Chapter 3: Exploring Bias in Artificial Intelligence in Digital Transformation and the Datafied Future of Work

Algorithms are increasingly affecting and making decisions on and around all of our daily lives, affecting everything from how much we pay for car insurance to transport schedules, what information we are presented with online, what healthcare we have access to among a wide range of other activities from the most personal to the mundane every day, and facial recognition, all of which are driven by code which contains inherent biases. Indeed, ‘AI systems can be trained to promote or discriminate, approve or reject’ (Gaskins, 2023, p. 423) contributing to the perpetuation or further entrenchment of particular bias/es. This chapter will critically explore and discuss HRM algorithms which are increasingly used to make recruitment and selection decisions (Carey & Smith, 2016) as well as informing workforce planning. Arguably, automated decision-making may drive HRM efficiency and value in organisations, but the potential implications for equality and diversity must be subject to scrutiny. Within the current extant literature, the use of such tools is largely referred to as ‘algorithmic management’ (Meijerink & Bondarouk, 2023). While there is a vastly growing body of literature surrounding the outputs of algorithmic decisions and how the data that are generated are then mobilised, oftentimes there is little discussion surrounding how human biases are inscribed in the code of HRM algorithms and, in turn, how the code itself which drives the algorithms serves to further entrench and perpetuate (intersectional) inequalities. This chapter will begin with a broad overview of the field of the current literature exploring the algorithmic perpetuation of inequalities in HRM decision-making, situating the importance of the debate for people in organisations, practitioners, as well as contemporary scholarly debates and discourse. A theorisation of AI bias/es in AI in digital transformation and the datafied future of work will be discussed in order to further situate research questions for the future field. It is argued here that in many organisations, algorithmic HRM and AI decision-making hold the potential to contribute to ideological ‘echo chambers and confirmation bias’ (Kalina, 2020) and, in turn, the further entrenching of existing inequalities and inequitable hiring decisions. This chapter will explore the role of AI in recruitment and selection, and what this means for diversity and inclusion, as well as the importance of the AI–human interface. The tensions between the efficiency gains in the process compared to the potential effects on the fairness of decision-making and, in turn, implications for equality and diversity must be balanced, both within digital transformation strategies in organisations and the datafied future of work more broadly.

When deliberating the vast shifts and developments in both technology and how HRM decisions are made in organisations, it is also important to consider some of the constants in organisational life, most notably the somewhat static nature of (intersectional) inequality, bias, and the fallibility of human and algorithmic decision-making, as well as accountability issues surrounding fairness and legal constraints (Tambe et al., 2019). Given that this chapter focuses on how human biases can be inscribed into the code of the HRM algorithms embedding and sustaining inequalities while assuming a veneer of objectivity, it is of value to consider a pre-algorithmic era of equality and diversity research to provide a scaffold for thinking about how the past has shaped the HRM landscape within which AI has been developed. In her 2006 seminal work, the late American sociologist Joan Acker (2006) argued, ‘Even organisations that have explicit egalitarian goals develop inequality regimes over time’ (p. 443). It is argued here that this notion still applies in most contemporary organisations. However, how inequality regimes are developed and perpetuated is now more commonly driven by algorithms and the written code, which analyses inputted data and generates decisions or results used for algorithmic management.

Given that algorithms crudely function on input to generate output, the coding element of algorithmic decision-making is a fundamental and crucial element of bias and (intersectional) inequality perpetuation. In short, biases written into the code will inform how the algorithm functions and the nature of the data and decisions generated. Particularly when loop algorithms, defined as a sequence repeated/executed several times, are considered, the potential for the inflation and sustaining of biases written into the code may be further intensified. Whereas ‘sequenced algorithms learn from users’ interactions, this process is called machine learning or ML. The ML process involves learning from algorithms to detect patterns that allow machines to make predictions of a given dataset’ (Gaskins, 2023, p. 417). However, it is also important to clarify the distinction between algorithms and AI, in that algorithms express decision-making and the decision-making process, whereas AI uses previously generated data to make decisions, albeit the two are inextricably linked.

