See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/372310583

Preparing to work with artificial intelligence: assessing WHS when using AI in

the workplace

Article in International Journal of Workplace Health Management · July 2023

DOI: 10.1108/IJWHM-09-2022-0141

CITATION READS

1 306

3 authors, including:

Andreas Cebulla Genevieve Knight

Flinders University Flinders University
121 PUBLICATIONS 795 CITATIONS 28 PUBLICATIONS 292 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Andreas Cebulla on 07 December 2023.

The user has requested enhancement of the downloaded file.

 Pre-Print Version (final accepted)

Preparing to work with artificial intelligence: assessing WHS when using AI in the workplace

Authors:

Andreas Cebulla (College of Business, Government and Law, Flinders University, Adelaide, Australia),

andreas.cebulla@flinders.edu.au

Zygmunt Szpak (The University of Adelaide, Adelaide, Australia), zygmunt.szpak@adelaide.edu.au

Genevieve Knight (The University of Adelaide, Adelaide, Australia), genevieve.knight@adelaide.edu.au

Published as: Cebulla, A., Szpak, Z. and Knight, G. (2023), "Preparing to work with artificial intelligence:
assessing WHS when using AI in the workplace", International Journal of Workplace Health Management,
Vol. 16 No. 4, pp. 294-312. https://doi.org/10.1108/IJWHM-09-2022-0141

Abstract

Purpose – Artificial Intelligence (AI) systems play an increasing role in organisation management, process

and product development. This study identifies risks and hazards that AI systems may pose to the work health

and safety (WHS) of those engaging with or exposed to them. A conceptual framework of organisational

measures for minimising those risks is proposed.

Design/methodology/approach – Adopting an exploratory, inductive qualitative approach, the researchers

interviewed 30 experts in data science, technology and WHS; 12 representatives of nine organisations using

or preparing to use AI; and ran online workshops, including with 12 WHS inspectors. The research mapped

AI ethics principles endorsed by the Australian government onto the AI Canvas, a tool for tracking AI

implementation from ideation via development to operation. Fieldwork and analysis developed a matrix of

WHS and organisational-managerial risks and risk minimisation strategies relating to AI use at each

implementation stage.

Findings - The study identified psychosocial, work stress and workplace relational risks that organisations and

employees face during AI implementation in a workplace. Privacy, business continuity and gaming risks were

also noted. All may persist and reoccur during the lifetime of an AI system. Alertness to such risks may be

enhanced by adopting a systematic risk assessment approach.

 Pre-Print Version (final accepted)

Originality – A collaborative project involving sociologists, economists and computer scientists, the study

relates abstract AI ethics principles to concrete WHS risks and hazards. It translates principles typically

applied at the societal level to workplaces and proposes a process for assessing AI system risks.

Keywords – Artificial intelligence, ethics principles, risk assessment, AI Canvas, WHS/OHS,

organisational/peer relations, job control/workload, future work

This article reports on research on how the use of artificial intelligence (AI) in workplaces may affect

the health and safety of workers, and which processes could be applied to avoid or contain harm. The

COVID-19 pandemic, which led to increased use of advanced technologies, including technologies driven

by AI, accentuated the urgency for such exploration. Contributors to a recent special issue of the

International Journal of Workplace Health and Management (IJWHM, Vol. 15 No. 3, 2022) illustrated the

health implications of the growth in teleworking/working from home (Pataki-Bittó and Kun, 2022),

associated novel forms of remote workplace management and monitoring (Jeske, 2022), and new challenges

to managing relationships with customers frequenting a workplace in healthcare (Cumberland et al., 2022)

and retail (Mayer et al., 2022).

The COVID-19 pandemic also changed how organisations were going about their business. Notably,

it accelerated the digitalisation of workplaces, including the use of AI, the exploitation of ‘big data’ in

production and service industries (McKinsey Analytics, 2021; PWC, 2021), and process methods no longer

guided by human input of instructions but driven by unsupervised machine coding (Chhillar and Aguilera,

2022).

The Oxford Reference (2023) dictionary defines artificial intelligence as “[t]he theory and

development of computer systems able to perform tasks normally requiring human intelligence, such as

visual perception, speech recognition, decision-making, and translation between languages”. Everyday

examples of tools constructed with AI include automatic image content generation, facial recognition and

predictive texting (now also including ChatGPT). These tools share a reliance on the processing of large
 Pre-Print Version (final accepted)

quantities of data for their construction and efficacy, which is facilitated by today’s computational

technology. AI promises faster speed, and greater precision beyond the capacity of human beings. AI

systems can assist humans by taking over or eliminating dangerous or tedious tasks in everyday and work

settings (Shubhendu and Vijay 2013). In workplaces, AI has found application in workforce management,

process management, and product and service development tools. In workforce management, AI

applications are reported in human resources (Black and van Esch, 2020) and monitoring and surveillance

(Mateescu and Nguyen, 2019). In process management, AI applications can be seen in automated

warehousing (Bustamante et al., 2020) and smart factories (Capgemini, 2020; Wilson and Daugherty, 2018).

In product and service development, AI has found applications, amongst others, in agriculture to improve

irrigation and seeding (Rural Industries, 2016; Talaviva et al., 2020), medicine to improve diagnostic

instrumentation (Choy et al., 2018; Davenport and Kalakota, 2019) and in care settings to assist individual

support tasks (Loveys et al., 2022). In manufacturing, automated product inspection tools built on AI are

challenging traditional manual product quality assurance processes (CBInsights, 2022; Azamfirei et al.,

2023).

Whilst the potential gains from AI are rarely disputed, albeit perhaps occasionally exaggerated, using

AI can contribute to new health risks (Todoli-Signes, 2021). This study examined AI’s potential health and

safety impacts on workers. It adopts a broadly preventative position, which echoed Karanika-Murray and

Ipsen (2022, p. 259), who, in their guest editorial to the IJWHM Special Issue, argued that: “…there is a need

for primary and proactive initiatives and a focus on the whole organisation. Preventative approaches are

future-focused and require a mentality that places people as a priority.”

The research focused on the organisational implications of introducing AI as a new organisational

mediator or actor. Workplace practices are typically subject to WHS regulation, new International

Organization for Standardization (ISO) standards and product liability regulations, but in the field of AI the

application of such rules and regulations is largely untested (Dignam, 2020). We thus asked what

commercial and other users of AI need to be aware of – and do – to avoid or contain risks 1 associated with

1
We use the term ‘risk’ to refer to the potentiality of an event without implying any a priori (typically negative) connotation.
 Pre-Print Version (final accepted)

the disruption that new technologies, such as AI, cause to workplace processes, safety, and worker

wellbeing. This paper builds on and develops research previously reported in Cebulla et al. (2021, 2022).

The Workplace Challenge of AI risks

AI adoption in workplaces involves changes to processes or products that likely also entail re-arranging,

fragmenting or substituting for human activity. AI adoption then shifts the boundaries of common job

descriptions, content and expertise, requiring new role assignments as the operation and management of the

AI change established working practices (Faraj et al., 2018). Time previously allocated to tasks that are now

automated gets re-assigned, while the servicing of the AI is delegated up or down the organisational

structure (Pepito and Locsin, 2018). Workers may experience such processes negatively, lowering job

engagement (Braganza et al., 2020). When appropriately applied, AI can improve the quality of materials or

services, but reports on accidents involving AI (Arnold and Toner, 2021) and on prediction failures

(Chakravorti, 2022) demonstrate current limits.

A conventional view is to assess AI adoption risks – along with business risks more generally – from

a leadership perspective, as this is where strategic decisions are taken and responsibility ultimately rests. But

because digitalisation leads to increased interconnectivity (Dery et al., 2017; Baethge-Kinsky, 2020), a

leadership perspective may not suffice to capture these risks. There is an emerging debate about how AI use

in business induces complacency in business leaders as they place unwarranted trust in the technology (de

Cremer, 2022; Walsh, 2019). AI systems may also desensitise shopfloor workers to the riskier aspects of AI

as they replace human cognitive and physical activity (Baethge-Kinsky, 2020).

