to settle EEOC discriminatory hiring suit [Press release]. https://www.eeoc.gov/newsroom/itutorgroup-pay-365000-settle-eeoc-discriminatory-hiring-suit

U.S. Equal Employment Opportunity Commission. (2024). Enforcement guidance on harassment in the workplace [Rescinded January 22, 2026]. https://www.eeoc.gov/laws/guidance/enforcement-guidance-harassment-workplace

U.S. Equal Employment Opportunity Commission. (2026, January 23). EEOC Commission votes to rescind 2024 harassment guidance [Press release]. https://www.eeoc.gov/newsroom/eeoc-commission-votes-rescind-2024-harassment-guidance

Wibmer v Fujifilm Data Management Solutions Pty Ltd [2026] FWC 835 (Austl.). https://www.fwc.gov.au/documents/decisionssigned/pdf/2026fwc835.pdf

Yao, J., Lim, S., Guo, C. Y., Ou, A. Y., & Ng, J. W. X. (2022). Experienced incivility in the workplace: A meta-analytical review of its construct validity and nomological network. Journal of Applied Psychology, 107(2), 193–220. https://doi.org/10.1037/apl0000870

This reading version was generated from the PDF by an AI conversion pipeline; the PDF remains the version of record.

📝 About this HTML version

This HTML document was automatically generated from the PDF. Some formatting, figures, or mathematical notation may not be perfectly preserved. For the authoritative version, please refer to the PDF.