Furthermore, it is also important when contextualising the functioning of organisations to consider and deliberate the wider external context within which they function. For example, the role of government at the macro-level, as well as the potential effects of AI-augmented public administration and policymaking which are playing an increasingly important role in the future of work and digital transformation. Clearly this has implications for equality and diversity, particularly given the linkage between HRM and workplace legislation. This chapter begins by exploring the code as a key driver for inequality and bias perpetuation in algorithmic decision-making. Then, it discusses HRM and algorithmic management more broadly, briefly drawing on the case of Amazon’s sexist AI tool, with practical macro-level recommendations then being made.

While research surrounding the importance of ethical and responsible code writing for algorithms is coming increasingly to the fore, much of this literature resides in the tech domain. Most notably, the work of Delecraz et al. (2022) makes the important contribution that AI should not be ‘the decision maker but a tool to assist human recruiters in their decision-making process’ (p. 7), and it is here where there is both opportunity and potential threat surrounding not only responsible human resource (HR) decision-making but also the datafied future of work, in that unconscious biases will remain. We must problematise the moral and actual agency ascribed to algorithms, both as standalone and how this shapes humans’ relationship and perception of the outputs and their interpretation. The attribution of agency to algorithms (Fritz et al., 2020) is an ongoing discussion that needs to be brought further to the fore in the wider HRM literature and in ‘algorithmic management’ scholarship more generally to understand further, as well as theorise the relationship between human decision-making elements and the datafied elements, the ‘human–computer’ interaction. It is here where further demystification of the ‘black box’ of ML models, AI, and the code behind it is further amplified. Indeed, Fritz et al. (2020) succinctly express the crux of the issue surrounding the inherent lack of transparency:

Anecdotally, the ascription of truth to insights or decisions that are generated by algorithms in many organisations are treated as truth or infallible as they have been non-human generated, with little consideration made of how the decision has been arrived at and the code behind it. However, Lee (2018) found that the nature of the tasks that are being completed algorithmically matters and influences perceptions of fairness regarding algorithms’ ‘perceived lack of intuition and subjective judgment capabilities contributed to the lower fairness and trustworthiness judgments’ (p. 1), but that human decisions invoked in participants increased trust and perceptions of fairness due to perceptions of human social recognition. It is here where notionally it is still of value to return to understandings of equality and diversity and inequalities in hiring decisions from a pre-algorithmic era to understand the biases that continue to exist and are written into the code of algorithms. It is widely acknowledged that models produced from ML and the code behind them are not guaranteed to be free from bias, and this is further exaggerated when the data they are built on come from discriminatory environments (Yohannis & Kolovos, 2022).

A well-known and deeply concerning example of algorithmic bias towards certain groups is the COMPAS tool, developed by Northpointe, Inc. (see Larson et al., 2016), a case management and algorithmic decision-making tool used by US courts, which assesses the likelihood of recidivism or re-offending. Most notably, it was found that ‘black defendants were far more likely than white defendants to be incorrectly judged to be at a higher risk of recidivism, while white defendants were more likely than black defendants to be incorrectly flagged as low risk’ (Larson et al., 2016). While there have been rebuttals as to the methodology and findings of the ProPublica’ (Larson et al., 2016) findings, predictive policing and legal decision-making are widely known to draw on ‘dirty data’, drawing on flawed, racially biased data (Richardson et al., 2019). Given that algorithms learn over time, the form of which is influenced by the type of ML and the data that ML algorithms are provided with, the importance of initial coding and source code data is critical; it is here where the key linkage to coders’ and data scientists’ human bias role comes to the fore. By the very nature of their role, data scientists must use their intuition and human perceptions to develop code and ML systems which contribute to algorithmic decision-making tools (Müller & Guido, 2016), the potential for unfairness and biased decision-making solidly remains. It is, however, also important when exploring bias in AI in digital transformation and the datafied future of work to consider the importance of diversity in the tech labour market.