Current regulation of safe working practices barely protects against such risks, taking a

predominantly mechanistic approach and connecting safety to good work design (Safe Work Australia,

2015). They are rooted in a notion of workplaces that are physically protected from accident risk and where

supervisory systems can judge human capacity to adapt to workflows and successfully manage workloads.

Data-driven applications of AI have less obvious physical safety implications but increase the scope for
 Pre-Print Version (final accepted)

psychosocial harms by raising the cognitive and emotional demands associated with workplace tasks and

activities over and above workers’ capacities.

Methods

Commencing in 2020, this study developed an evidence-based conceptualisation of the implications of AI

adoption in a workplace for workers and management health and safety, involving data science and

technology experts, including AI specialists, and AI users in academia, the public sector and business, and

WHS inspectors. Given the novelty of the topic, the research adopted an exploratory, inductive qualitative

approach (Bingham and Witkowsky 2022; Swedberg 2020). The data collected from individuals familiar

with the uses of AI inside and outside of workplaces integrated impressions, stipulations and projections

with actual experiences from within workplaces. Based on the findings from that research, a framework of

actions for identifying and containing WHS risks when using AI in workplaces was developed.

Three phases of fieldwork

The research with data science and technology experts, AI users, and WHS inspectors was conducted as a

three-phased fieldwork program between July 2020 and February 2021, namely:

• Interviews with data science and technology experts, and WHS inspectors; and two online

workshops seeking input from an audience beyond technology and WHS specialists (phase 1);

• Interviews with representative of organisation using or preparing for the use of AI in the workplace

(phase 2); and

• An online workshop with WHS specialist to discuss the findings from our study (phase 3).

Sampling, participants and fieldwork conduct

During phase 1, interviews were conducted, individually, with data science and technology experts,

and WHS inspectors with interest and expertise in, or responsibility for, workplaces uses of AI. Participants
 Pre-Print Version (final accepted)

were identified via social media, notably Linkedin, and professional networks, such as regional public fora.

Thirty interviews were completed by October 2020 with participants from industry (16), government (5),

WHS specialists (4), academia (2), research organisations (2), and one representative of an AI professional

network. Interviews lasted between 40 minutes and about one hour; were conducted by phone or via video

conferencing tools (Zoom, MS Teams); and were recorded with the participant’s consent, which 15

participants granted. About one-third of the interviews were co-attended by two or more researchers, who

also took notes of the conversations, with particular attention given to detailing those that could not be

recorded. Whilst we took a pragmatic approach to determining the number of interviews, guided by

resources and deadlines framing the commissioned part of this research, the data collection achieved

saturation before all interviews had been collected (Baker and Edwards, 2012).

Parallel with these interviews, two public online workshops were conducted in August 2020,

advertised via the researchers’ and their employers’ social media sites. A selection of AI interest groups

identified in web searches was also approached directly. The workshops were intended to reach a wider non-

technical, non-specialist audience than the interviews, although data science and technology experts again

made up most of the 22 participants. A detailed participant breakdown is unavailable since no further

personal or professional data were collected besides registering emails and participants’ names.

Phase 2 involved in-depth interviews with organisations using, or having put into place structures

and process in preparation for using, AI, and further expert interviews. Participants in this phase were

selected for their experience in introducing AI in the workplace. In the absence of a register of AI users,

participants were identified through contacts established during phase 1 and additional searches of publicly

available sources such as AI industry networks, innovation centres and innovation labs (mostly university-

based). Other sources included websites advertising, promoting, selling or otherwise exploring and

discussing AI, as well as professional social media platforms (notably, LinkedIn).

Twelve individuals from nine organisations participated. They included senior managers and data

scientists in local government; data mining scientists, chief executives and WHS managers from three

Australian state and territory governments; chief executives and senior managers of a specialist
 Pre-Print Version (final accepted)

manufacturing company, a software company and a disability service provider. Semi-structured interviews

were conducted by phone or video conferencing, lasting between 40 and 80 minutes; about half of the

interviews were attended by two researchers, the remainder by one interviewer only. Interviews were

recorded where permission had been granted, and researchers took notes. As in phase 1, resources and

deadlines defined the extent of this fieldwork.

The final phase 3, conducted during March 2021, engaged 12 WHS inspectors identified by the

funding organisation prior to meeting in an online workshop. The workshop was conducted via

teleconference, lasted approximately 1.5 hours and was recorded with participants’ consent. All five

researchers attended the workshop, taking notes.

Sequential focus of research phases

The fieldwork phases adopted a sequential approach, whereby phases 2 and 3 built on the findings of their

respective prior phases.

In the phase 1 interviews, participants were asked to explore current or potential uses of AI in their

workplace and workplaces more generally, and any ethical or WHS matters that these uses raised or might

raise. Participants were invited to provide examples based on their own experience or observation, or as they

might have seen them reported elsewhere.

The workshops of phase 1 started with a short presentation about the research objectives. Brief

illustrations of current uses of AI in workplaces and the generic debate about the ethics of AI followed to

guide the conversation, which sought to generate further examples of AI risks in workplaces.

Phase 1 resulted in an initial itinerary of WHS hazards and risks when using AI in the workplace

(hereafter referred to as AI WHS risks), which was then further developed and populated with case

examples and proposed responses or preventions in phase 2 of our fieldwork between November 2020 and

February 2021.

This second phase sought to (1) explore how organisations used or intended to use, and prepared the

application of, AI at work, (2) understand management and employee roles in those processes, (3) identify
 Pre-Print Version (final accepted)

risk factors affecting worker safety, and (4) establish principles to safeguard their health. At the end of phase

2, a more detailed AI WHS risk list had emerged, including examples of how businesses had managed those

risks, distinguishing between different stages from the conception and development to the use of AI in the

workplace.

In phase 3, the findings were taken to the workshop with WHS experts for discussion and

exploration of implications of WHS management and monitoring in workplaces.

Conceptual framework

The fieldwork’s conceptual starting point was the AI ethics principles endorsed by the Australian

Government and developed in 2019/20 by the country’s Commonwealth Scientific and Industrial Research

Organisation (CSIRO), the principal government agency responsible for scientific research (Dawson et al.,

2019). They were presented to study participants to guide the conversations and stimulate discussion about

uses of AI in workplaces that may jeopardise meeting these ethics principles and what could be done to

reduce that risk.

The Australian Government’s AI ethics principles concerned the need for AI and its applications to

protect, preserve or instil: (i) human, societal and environmental wellbeing, (ii) human-centred values, (iii)

fairness, (iv) privacy protection and security, (v) reliability and safety, (vi) transparency and explainability,

(vii) contestability, and (viii) accountability (Figure I).

---------------------------

FIGURE I ABOUT HERE

---------------------------

Further to assist interviews and group conversations about AI WHS risks, we introduced the AI Canvas, a

conceptual model of the stages of AI development in a business context developed by Agrawal et al. (2018)

(see Figure II). The AI Canvas model identifies seven stages, from the initial ideation of potential AI uses in

an organisation, via design, set up and testing of the AI system, to full operation. This sub-dividing sought to

help reflections by focussing on specific AI implementation stages.

 Pre-Print Version (final accepted)

Combining the ethics principles and the AI Canvas into a matrix added a temporal, sequential

dimension, thus, to explore when as well as how AI systems implementation may contravene ethical

principles and, more specifically, pose actual WHS risks. That matrix was shared for reference and

discussion with study participants before (via email) or during (online screen sharing) the interviews and

workshops.

---------------------------

FIGURE II ABOUT HERE

---------------------------

Data analysis and framework development

Data analysis revisited the interview and workshop recordings, and aggregated and compared the notes

researchers had taken of the conversations. At least two researchers analysed interviews before the findings

were summarised collectively, drawing on all researchers’ notes.

The analysis had an emic focus (Gaber and Gaber, 2010), populating the matrix of AI Ethics

Principles and the AI Canvas with AI WHS risks and possible pre-emptive measures suggested by study

participants. The Ayoa Mind Mapping software was used to assist in this process. The matrix was updated

as new data were generated.

The researchers then applied the information collected in the fieldwork to develop a framework of

actions for identifying and minimising AI WHS risks, building on the AI Ethics/Canvas matrix and also

drawing on extant literature on topic.

Findings

Understanding AI ethics in the workplace

The workshops and interviews (phase 1) revealed early on that participants were struggling with giving

concrete meaning to some of the AI ethics principles, which were seen to be referring to similar or related

ethical challenges and values. In discussions, participants condensed the eight principles to just three, which

they considered to capture the range of principles effectively. The three aggregates reflected a concern for
 Pre-Print Version (final accepted)

human wellbeing (combining the first three of the ethics principles in Figure 1), safety (principles 4 and 5)

and accountability and oversight (principles 6 to 8).

The explorations then turned to applying these three overarching principles to workplace situations

to give them meaning specific to that context. In discussing the principle of human wellbeing, participants

identified potentially psychologically or physically harmful effects of ill-conceived or inadequately

developed AI systems. This included AI’s potential impact on workers’ control over their job and workload,

including changing task complexities and duration of work activities.

The safety principle, in contrast, was discussed in terms of potential threats to personal data security

and the manipulation of AI systems for personal advantage by those familiar with its functions (‘gaming’,

notably, in the context of performance measurement). Worker safety was, above all, understood to mean the

protection of privacy from AI intrusion and the prevention of associated harm. But it also included harm that

may result from the physical characteristics of workplaces changing with the installation of AI driven

technologies.

Finally, participants linked the accountability and oversight principle of AI systems to autonomous

decision-making systems changing human-machine interactions as those systems determine and potentially

overrule decisions previously taken by a human. Moreover, these systems may replace or alter human-

human (e.g., worker-supervisor) interactions (e.g., in human resource management applications) or reporting

lines when adding a third medium to traditionally two-person supervisor/peer relations.

In short, participants translated the abstract AI ethics principles to the workplace context as follows:

• Human wellbeing → job control and workload

• Safety → privacy and harm

• Accountability and oversight → supervisor/peer relations.

To distinguish this concretisation from the original AI ethics principles, we refer to it as the work

arrangements and governance (WAG). Because of their conceptual proximity, in the sections below, in the
 Pre-Print Version (final accepted)

following, the sequence of the WAG is changed, still commencing with job control and workload, but now

followed by supervisor/peer and organisational relations, then concluding with privacy and harm.

Detailing AI WHS risk using AI Canvas stages

With the three WAGs, the study defined the principal areas in which sound, ethical workplace relations and

conditions are worked out. The research next sought to establish how and when AI applications might

challenge these. The AI Canvas had initially been chosen to facilitate this part of the investigation because

its seven-stage model offered a focus on AI WHS risk specific to each stage and thus a granular account of

potential AI WHS risks. It was still possible to do so, but in working with the AI Canvas, participants again

sought simplifications to the instrument. Those less familiar with AI or machine learning (ML) terminology

or how business may approach their application from a commercial perspective found the specialist

language and technical connotations in the AI Canvas problematic. The solution was to retain the seven

stages of the AI Canvas but not to insist on maintaining all distinctions between stages when seeking to

identify AI WHS risks. The seven stages were thus aggregated into three broader categories, labelled

ideation (consisting of the AI Canvas stages of prediction, judgement, action), development (outcome,

training), and operation (input, feedback).

Table 1 shows the most frequently mentioned occasions and circumstances that participants thought

might produce WHS risks and hazards in relation to each WAG and AI Canvas stage. These examples were

collected during fieldwork phases 1 and 2. For most of the AI stages, participants identified a mix of

psychological (e.g., communication, work stress) or physical (e.g., changes to the workplace environment)

risks pertaining to employees alongside business continuity and reputational risks that may affect employers.

Contributions also repeatedly noted risks associated with the opaqueness of complex data systems

employed by AI and different levels of access and understanding of these amongst a workforce, potentially

giving rise to conflict, manipulation and inappropriate, unintended use. Few, though, remarked on the risk of

poor or inadequate coordination of human-AI interaction and ambiguous determination of ultimate

responsibility and oversight (Elish, 2019).

 Pre-Print Version (final accepted)

In several instances, it was found that AI WHS risks may be present at different stages of AI

implementation, either because they persist or impact later stages. For example, not giving sufficient

consideration to employee job satisfaction may originate during early implementation stages because of a

failure to conduct (adequate or any) pay-off assessments but emerge only during full operation. A

consequence of AI scope transgressions is that risks may feature as ‘boundary creep’ or ‘worker resistance’

during AI operations. These lasting consequences pointed to the importance of reflecting early on the

potentially multiple operational and organisational factors that AI implementation in a workplace affects.

In Table 1, the scenario of AI being used “out-of-scope” resulting in disgruntled employees is

captured as part of the Ideation and Operation phases. The risk in the Ideation phase is at the intersections of

“Supervisor/peer and organisational relations” and “Judgement”, and of “Job control and workload” and

“Judgement”. In the Operation phase, the risk is captured at the intersection of “Supervisor/ peer and

organisational relations” and the “Input” heading.

---------------------------

TABLE I ABOUT HERE

---------------------------

A more detailed version of Table 1 than can be presented here was developed into an AI risk assessment

scorecard, which may be used to support users in identifying and rating the likelihood and impact of AI

WHS risks, and thus to prioritise preventative or remedial measures (Cebulla et al., 2021).

AI Action Framework – applying findings to develop a model for containing AI WHS risks

The question for practitioners remained as to what to do to avoid or at least contain AI WHS risks.

To answer this question, we have used the data collected in this study, including preventative and remedial

measures suggested in interviews and workshops, to construct a framework of actions for identifying and

containing AI WHS risks. We took turns reviewing items included in the matrix, interview notes and
 Pre-Print Version (final accepted)

recordings for examples of identification or minimisation strategies that had been proposed or adopted,

yielding recommendations for pre-implementation preventative actions and post-implementation remedial

actions.

In addition, the framework development drew on a review of the literature undertaken concurrently

with the fieldwork (for details, see Cebulla et al., 2021), which contributed to populating Table I with risk

examples or confirming others identified in interviews and workshops.

The following sections summarise the action framework’s key messages. In doing so,

each AI Canvas category (in the columns of Figure III) is addressed sequentially, elaborating in each

instance on the three WAG (identified in the first (left) column). The WAG and the actions proposed

regarding an AI Canvas category are identified in italics at the start of each section, corresponding to entries

in Figure III.

---------------------------

FIGURE III ABOUT HERE

---------------------------

Ideation (prediction, judgement, action)

Job control and workload - selecting solutions offering maximum benefit at minimum loss.

The intersection of “Job control and workload” and “Prediction” in Table 1 triggers consideration of

how an AI system may impact the workforce. AI systems that significantly alter, reduce or displace jobs,

i.e., AI used mainly for cost reduction purposes, especially if they also fail to enhance working conditions,

bear psychosocial risks for those affected. The ideation stage defines the business problem an AI system is

intended to address. The solution that AI may offer can generate new value propositions with profound

implications for what the organisation and its parts do and how individual members, i.e., the workers, make

sense of what their organisation is doing (Wessel et al., 2021). This research suggests value propositions be

explored during the ideation stage.

 Pre-Print Version (final accepted)

Consideration should be given to AI’s potential for augmentation beyond the immediate commercial

objectives for AI use, such as more reliable or customised products and services (IEEE, undated; WEF,

2020). Could the greater productivity that AI use promises also translate into labour time savings to benefit

work-life balance (Crane, 2021)? What other potential might there be for promoting better working

conditions, including less physically or psychologically demanding work? How can job autonomy be

enhanced so as to improve mental wellbeing? Any reflection of AI use in this context would consider the

need and scope for rewriting job descriptions, modifying work tasks and work patterns and schedules, and

any implications for skills training and development. Some modifications may be more practically realisable

and acceptable to the workforce than others (ADAPT, 2017; van de Poel, 2016).