There is a rampant disconnect and imbalance between diversity in the tech industry and the societal impact of artificial intelligence (AI) within organisations and society. If bias is to be mitigated, reduced, or even eradicated from, for example, HRM algorithms, the current pool of coders, those who write the code behind the algorithms, must be widened. Currently, the discourse focuses on the underrepresentation of women. To an extent, ethnic diversity, but at the top level, it is still disproportionately white men who hold the power of the industry. While this is not surprising, indeed, it is the status quo in most organisations, if algorithmic bias perpetuation is to be reduced through the writing of code that is more equitable, ergo ‘coding for equity’ and equitable outcomes, the tech sector more widely needs to become more inclusive, not only for social justice but also because of the (intersectional) biases that are already embedded in new technologies and which actively contribute to algorithmic oppression (Alegria, 2020; Benjamin, 2019; Noble, 2018).

However, there are a number of models being developed which aim to reduce and account for algorithmic biases; for example, Yohannis and Kolovos (2022) propose a FairML metamodel, which aims to automate bias measurement and mitigation in ML (p. 143). While there are several metrics available to coders both for bias measurement and reduction, critically, the human element of the perception of these biases remains perceptual; differences between equal and proportional fairness remain at the discretion of the coder themselves, as well as the decisions surrounding which debiasing algorithms are applied or mobilised again fall to the coder. The salience of this can also be highlighted by the identification that ‘users still need to understand all the supported debiasing algorithms and bias metrics in order to use them properly. Being able to help them construct a FairML model using a wizard that asks questions in natural language is important’ (Yohannis & Kolovos, 2022, p. 149). Hence, it again demonstrates the individual coder’s fundamental role and interpretation and their conscious and unconscious biases. Indeed, ‘Social justice in the data age demands not only further scholarly attention but urgent, increased societal awareness of systems of oppression’ (Yarrow, 2023, p. 1149), as well as more holistic approaches to systems architecture in regard to equality and diversity, algorithmic hygiene (Vassilopoulou et al., 2022), and systems bias prevention. Opportunities for voice building through code and community-led design processes are a domain where there is both hope and scope for positive change and inclusion; ‘design justice’ is a framework for analysis ‘of how design distributes benefits and burdens between different groups of people’ (Costanza-Chock, 2020, p. 23) calling for designers to give ongoing attention to how matrices of domination affect minoritised groups.

In terms of HRM and ‘algorithmic management’ more broadly, it is arguable that HRM by algorithms is more complex than reinforcing negative outcomes for workers only. Instead, HRM algorithms can simultaneously offer value to and foster worker autonomy (Meijerink & Bondarouk, 2023, p. 2). Indeed, when exploring bias in AI in digital transformation and the datafied future of work, it is important to also pay heed to the potential opportunities for positive impacts on how biases in not only algorithmic decision-making may be reduced but also consider opportunities for positive change and the reduction of inequalities. Given the global corporate interest in and increasing adoption of algorithmic management and the impacts that this will have not only on HR as a function but also, by implication, organisational practices and policies, there is an ever-pressing need to contribute to understanding how organisations can leverage AI for driving positive change and inclusion in their organisations, through ‘algorithmic inclusion’, whereby AI needs to be actively designed for inclusion (Kelan, 2023) to prevent the further replication of existing biases, which may also be further enhanced by liberatory design thinking (Gaskins, 2023). However, it is here where in our thinking, we can return to wider, pre-AI equality, diversity, and inclusion thinking and theory. It may be argued that algorithms serve as a contemporary algorithmic, datafied expression of existing inequality regimes (Acker, 2006), (biased) practices and policies in organisations, and notions of intersectional inequalities (Crenshaw, 1989) whereby existing inequalities remain. Rather than being perpetuated and further entrenched solely by humans, inequalities are further wrought by AI tools which are often perceived to be more objective, as well as efficient and financially advantageous. The black box of HRM is a well-documented and explored phenomenon in the extant literature and within scholarly debates. However, it may be asserted that there is a contemporary augmentation of the linkage between HRM practices and firm performance through the increasing use of AI in HRM, specifically around recruitment and selection automation. When deliberating the context within which AI, for example, is being adopted by organisations, specifically focussing on HRM, distinctions must also be made between the automation of data analytics and data management and active AI use for decision-making.