Supervisor/peer and organisational relations – configuring modes for communicating the purpose and

functioning of AI. The intersection of “Supervisor/peer and organisational relations” and “Prediction” in

Table 1 captures the need to think about how the purpose of the AI will be communicated. Detection of AI

risks is aided by inclusive communication, which is already used to stimulate positive attitudes to change

and AI buy-in (Baethge-Kinsky, 2020; Matsumoto and Ema, 2020; Makarius et al., 2020). It could also

benefit AI WHS risk awareness early in the ideation process. Communication across organisations would be

about more than providing information. It becomes about employee engagement, which includes fact-

finding, enabling organisations and their members to identify AI risks collectively, using two-way

communication channels. Interactive risk communication helps to anticipate unintended consequences,

which ought to cascade to the operation stage.

Phase 2 interviews indicated that the impact of AI on workplace communication should be

considered during the AI ideation stage, especially if new job roles or task delegation are involved, rather

than only minor changes to an employee’s daily activities. An AI system thus may become a new

intermediary, for instance, when wearable devices, such as augmented reality (AR) glasses, provide instant

access to information, which substitutes for consultation with colleagues, likely also affecting lines of

accountability. Automated performance monitoring systems are another example, which may not only add to
 Pre-Print Version (final accepted)

but replace conventional, inter-personal performance reviews. How is this likely to affect intra-

organisational relations? Human resource management provides early insights into the current limits of AI

use with lessons for other sectors (Charlwood and Guenole, 2022; Giermindl et al., 2021; Robert et al.,

2020; Tambe et al., 2019).

Privacy and harm - identifying ethical, moral, social principles, and personal and collective conditions and

rights. The intersections of “Privacy and harm” with “Prediction” and “Judgement” in Table 1 cue attention

to reputational risks and encourage one to go beyond financial motivations to determine if an AI solution is

worth pursuing. Aloisi (2020, p.52) has argued that new technology deepens the risk of ‘hierarchy and

control over the workforce’, with associated physical or psychological harm and infringement of privacy.

Ideation should consider these risks, their probability, and scope for prevention. Could AI-driven

organisational innovation or an over-reliance on it adversely affect the workers’ right to a healthy and safe

workplace? How is this likely to affect trust within organisations and amongst workers? How might this

affect organisational processes internally or reputation externally?

AI risks may also be introduced from the outside, via external contractors (Duke, 2022). Outsourcing

system development and management adds communication layers and complexity. Involved actors may not

speak the same language, have different agendas, knowledge, expectations, and (technical) understanding,

and aspire to other benefits/costs. Due diligence and risk management become of primary importance with

additional external actors and an added complexity of interactive risks.

Development (outcome, training)

Job control and workload - mapping potential disruptions to work processes and needs for changes. The

intersections of “Job control and workload” with “Outcome” and “Training” in Table 1 focus on how an AI

solution will alter the way people work. The challenge to job control and workload during the development

stage is the risk of new unwelcome outcome measures and the collection of uneven or intrusive employee
 Pre-Print Version (final accepted)

(performance) data to support AI development. These measures often signal the redefining of job roles and

uncertain, possibly inequitable impacts on individuals (Fountaine et al., 2019). The result may be resistance

to change (Holmström and Hällgren, 2021). Resistance is likely when the ideation stage has neglected

engagement with those directly or indirectly affected by the innovation.

Changing job roles will affect working patterns, and new divisions of labour may need to be planned

and mapped out. New job roles may emerge that require new positions and, thus, resourcing. Organisations

may need new staff or face skill shortages and an abundance of ‘old’ skills. Although primarily about

training the AI, this stage should also be used to identify secondary impacts on workplace capacities not

previously apparent.

Supervisor/peer and organisational relations - delineating lines of reporting and accountability. Table 1

shows that the intersections of “Supervisor/peer and organisational relations” with “Outcome” and

“Training” emphasise the importance of considering how AI can affect worker incentives and relationships.

In defining expected outcomes from AI and commencing testing, questions of judgement, decision-

making and authority edge to the fore. To the extent that role differentiation occurs, some job roles will have

a greater capacity to shape outcomes and to benefit from them than others. Those holding those roles may

have different dispositions towards working with new technologies, especially fundamentally opaque AI

technologies (e.g., Lebovitz et al., 2022). There may be direct (physical) and relational risks, such as the

gaming of AI systems (Beard and Longstaff, 2018), that may have beneficial effects for some but

detrimental effects on others – those not gaming the system and hence losing out.

To the extent that new technology forces individual workers to adapt, it affects intra-organisational

expectations and processes, including amongst peers and, notably, concerning supervisors. The training

stage may reveal new complexities of the proposed AI system, including its interconnectivity across the

organisation, requiring changes to traditional reporting lines, responsibilities and accountability.

Accountability requires mechanisms to address AI predictions and recommendations that conflict with

workers’ experience, intuition, or sense of justice, potentially causing dissonance and psychosocial harm.
 Pre-Print Version (final accepted)

Privacy and harm - anticipating direct and indirect effects of AI systems use on individual task performance.

The intersection of “Privacy and harm” with “Training” in Table 1 emphasises that AI systems may gather

or produce sensitive data. As organisational or product data are finally used during AI system testing, the

focus extends from personal physical and psychosocial harms to include those explicitly concerned with

privacy, which, in turn, may also relate to psychosocial harms. With data gathering, cyber security risks also

emerge. Workplace-specific AI systems likely incorporate process data with a discrete risk of disclosing

personal data, be they about how tasks are completed, work processes chosen, or concern socio-

demographics, pay rates and other data deemed critical to organisational efficiency calculations.

This type of data gathering and use entails surveillance risks. Information flowing into AI systems

may be used to improve an organisation’s aggregate performance and individual workers’ performance by

monitoring specific behaviours, ranging from toilet breaks to typing speeds (Gartner, 2019; Scassa, 2021;

Yu et al., 2018). With the use of wearables, intended or incidental surveillance may not end at the factory

gate or the office door. This additional data collection may not have been intended nor disclosed and thus

potentially breach prior agreement or legislation. Strict data use protocols and oversight should limit ‘usage

creep’.

Operation (input, feedback)

Job control and workload - maintaining competencies, capabilities and capacity of workers.

During AI operation, the focus shifts to the longer-term impacts of AI systems on operators and end-users

and their interactions with the system. Failure to ensure that relevant competencies are in place may

undermine business continuity should new processes be disrupted. The intersection of “Job control and

workload” with “Input” in Table 1 recognises the fact that AI systems may fundamentally change how

people work. The impact of AI on working conditions, such as changes in speed, tasks, decision-making,

and technical skill requirements, will become evident, emphasising the importance of providing AI skills
 Pre-Print Version (final accepted)

training to ensure safe and efficient use of AI systems and prepare individuals for collaborative work with

AI (Kolbjørnsrud et al., 2017). Makarius et al. (2020) conceive this process as one of the growing utility of

AI systems to organisations, with opportunities for workplace improvement, notably enhancing employee

autonomy for greater job satisfaction.

Supervisor/peer and organisational relations - securing impact monitoring, conflict detection,

communication and resolution. As inherited job roles and responsibilities change, new risks and hazards

emerge associated with worker unfamiliarity, and status change, including loss/degradation and emotional

reactions to working with AI (Hornung and Smolnik, 2022; Mirbabaie et al., 2022). As pointed out in the

intersection between “Supervisor/peer and organisation relations” and “Input” (Table 1), employees may

react with resistance.

Almost unique to AI-enabled work environments is ‘algorithmic distance’, which emerges when

‘organisational power [is] exercised through automated routines or algorithms’ (Bartley et al., 2019, p.8).

Supervisor/peer relations may, for instance, be automated via AI in performance review contexts, when

remote monitoring (e.g., movement sensors) replaces supervisor assessment, significantly modifying

feedback and potentially also appeals processes. Our research participants additionally alerted to a physical

and relational gap in workplaces when human interaction is replaced with machine-generated AI messaging

or team virtuality (Kirkman and Mathieu, 2005; Trenerry et al., 2021).

Participants disagreed and the literature does not provide a clear answer about whether physical or

relational gaps have a positive, negative, or neutral effect on WHS and worker wellbeing. The lack of

consensus may reflect a gap in current knowledge but also an ambiguity of AI systems’ impacts, which may

have ‘good’ and ‘bad’ effects depending on the mode of implementation, use and the worker demographic

affected (Lee, 2018; Mayer et al., 2020; 2022).

Privacy and harm - embedding system protections, quality assurance and wellbeing checks. Krzywdzinski

et al. (2022) found that workers expect new assisting AI technologies to contribute to WHS, and stress
 Pre-Print Version (final accepted)

prevention and seek evidence of that effect (Table 1, the intersection of “Privacy and harm” with “Input”).

The same arguably applies to personal data. Both call for feedback controls to be put in place to record the

AI systems’ human and organisational impact, such as for responding to technostress (Graveling, 2020;

Tarfdar et al., 2015); or modifying the speed, sequencing or monitoring of work where they cause new

endurance challenges (Jarrahi, 2021).

Agrawal et al.’s (2018) concept of feedback within their AI Canvas is about using outcomes to

improve the AI algorithm to achieve maximum benefits to the business. To this, we would add that this

objective should be achieved within its original confines and intended duration (Table 1, the intersection of

“Privacy and harm” and “Feedback”). A more comprehensive review than of the AI algorithm alone is

required to ascertain the continuing value of the AI outcomes in a workplace context. Feedback loops would

need to return to the initial stages of the risk and hazard assessment process at the ideation stage. It will need

to ask whether the implementation of the AI system has met or at least contributed to achieving a ‘good

workplace’ WHS objectives (specifically, enhanced workplace wellbeing as would have been defined at that

stage). The questions to ask at this point are: are those initial objectives still valid and desirable, and what

else should be done to use AI for ‘social good’ within the workplace?

Discussion

The framework presented here is a proposal for how AI risk may be assessed systematically and

preventatively in a workplace, the questions that ought to be asked and when they ought to be asked. It

invites self-reflection alongside pre-emptive impact assessment. The framework does not and cannot claim

to have conclusively covered known risks, whilst we also suspect new, unknown risks to emerge in the

future as AI development continues apace. In conjunction with the detailed risk assessment scorecard

previously developed during this study (Cebulla et al., 2021), the framework is a tool to assist with

recognising emerging AI WHS risks.

 Pre-Print Version (final accepted)

An important message emerging from our fieldwork and the literature to date is that consultation and

communication are central to countering organisational complacency when working with AI. It prescribes

how the framework may be used most productively. Sanderson et al. (2021, p.5) stress the importance of

‘diverse perspectives and collaborative teams in the design, development and use of AI technology’ to

promote the beneficent use of AI. To ensure comprehensive involvement, management, data managers, and

relevant business units (including those some steps removed from the area directly impacted by the AI

system) should be purposely engaged (Krzywdzinksi et al. 2022).

As our research noted the temporality of the AI WHS risks, that is, the same risk carrying weight

beyond its initial occurrence, persisting in similar or different shape later on, workforce engagement may

need to be embedded in organisational processes and culture. This need is accentuated by the AI Canvas’

inclusion of a feedback stage for reflecting on how an AI system has been implemented and is being used.

AI implementation is here understood as a loop of checks and balances, and checks again.

AI may thus change not only WHS but also how workplaces ought to prepare for accelerating change

and emergence of AI WHS risks which, like AI risks more generally, are incompletely understood and, for

now, hard to predict.

The use of AI in workplaces has then wider societal and regulatory implications. Societal, because

AI changes workplaces and communications and interactions in those workplaces, but also beyond, for

instance, through new ways of supply chain management. Like any technology, AI affects social relations.

The difference here is that, in the past, humans used technology to assist them in doing their work; today AI

may be used to determine what that work is and how it is to be done. Job roles and responsibilities become

interchangeable or inversed. Accountability, however, looks set to remain as is, i.e., with the performing

human – at least for now, creating a new imbalance.

In the absence of bespoke statutory WHS regulation of psychosocial harm from AI and impact on

workers beyond workplace boundaries, the fallback for keeping AI WHS risks in check is existing

legislation, such as anti-discrimination laws used to prevent or halt algorithmic bias discriminating based on

gender, ethnicity/colour or disability. In some instances, AI workplace impacts may fall under the auspices
 Pre-Print Version (final accepted)

of current WHS regulation, despite its limitations. These may be imperfect controls, but they nonetheless

require businesses to demonstrate to the outside world that their use of AI is fair and legal – and that they

have the means and evidence to demonstrate both.

As AI evolves, now reshaping work processes with the proliferation of generative AI, research must

continue to dissect the social and relational as well as commercial cost-benefits of that technology and

develop recommendations for managing those newly emerging risks, including through the means of WHS

oversight. This would be supported by a living inventory of AI risks and their impact on WHS, and risk

management and available risk mitigation strategies.

Conclusion

This paper has presented an action framework for assessing and containing WHS risks associated

with the use of AI in workplaces and identified in interviews and groups discussions with AI and WHS

specialists, and organisations using or preparing to use AI in the workplace. Study participants identified a

workplace risk dimension of AI beyond physical or psychological, work stress-related WHS risks and

hazards commonly identified in and monitored by safe-work guidelines. They stressed the link between AI

systems, their intended benefits or purpose, and their effects on workplace relations, (data) privacy and

business continuity. The risk assessment framework developed from those discussions suggests a sequence

of intra-organisational ‘Q&A’ (questions and answers) and engagement with a view to achieving a socially

and collegially aware workplace with sound WHS and employee welfare processes in place. It makes the

case for a systematic, continuous AI risk and hazard review process.

References

Agrawal, A., Gans, J. and Goldfarb, A. (2018), “A Simple Tool to Start Making Decisions with the Help of

AI”, Harvard Business Review. 17 April, available at: https://hbr.org/2018/04/a-simple-tool-to-start-making-

decisions-with-the-help-of-ai (accessed 3 March 2020).

 Pre-Print Version (final accepted)

Aloisi, A. and de Stefano, V. (2020), “Regulation and the future of work: The employment relationship as an

innovation facilitator”, International Labour Review, Vol. 159, No. 1, pp.47-69.

https://doi.org/10.1111/ilr.12160.

Arnold, Z. and Toner, H. (2021), “AI Accidents: An Emerging Threat”, CSET policy brief, Center for

Security and Emerging Technology, Georgetown University, Washington, DC, United States.

https://doi.org/10.51593/20200072.

Azamfirei,V., F. Psarommatis and Y. Lagrosen (2023), “Application of automation for in-line quality

inspection, a zero-defect manufacturing approach”, Journal of Manufacturing Systems, Vol. 67, pp. 1-22.

https://doi.org/10.1016/j.jmsy.2022.12.010.

Baethge-Kinsky, V. (2020), “Digitized Industrial Work: Requirements, Opportunities, and Problems of

Competence Development”, Frontiers in Sociology, Vol. 5 No. 33. https://doi.org/10.3389/fsoc.2020.00033.

Baker, S. and Edwards, R. (2012) “How many qualitative interviews is enough? Expert voices and early

career reflections on sampling and cases in qualitative research”, available at:

https://eprints.ncrm.ac.uk/id/eprint/2273/4/how_many_interviews.pdf (accessed 28 March 2023).

Bartley, T., Soener, M. and Gershenson, C. (2019), “Power at a distance: Organizational power across

boundaries”, Sociology Compass, Vol. 13: e12737. https://doi.org/10.1111/soc4.12737.