Interestingly, the Chartered Institute of Personnel and Development (CIPD) found in a recent survey of UK business leaders that the focus on investment for the organisations surveyed was greatest for HR information systems, but that critically, there is also a strong focus on the automation of recruitment and selection (CIPD, 2022). Given that many organisations source their information technology (IT) systems provisions from outside the organisation, this poses important questions for the level of control and understanding of how the systems function and on what data they are trained to function for what business-critical decisions are. Ultimately, it may be argued that due to the convoluted nature of algorithm and code supply chains, that organisations will be faced with high levels of opacity surrounding the origins of the data driving their decision-making. This is particularly problematic in automating hiring decisions, whereby existing inequalities may be perpetuated or even further worryingly, new, and different biases introduced. It is argued here that traceability in source coding¹ and a holistic approach to embedding equality, diversity and inclusion (EDI) consideration in systems architecture is as important as ensuring that the source code itself and the algorithms are designed and developed to account for biases and ensure their mitigation to foster accountability and corporate social responsibility (CSR) in the usage of AI in organisations.

It is argued here that this makes for a dangerous contemporary cocktail of AI and human-biased decision-making, which may be termed ‘Human-AI symbiosis in organisational decision-making’ (Jarrahi, 2018) that holds the potential not only to further deepen and entrench existing inequalities but to also accelerate their perpetuation, due to (mis)perceptions of objectivity and algorithmic inequality perpetuation. It is this mutual continuativeness where further understandings and research need to be focussed to not only contribute to existing scholarly debates and discourse but also, more urgently, to gain understanding as to how organisational equality regressions may be effectively countered to mobilise AI, algorithmic management in HRM for positive organisational change. It is herein where the challenges lie for scholars and practitioners alike, as well as for organisations to better understand algorithmic decision-making’s role in perpetuating inequalities. As discussed earlier, the inclusion of not only bias metrics and mitigation algorithms but also a streamlined approach in ML is imperative, particularly where there is little experience in fairness ML and programming (Yohannis & Kolovos, 2022, p. 143); as such it is argued here that the importance of bias training for coders must be further emphasised, as well as to contribute further to algorithmic hygiene for bias proofing (Vassilopoulou et al., 2022).

However, this chapter contributes to understanding how HR algorithms may mask inequality and discrimination and reinforce and replicate social and organisational inequalities. Currently, we are at a place where organisations must better understand how algorithms and models can be developed that mitigate, reduce, and ultimately, as far as possible, eradicate bias to ensure more inclusive and egalitarian outcomes while reducing or eradicating the biases present in the data that they are trained on examining. Anti-bias controls are urgently needed as an industry standard, and this is discussed in further detail at the end of this chapter as a practical recommendation. However, algorithmic anti-bias controls may be argued to be limited in their utility by the fact that they are not a root cause fix in the sense that it may be more effective to actively ‘code-in’ for certain biases and organisational-level inequalities when the decision-making tools are being developed, and then as second-layer add-on algorithmic anti-bias controls, to build a double-layer control mechanism into systems to further mitigate for bias introduction and perpetuation, potentially aiding a shift to more liberatory design practices (Gaskins, 2023).