Beard, M. and Longstaff, S. (2018), “Ethical by Design: Principles for good technology”, available at:

https://ethics.org.au/ethical-by-design/ (accessed 17 April 2021)

Bingham, A.J. and Witkowsky, P. (2022), “Deductive and inductive approaches to qualitative data analysis.”

In C. Vanover, P. Mihas, & J. Saldaña (Eds.), Analyzing and interpreting qualitative data: After the

interview (pp. 133-146). SAGE Publications.

Black, J.S. and P. van Esch (2020), “AI-enabled recruiting: What is it and how should a manager use it?”,

Business Horizons, Vol. 63, Issue 2, pp. 215-226, https://doi.org/10.1016/j.bushor.2019.12.001.

 Pre-Print Version (final accepted)

Bustamante, F., A. Dekhne, J. Herrmann and V. Singh (2020), Improving warehouse operations—digitally,

available at : https://www.mckinsey.com/capabilities/operations/our-insights/improving-warehouse-

operations-digitally (accessed 24 February 2020)

Capgemini (2020), Smart Factories. How can manufacturers realize the potential of digital industrial

revolution, available at:https://www.capgemini.com/wp-content/uploads/2017/07/smart_factories-

how_can_manufacturers_realize_the_potential_of_digital_industrial_revolution.pdf. (accessed 9 July 2020)

CBInsight (2022), 12 Tech Trends To Watch Closely In 2022, available at:

https://www.cbinsights.com/research/report/top-tech-trends-2022/ (accessed 31 January 2023)

Cebulla, A., Szpak, Z., Knight, G., Howell, C., and Hussain, S., (2021), “Ethical use of artificial intelligence

in the workplace”, available at: https://www.centreforwhs.nsw.gov.au/knowledge-hub/ethical-use-of-

artificial-intelligence-in-the-workplace-final-report (accessed 6 April 2023)

Cebulla, A., Szpak, Z., Howell, C., Knight, G., & Hussain, S. (2022). “Applying ethics to AI in the

workplace: the design of a scorecard for Australian workplace health and safety”, AI & Society, 38, pp.919–

935, https://doi.org/10.1007/s00146-022-01460-9 .

Chakravorti, B. (2022), “Why AI Failed to Live Up to Its Potential During the Pandemic”, Harvard Business

Review, 17 March.

Charlwood, A. and Guenole, N. (2022), “Can HR adapt to the paradoxes of artificial intelligence?,” Human

Resource Management Journal, https://doi.org/10.1111/1748-8583.12433.

Chhillar, D. and Aguilera, R. V. (2022), “An Eye for Artificial Intelligence: Insights into the Governance of

Artificial Intelligence and Vision for Future Research”, Business & Society.

https://doi.org/10.1177/00076503221080959.

Choy, G., O. Khalilzadeh, M. Michalski, S. Do, A. E. Samir, O.S. Pianykh, J.R. Geis, P.V. Pandharipande,

J.A. Brink, and K.J. Dreyer (2018), “Current Applications and Future Impact of Machine Learning in

Radiology”, Radiology, Vol. 288, pp.318–328. https://doi.org/10.1148/radiol.2018171820

 Pre-Print Version (final accepted)

Crane, J. (2021), “How AI in the Workplace Could Shorten Your Workweek”, available at:

https://www.spiceworks.com/hr/hr-strategy/guest-article/how-ai-in-the-workplace-could-shorten-your-

workweek/ (accessed 26 January 2023)

Cumberland, D.M., Ellinger, A.D. and Deckard, T.G. (2022), “Listening and learning from the COVID-19

frontline in one US healthcare system”, International Journal of Workplace Health Management, Vol. 15

No. 3, pp.410-426. https://doi.org/10.1108/IJWHM-03-2021-0054.

Davenport, T.A. and R. Kalakota (2019), “The potential for artificial intelligence in healthcare”, Future

Healthcare Journal, Vol 6, No 2, pp. 94–98.

Dawson, D., Schleiger, E., Horton, J., McLaughlin, J., Robinson, C., Quezada, G., Scowcroft, J. and

Hajkowicz, S. (2019), “Artificial Intelligence: Australia’s Ethics Framework”, Data61 CSIRO, Australia,

available at: https://www.industry.gov.au/publications/australias-artificial-intelligence-ethics-framework

(accessed 28 March 2023)

De Cremer, D. (20220, “With AI entering organizations, responsible leadership may slip!”, AI Ethics, Vol.

2, pp.49–51. https://doi.org/10.1007/s43681-021-00094-9.

Dery, K., Sebastian, I. M. and van der Meulen, N. (2017), “The Digital Workplace Is Key to Digital

Innovation”, MIS Quarterly Executive, Vol. 16 No. 2, Article 4, available at:

https://aisel.aisnet.org/misqe/vol16/iss2/4.

Dignam, A. (2020), “Artificial intelligence, tech corporate governance and the public interest regulatory

response”, Cambridge Journal of Regions, Economy and Society, Vol. 13 No.1, pp.37–54.

https://doi.org/10.1093/cjres/rsaa002.

Duke, S.A. (2022), “Deny, dismiss and downplay: developers’ attitudes towards risk and their role in risk

creation in the field of healthcare‑AI”, Ethics and Information Technology, Vol. 24 No. 1.

https://doi.org/10.1007/s10676-022-09627-0.
 Pre-Print Version (final accepted)

Elish, M.C. (2019), “Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction”, Engaging

Science, Technology, and Society, Vol.5, pp.40-60. https://doi.org/10.17351/ests2019.260.

Faraj, S., Pachidi, S. and Sayegh, K. (2018), “Working and organizing in the age of the learning algorithm”,

Information and Organization, Vol. 28, pp.62-70. https://doi.org/10.1016/j.infoandorg.2018.02.005.

Fountaine, T., McCarthy, B. and Saleh, T. (2019), “Building the AI-powered organization”, Harvard

Business Review, July-August, available at: https://hbr.org/2019/07/building-the-ai-powered-organization

(accessed 25 February 2022)

Gaber J. and Gaber, S. (2010), “Using face validity to recognize empirical community observations”,

Evaluation and Program Planning, Vol. 33, pp. 138–146.

Gartner (2019), “The Future of Employee Monitoring”, 3/5/2019, available at:

https://www.gartner.com/smarterwithgartner/the-future-of-employee-monitoring/ (accessed 14 September

2022)

Giermindl, L.M., Strich, F., Christ, O., Leicht-Deobald, U. and Redzepi, A. (2021), “The dark sides of

people analytics: reviewing the perils for organisations and employees”, European Journal of Information

Systems, https://doi.org/10.1080/0960085X.2021.1927213.

Graveling, R. (2020), “The mental health of workers in the digital era: how recent technical innovation and

its pace affects the mental well-being of workers”, European Parliament, Directorate-General for Internal

Policies of the Union, Brussels, Belgium. https://doi.org/ 10.2861/986378.

Holmström, J. and Hällgren, M. (2021), “AI management beyond the hype: exploring the co-constitution of

AI and organizational context”, AI & Society. https://doi.org/10.1007/s00146-021-01249-2.

Hornung, O. and Smolnik, S. (2022), “AI invading the workplace: negative emotions towards the

organizational use of personal virtual assistants”, Electron Markets, Vol. 32, pp.123–138.

https://doi.org/10.1007/s12525-021-00493-0.
 Pre-Print Version (final accepted)

IEEE (undated), “A Call to Action for Businesses Using AI. Ethically Aligned Design for Business”,

available at: https://standards.ieee.org/content/dam/ieee-standards/standards/web/documents/other/ead/ead-

for-

business.pdf?utm_medium=undefined&utm_source=undefined&utm_campaign=undefined&utm_content=u

ndefined&utm_term=undefined (accessed 17 April 2021)

Jarrahi, M. H., Newlands, G., Lee, M. K., Wolf, C. T., Kinder, E. and Sutherland, W. (2021), “Algorithmic

management in a work context”, Big Data & Society. https://doi.org/ 10.1177/20539517211020332.

Jeske, D. (2022), “Remote workers' experiences with electronic monitoring during Covid-19: implications

and recommendations”, International Journal of Workplace Health Management, Vol. 15 No. 3, pp.393-

409. https://doi.org/10.1108/IJWHM-02-2021-0042.

Karanika-Murray, M. and Ipsen, C. (2022), “Guest editorial: Reshaping work and workplaces: learnings

from the pandemic for workplace health management”, International Journal of Workplace Health

Management, Vol. 15 No. 3, pp.257-261. https://doi.org/10.1108/IJWHM-06-2022-209.

Kirkman, B. L. and Mathieu, J. E. (2005), “The Dimensions and Antecedents of Team Virtuality”, Journal

of Management, Vol. 31 No. 5, pp.700–718. https://doi.org/10.1177/0149206305279113.

Kolbjørnsrud, V., Amico, R. and Thomas, R.J. (2017), "Partnering with AI: how organizations can win over

skeptical managers", Strategy & Leadership, Vol. 45 No. 1, pp.37-43. https://doi.org/10.1108/SL-12-2016-

0085.

Krzywdzinski, M., Pfeiffer, S., Evers, M. and Gerber, C. (2022), “Measuring work and workers. Wearables

and digital assistance systems in manufacturing and logistics”, Discussion Paper SP III 2022–301. Berlin,

Germany: WZB Berlin Social Science Center.

Lebovitz, S., Lifshitz-Assaf, H. and Levina, N. (2022), “To engage or not to engage with AI for critical

judgments: How professionals deal with opacity when using AI for medical diagnosis”, Organization

Science. https://doi.org/10.1287/orsc.2021.1549.
 Pre-Print Version (final accepted)

Lee, M. K. (2018), “Understanding perception of algorithmic decisions: Fairness, trust, and emotion in

response to algorithmic management”, Big Data & Society. https://doi.org/ 10.1177/2053951718756684.

Loveys, K. M. Prina, C. Axford, Ò. Ristol Domènec, W. Weng, E. Broadbent, S. Pujari, H. Jang, Z. A Han,

J. A. Thiyagarajan (2022), “Artificial intelligence for older people receiving long-term care: a systematic

review of acceptability and effectiveness studies”, Lancet Health, 3: e286–97.

McKinsey Analytics (2021), “The state of AI in 2021”, available at: https://www.mckinsey.com/business-

functions/quantumblack/our-insights/global-survey-the-state-of-ai-in-2021 (accessed 11 May 2021)

Makarius, E.E., Mukherjee, D., Fox, J.D. and Fox, A.K. (2020), “Rising with the machines: A

sociotechnical framework for bringing artificial intelligence into the organization”, Journal of Business

Research, Vol. 120, pp.262-273. https://doi.org/10.1016/j.jbusres.2020.07.045.

Mateescu, A., and A. Nguyen (2019), “Workplace Monitoring & Surveillance”, available at:

https://datasociety.net/library/explainer-algorithmic-management-in-the-workplace/ (accessed 6 July 2022)

Matsumoto, T. and Ema, A. (2020), “Proposal of the Model Identifying Risk Controls for AI Services”, The

34th Annual Conference of the Japanese Society for Artificial Intelligence, Kumamoto, Japan, June 12,

2020, available at: https://confit.atlas.jp/guide/event/jsai2020/subject/4N2-OS-26a-02/tables?cryptoId=

(accessed 29 September 2020)

Mayer, A.-S., Strich, F. and Fiedler, M. (2020), “Unintended Consequences of Introducing AI Systems for

Decision Making”, MIS Quarterly Executive, Vol. 19 No. 4, Article 6, available at:

https://aisel.aisnet.org/misqe/vol19/iss4/6 (accessed 17 May 2022)

Mayer, B., Helm, S., Barnett, M. and Arora, M. (2022), “The impact of workplace safety and customer

misbehavior on supermarket workers’ stress and psychological distress during the COVID-19 pandemic”,

International Journal of Workplace Health Management, Vol. 15 No. 3, pp.339-358.

https://doi.org/10.1108/IJWHM-03-2021-0074.
 Pre-Print Version (final accepted)

Mirbabaie, M., Brünker, F., Möllmann (Frick), N.R.J. & Stieglitz, S. (2022), “The rise of artificial

intelligence – understanding the AI identity threat at the workplace”, Electron Markets, Vol. 32, pp.73–99.

https://doi.org/10.1007/s12525-021-00496-x.

Oxford Reference (2023), “Artificial Intelligence”, available at:

https://www.oxfordreference.com/view/10.1093/oi/authority.20110803095426960 (accessed 25 January

2023)

Pataki-Bittó, F. and Kun, Á. (2022), “Exploring differences in the subjective well-being of teleworkers prior

to and during the pandemic”, International Journal of Workplace Health Management, Vol. 15 No. 3,

pp.320-338. https://doi.org/10.1108/IJWHM-12-2020-0207.

Pepito, J. A. and Locsin, R. (2018), “Can nurses remain relevant in a technologically advanced future?”,

International Journal of Nursing Sciences, Vol. 6 No.1, pp.106–110.

https://doi.org/10.1016/j.ijnss.2018.09.013.

PWC (2021), “AI Predictions 2021”, available at: https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-

predictions.html (accessed 11 May 2022)

Robert, L.P., Pierce, C., Marquis, L., Kim, S. and Alahmad, R. (2020), “Designing fair AI for managing

employees in organizations: a review, critique, and design agenda”, Human–Computer Interaction, Vol. 35

No. 5-6, pp.545-575. https://doi.org/10.1080/07370024.2020.1735391.

Rural Industries (2016), “Artificial intelligence”. Rural Industries Research & Development Corporation,

Charles Sturt University, Australia.

Safe Work Australia (2015), “Principles of good work design. A work health and safety handbook”,

available at: https://www.safeworkaustralia.gov.au/doc/handbook-principles-good-work-design (accessed 14

September 2022)
 Pre-Print Version (final accepted)

Sanderson, C., Douglas, D., Lu, Q., Schleiger, E., Whittle, J., Lacey, J., Newnham, G., Hajkowicz, S.,

Robinson, C. and Hansen, D. (2022), “AI Ethics Principles in Practice: Perspectives of Designers and

Developers”. arXiv:2112.07467 [cs.CY].

Scassa, T. (2021), “Privacy in the Precision Economy: The Rise of AI-Enabled Workplace Surveillance

during the Pandemic”, available at: https://www.cigionline.org/articles/privacy-in-the-precision-economy-

the-rise-of-ai-enabled-workplace-surveillance-during-the-pandemic/ (accessed 8 June 2021)

Shubhendu, S., & Vijay, J.F. (2013), “Applicability of Artificial Intelligence in Different Fields of Life”,

International Journal of Scientific Engineering and Research, Vol. 1 Issue 1, pp. 2347-3878.

Swedberg, R. (2020) “Exploratory Research,” In Elman, C., Gerring, J. and Mahoney, J. (eds.) (2020) The

Production of Knowledge: Enhancing Progress in Social Science, Strategies for Social Inquiry, Cambridge,

Cambridge University Press, pp. 17–41.

Talaviya, T., D. Shah, N. Patel, H. Yagnik and M. Shah (2020), “Implementation of artificial intelligence in

agriculture for optimisation of irrigation and application of pesticides and herbicides”, Artificial Intelligence

in Agriculture, Vol 4, pp.58-73, https://doi.org/10.1016/j.aiia.2020.04.002.

Tambe, P., Cappelli, P. and Yakubovich, V. (2019), “Artificial Intelligence in Human Resources

Management: Challenges and a Path Forward”, California Management Review, Vol. 61 No. 4, pp.15–42.

https://doi.org/10.1177/0008125619867910.

Todolí-Signes, A. (2021), “Making algorithms safe for workers: occupational risks associated with work

managed by artificial intelligence”, Transfer: European Review of Labour and Research, Vol. 27 No. 4, pp.