In the following section, a case study of Amazon is discussed, drawing initially on the now infamous ‘sexist AI tool’, which was used for hiring employees (Dastin, 2018). The tool was found to discriminate against women, with the algorithm drawing on data from previously collected applicant data, which was disproportionally from men. However, little is known about other forms of bias and intersectional inequalities that may have been perpetuated. Amazon is, of course, not the only example of the application of AI tools for hiring decisions, but this case was one of the most widely reported instances globally in the last few years. Given the size and prominence of the organisation, this held significant potential to resonate with a wide number of other organisations and indeed customers and potential applicants more widely and raise awareness of the inherent biases in AI tools more broadly. While it was reported by various sources at the time that the tool was not solely used for hiring decisions, it was clear that the data on which the tool had been trained was predominantly from male applicants, which is unsurprising given that at the corporate level, men make up around 66% of Amazon’s employees, both globally and in the United States, around 68% of people managers, rising to 75% of senior leaders globally (Amazon, 2023b). Ultimately, AI tools become biased because of how they are trained and the data upon which they are trained. In the case of Amazon, in its simplest form, the algorithm will have picked up male dominance and interpreted male profiles as a success factor, predisposing it to select male candidates and then developing a pattern of bias towards women candidates. In turn, it may be argued that tools such as those designed to aid efficiency in the hiring process hold significant potential to perpetuate existing workforce representation patterns, such as gender bias, racial inequalities, and predispositions or resistance to certain groups or identity profiles. There is also an endemic lack of how algorithms could be audited, appraised, and regulated to ensure equality and inclusion (Kelan, 2023).

Further notable is that there is still a very high level of opacity surrounding the use of AI in hiring decisions, not only at Amazon but within many different organisational settings globally. In the case of Amazon, there is a disconnect in the rhetoric between striving to be an inclusive organisation that claims to be aware of multiple biases and how AI could be a positive force for change. Rather, AI hiring tools are shrouded in secrecy, with the organisation always refusing to comment. Furthermore, what is further pertinent and noteworthy of the case study is that there is (leaked) information available, which suggests that Amazon is still forging ahead with automated application evaluation (AAE), developed within its AI recruitment team. Naturally, there is little discussion on the Amazon website about the potential drawbacks of AI and ML-based hiring tools, but what is clear is that there is an acknowledgement of the need for elements of human interaction for some roles, stating:

Recruitment, selection, and, in turn, retention of employees are a core organisational strategic endeavour but also a resource-intensive activity where opportunities for bias/es and inequitable practices and processes are rife, thus rendering the automation of recruitment and selection and the associated decision-making processes a vast area of potential opportunity for cost savings and efficiency gains. However, we must deliberate on the human cost regarding equality of opportunity and the potential for bias entrenching and perpetuating. There is a growing body of literature that explores the equality and diversity issues surrounding AI, automation and the future of work, as well as the use of AI in recruitment, selection, and retention processes, including much hyperbole about the opportunities for cost savings, as well as a vast rise in providers of AI recruitment tools. Yet, there is far less insight into the perceptions and experiences of candidates/users (Horodyski, 2023) and ethical perceptions of AI usage in hiring (Figueroa-Armijos et al., 2023). Ethics, trust, and CSR are inextricably linked with equality and diversity, and in order to better understand how we can create accountability and CSR in the usage of AI in organisations, the end-user experience/candidate experience also requires deeper consideration. Currently, the commercial and media discourse around the adoption of AI in recruitment and selection focuses largely on economic gains, time savings, and overarching organisational benefits, yet discussions beyond potential bias in algorithms are somewhat underdeveloped. Further the scholarly discourse surrounding the impact of AI on marginalised employees’ experiences of AI is currently somewhat underdeveloped, which holds the potential to further silence the already marginalised, particularly the intersectionally disadvantaged, and further deepen inequalities.

The following conceptual model (Fig. 3.1) outlines the factors which need to be considered both at the meso organisational and macro-levels for the development of (intersectionally) inclusive and sustainable AI in digital transformation and the datafied future of work, whereby it is argued that for an overall reduction in AI opacity, building of end-user trust, and the reduction of the algorithmic bias perpetuation, a holistic approach is needed.

Many organisations are also using AI tech to automate decision-making around applications through video analysis using platforms such as hirevue.com, which utilises facial recognition software to inform hiring decisions and provides a range of automated solutions such as video interviewing, conversational AI, interview building and text recruiting and includes several large, well-known clients such as JustEat, Santander, Randstad, Cathay Pacific, Rio Tinto, Kraft Heinz, and Unilever to name but a few, demonstrating the widespread and ever-increasing usage of such AI tools across the globe. While organisations such as Hirevue and others have a range of AI ethical principles, there is little information publicly available as to how biases are accounted for and the algorithms’ debiasing regarding a number of intersectional inequalities. The gendered and racialised (Noble, 2018) issues surrounding algorithmic decision-making are well established in the literature and are a vastly growing area of concern. Given the ever-increasing popularity of facial recognition software in hiring decisions, it is imperative to discuss and bring to the fore the inherent issues in mobilising such software.