433–452. https://doi.org/10.1177/10242589211035040.

Trenerry, B., Chng S., Wang Y., Suhaila, Z.S., Lim, S.S., Lu, H.Y. and Oh, P.H. (2021), “Preparing

Workplaces for Digital Transformation: An Integrative Review and Framework of Multi-Level Factors”,

Frontiers in Psychology, Vol. 12 No. 620766. https://doi.org/10.3389/fpsyg.2021.620766.

 Pre-Print Version (final accepted)

Walsh, M. (2019), “When Algorithms Make Managers Worse”, Harvard Business Review, 8 May. Product

#: H04XR1-PDF-ENG, available at: https://hbr.org/2019/05/when-algorithms-make-managers-worse

(accessed 14 August 2020)

WEF (2020), “Companion to the Model AI Governance Framework. Implementation and Self-Assessment

Guide for Organizations”, available at: https://www.pdpc.gov.sg/-/media/Files/PDPC/PDF-Files/Resource-

for-Organisation/AI/SGIsago.pdf (accessed 10 September 2020)

Wessel, L. K., Baiyere, A., Ologeanu-Taddei, R., Cha, J. and Jensen, T. B. (2021), “Unpacking the

Difference between Digital Transformation and IT-enabled Organizational Transformation”, Journal of the

Association for Information Systems, Vol. 22 No. 1, pp.102-129. https://doi.org/10.17705/1jais.00655.

Wilson, H.J. and P. R. Daugherty (2018), “Collaborative Intelligence: Humans and AI Are Joining Forces”,

Harvard Business Review, July–August 2018, available at: https://hbr.org/2018/07/collaborative-

intelligence-humans-and-ai-are-joining-forces (accessed 11 February 2022)

Yu, Z., Du, H., Xiao, D., Wang, Z., Han, Q. and Guo, B. (2018), “Recognition of human computer

operations based on keystroke sensing by smartphone microphone”, IEEE Internet of Things Journal, Vol.

5, pp.1156-1168. https://doi.org/10.1109/JIOT.2018.2797896.
 Pre-Print Version (final accepted)

Figure I Australian Government AI Ethics Principles

Principles at a glance

1. Human, societal and environmental wellbeing: AI systems should benefit individuals, society and the
environment.
2. Human-centred values: AI systems should respect human rights, diversity, and the autonomy of
individuals.
3. Fairness: AI systems should be inclusive and accessible, and should not involve or result in unfair
discrimination against individuals, communities or groups.
4. Privacy protection and security: AI systems should respect and uphold privacy rights and data
protection, and ensure the security of data.
5. Reliability and safety: AI systems should reliably operate in accordance with their intended purpose.
6. Transparency and explainability: There should be transparency and responsible disclosure so people
can understand when they are being significantly impacted by AI, and can find out when an AI system
is engaging with them.
7. Contestability: When an AI system significantly impacts a person, community, group or environment,
there should be a timely process to allow people to challenge the use or outcomes of the AI system.
8. Accountability: People responsible for the different phases of the AI system lifecycle should be
identifiable and accountable for the outcomes of the AI systems, and human oversight of AI systems
should be enabled.

Source: adapted by the authors from https://www.industry.gov.au/data-and-publications/australias-artificial-

intelligence-ethics-framework/australias-ai-ethics-principles

Figure II Modified AI Canvas
Ideation
Prediction: Identify the key
uncertainty that you would like to
resolve.
Judgement: Determine the
payoffs to being right versus being
wrong. Consider both false
positives and false negatives.
Action: What are the actions that
can be chosen?
Source: adapted by the authors from Agrawal et al., 2018
Pre-Print Version (final accepted)
Development
Outcome: Choose the measure of
performance that you want to use to
judge whether you are achieving
your outcomes.
Training: What data do you need on
past inputs, actions and outcomes in
order to train your AI to generate
better predictions?
Operations
Input: What data do you
need to generate predictions
once you have an AI
algorithm trained?
Feedback: How can you use
the outcomes to improve the
algorithm?
 Pre-Print Version (final accepted)

Table I Risks and hazards of AI in the workplace

Ideation
Work Prediction: Identify the key Judgement: Determine the payoffs Action: the actions to choose
arrangements uncertainty
and
governance
Job control and -Displacement: with or without non- -Insufficient consideration of unintended -Adverse effects on worker or worker
workload displacing alternative consequence rights, incl. safe workplace, pay rates
-Differential augmentation – differential -Undermining human capabilities -Inequitable or burdensome treatment of
impact on workers -Personal flourishing undermined by worker
organisation gain being favoured

Supervisor/peer -Inadequate, or inadequately defined and -AI used out of scope -Inadequate chain of accountability,
and communicated, purpose of AI -AI undermining company core values reporting or governance structure,
organisational -Insufficient transparency, contestability outsourcing of design responsibility
relations -Gaming risk
-Context stripped from communication
between employees, replaced by
management by algorithm
-Worker manipulation or exploitation
-Undue reliance on AI decisions
-Lack of process for triggering oversight

Privacy and -Resolutions affecting ethical, moral and -Technical failure, human error, security -Physical and psychosocial hazards,
harm social principles, e.g., predicting health breach, processes/essential services unnecessary harm, avoidable death
conditions/pregnancy contravening impacted
privacy (in case employee monitoring for -Reputational risk
greater productivity)
-Over-reliance leading to diminished
diligence on-site

Development
Outcome Choose Training Data to train AI for
performance measures
Job control and -Workers impeded from
workload modifying outcomes
of/challenging AI
recommendations (e.g.,
automated advice processes
leaving queries unanswered)
Supervisor/peer -Outcome measures not
and aligned with healthy
organisational workplace dynamics (e.g.,
relations efficiency vs competition,
equity/fairness)
-Worker-AI interface
adversely affecting the status
of workers (e.g., differential
rewards)
better predictions
-Insufficient consideration given to
interconnectivity/interoperability
of AI systems, and their secondary
effects on job roles, tasks and
responsibilities
-Unrepresentative training data
-Data not for purpose (e.g.,
untrusted past indicators)
-Continuity and change in
responsibilities and accountability
Operation
Input Data for predictions after Feedback Using
training the AI algorithm outcomes to improve
algorithm
- Worker competences, skills (not) -Irreversible impacts on
meeting AI job responsibilities
requirements/overburdening -Inadequate integration
-Discontinuity of service (e.g., of AI into mechanical
failure to anticipate shock events, or electrical processes
seasonal factors) -No offline systems or
processes to review
veracity of AI
predictions/decisions
-Worker resistance (e.g., to data -Assessment process
sharing) review requirement
-Insufficient safety understanding
e.g., resulting from acceleration)
-Boundary creep (e.g., data
collection outside workplace)
-Supervisor/peer mediation
via/supervisor substitution with AI

Privacy and -Adverse, differential,
harm additional (unforeseen) work
task and process effects (e.g.,
acceleration across internal
supply chain)
-Cyber security vulnerability -Physical workplace impact (e.g., -Personal data storage
-Data leaks and disclosure of design/cobot, temperature) beyond time
personal information -Insecure data storage, cyber
security vulnerability

Source: Authors’ own creation based on interviews with AI, computer science, WHS experts, online
workshops & case studies
 34

Figure III Framework of actions for identifying and minimising AI WHS risks

Ideation Development Operation

Selecting solutions Mapping potential Maintaining

Job control and offering maximum disruptions to work competences,
workload benefit at minimum
loss
processes and needs
for changes
capabilities and
capacity of workers

Supervisor/peer Securing impact

Configuring modes for
and communicating
Delineating lines of
reporting and
monitoring, conflict
detection,
organisational purpose and
functioning of AI
accountability communication and
resolution
relations

Identifying ethical, Anticipating direct

Embedding system
Privacy and moral, social
principles, and
and indirect effects of
AI systems use on
protections, quality
harm personal or collective individual task
assurance & wellbeing
checks
conditions and rights performance

Source: Author's own creation.

View publication stats