In her now infamous Netflix documentary: ‘Coded Bias’, Joy Boulamwini, who is also the founder of the Algorithmic Justice League and a leading scholar in the field of bias in facial recognition software, uncovers the racial biases/racism inherent in many algorithms, whereby the algorithm cannot and does not accurately recognise black faces. This is another example of the stark biases inherently written into algorithms.

Hunkenschroer and Luetge (2022) explore the ethics of AI recruiting, which they define as: ‘any procedure that makes use of AI for the purposes of assisting organizations during the recruitment and selection of job candidates’ (p. 977), but most notably, in their exploration of the extant literature, they map the ethical opportunities, risks, and ambiguities inherent in the use of AI in recruitment and selection decisions, including the introduction of algorithmic bias and obfuscation of accountability, as well as a potential loss of human oversight. When considering the obfuscation of accountability, it is here where we must also consider the role of (human) trust in algorithmic decision-making and the tensions between perceptions of improvement in the speed and efficiency of decision-making (van Esch & Black, 2019). Critically, they argue that ‘AI-enabled recruiting has moved from a peripheral curiosity to a critical capability’ (van Esch & Black, 2019, p. 730), given the potential efficiency gains and cost savings. Efficiency gains and cost savings may present new and different opportunity costs for marginalised or minoritised groups given the widespread gendered, racialised, and classed inequality perpetuation widely associated with algorithmic decision-making and algorithmic management.

However, it is also evident that while such critical capability is indeed important, there is distrust, and that for many organisations, there is, rightly, a reluctance to hand over all aspects of their recruitment and selection strategies to outsourced AI companies or indeed overall algorithmic control. Lacroux and Martin-Lacroux (2022), in their exploration of recruiter trust and perceptions of AI decisions in the resumé screenings, found that there are high levels of distrust in algorithmic decisions in this domain and that the ‘black box’ problem surrounding the use, understanding of, and trust in AI is coming more and more to the fore. It is here where pedagogical opportunities for HR educators and future HR practitioners become further evident. However, it would be remiss not to acknowledge and discuss the convoluted nature of code development and international code and AI tool supply chains, whereby coding is carried out and sourced from multiple locations around the world, whereby coding is increasingly outsourced to contexts such as India where labour costs are typically much lower, and in turn, demand for such services is booming (Tejaswi, 2021). It is here where fallacies surrounding potential differences in cultural values and norms, differences in legislative contexts, particularly surrounding equality and diversity, as well as differences in awareness of the nature of types of and severity of biases become further convoluted by global code supply chains and where opportunities for nuance within the coding for hiring tools, for example, may also be misguided and not sufficiently contextually bounded. Furthermore, what is also concerning is the notion that AI may instead replace human coders, which will present new and different challenges and threaten the jobs and livelihoods of outsourced coders in countries such as India, which is in and of itself a labour market equality issue.

Hunkenschroer and Luetge (2022) succinctly bring together the current dilemmas, issues, and opportunities in the field, arguing that:

However, what is notable in current scholarly and industry debates is the lack of formal governance surrounding algorithmic design, ML and AI, and while there are organisations, such as the aforementioned Algorithmic Justice League, which lobby for not only greater awareness of the inherent issues but also to raise awareness of the importance of the need for equitable and accountable AI, and

It is herein where a fundamental issue lies on a global scale; the lack of governance surrounding coding, programming, as well as the development of ML and algorithmic decision-making is palpable and should be of pressing global concern, both for corporate leaders and government,² which may also be argued to be compounded by the complexity and convoluted nature of global outsourcing of coding and AI supply chains.

In terms of future research directions, there is much opportunity to explore how different groups of (human) decision-makers perceive algorithmic decision-making and bias regarding the need for a hybrid approach between algorithmic and human decision-making and levels of trust in an algorithm and its code. Qualitative studies, in particular, will be important in addressing contemporary questions surrounding the extent of trust, scrutiny of code origin, and how organisations adopt algorithmic decision-making on a large(r) scale in the context of digital transformation and the datafied future of work. Further, there is much scope for research surrounding the origin of code and levels of awareness of bias(es), both conscious and unconscious, among coders themselves, particularly given that it is well known that coders are globally disproportionally male (Statista, 2022), as well as there is a lack of diversity in the tech industry more widely.

Further, and as mentioned earlier in this chapter, there are important pedagogical opportunities in the HR teaching domain. When we consider the skills requirements of HR practitioners of the future, it is imperative to develop courses and programmes which explore the ethics, limitations, and effects of equitable and meaningful hiring decisions in the context of algorithmic recruitment and selection to ensure that future practitioners are not only equipped with the skills to understand the implications of AI decision-making but also understand how they can counter the inherent algorithmic biases that are inherent in many tools today.

Ultimately, whether coding, programming, and algorithmic development is carried out by humans or AI, what remains clear is the need for code that has equity as a core control measure built in from the outset, where AI is actively designed for inclusion (Kelan, 2023), and which serves to analyse and mitigate in-built biases critically. Arguably, this begins with human vigilance (Silberg & Manyika, 2019) albeit this in and of itself is problematic given the pervasiveness of human biases and inequity. The core issues surrounding the algorithmic amplification of human bias are some of the most pressing issues in the development of contemporary mechanisms through which to mitigate inequalities and biases and contribute to a more equitable world.

Developing solutions and working practices for coders is one of the most pressing contemporary challenges for organisations committed to equality and diversity, as well as for society more broadly, given the pervasiveness of algorithmic decision-making in almost all facets of life.

The following practical recommendations, which focus on the root cause of inequitable algorithmic outcomes, the writing of and code building for algorithms, as well as strategic recommendations for stamping out biases present in the data they are trained on examining, seek to serve as a meaningful contribution to existing scholarly debates but also as a guide for practitioners:

• Training for coders on the importance of writing code that is sensitive to intersectional biases that are prevalent in society, such as gender, race, and class, as well as accounting for any organisation-specific over- or underrepresentations.

• Training on identifying and measuring bias in ML and algorithmic decision-making.

• Mandatory training and modules for programming students in universities to understand systemic biases and their role as coders and developers in ‘coding for equity’, as well as to counter the biases students potentially also may have previously encountered through EdTech in their own previous learning (see Gaskins, 2023).

• Code development for the checking of code for such biases.

• Promotion in the wider industry of ‘DevOps’ (see also Díaz et al., 2021) approach to testing for bias in ML and algorithms before it is ‘put to market’ or mode widely rolled out, to ensure ‘equity’ is a key metric, much like quality or performance for example.

While there are a growing number of studies that provide models that serve to mitigate bias (Bellamy et al., 2019; Fritz et al., 2020; Van Giffen et al., 2022), naturally, there is a lag in the academic literature, and it is fruitful also to be cognisant of wider industry debates and suggestions surrounding the area of ‘coding for equity’. For example, Christopher Bergh, CEO of Datakitchen, a US-based data software and data analytics company, champions an approach of ‘equity as code’ whereby ‘the tests that enforce equity are built into automated software applications that test, deploy and monitor the model 24/7’ (Bergh, 2021). This chapter argues that the approach of ‘equity as code’ holds significant potential for wider adoption in practice, but for equity to be built into both testing applications and algorithms, there need to be industry standards, which could potentially be built into international standards for organisations such as ISO 9001 quality management systems requirements (see ISO, 2021). This could potentially serve as an international impetus for organisations, and by implication, coders, and developers to adopt a proactive, measurable, process-oriented, and governed approach to building in equity measures into their algorithms.