From Greenwashing to Machinewashing: A Model and Future Directions Derived from Reasoning by Analogy

Automation, digitization, and machine learning have irreversibly entered the scene. However, at the moment, it remains unpredictable what impact, what level of disruption, and which threats and benefits digital technology will contribute to business and society. Hopes and fears, utopian and dystopian visions are equally discussed. Particularly, ethical questions arising from machine intelligence are prevalent and trending. But it is not all (ethical) gold that shines: Kevin Roose (2019) reported for the New York Times, from the World Economic Forum in 2019, describing the “hidden automation agenda of the Davos Elite” where business leaders were publicly praising and discussing ethical and “human-centered AI,” whereas in private talks with other managers, consultants and investors, shared that “they are racing to automate their own workforces to stay ahead of the competition, with little regard for the impact on workers” (Roose, 2019).

Corporations developing or working with machine intelligence respond to this unease among humans and their politicians trying to calm the worries with ethics programs and the hiring of ethics experts to “navigate the moral hazards presented by artificial intelligence without press scandals, employee protests, or legal trouble” (Knight, 2019). Business consulting company KPMG, for example, named “AI ethicist” among the “[t]op 5 AI hires companies need to succeed in 2019” (Fisher, 2019). Hitherto, the recent exit of Google’s Ethical AI co-leader casts at least doubt on the seriousness of such ethical engagement. Timnit Gebru, Google’s former researcher, “said she was fired by the company after criticizing its approach to minority hiring and the biases built into today’s artificial intelligence systems” (Metz & Wakabayashi, 2020). The media scandal that followed her exit and the firing of her co-head Margaret Mitchell has provoked broad skepticism about the actual reality behind corporate AI ethics programs (Johnson, 2021; Vincent, 2021). As a consequence, critical observers have drawn associations with a well-established business ethics concept, labeling AI ethics as “machinewashing”—derived from greenwashing as misleading environmental communication by companies (Obradovich et al., 2019). Now, what is machinewashing? The term has first been coined by Wagner (2018) and further refined by researchers from MIT Media Lab in a Boston Globe article, stressing that it represents a new form of greenwashing used to “[a]ddressing widespread concerns about the pernicious downsides of artificial intelligence (AI)—robots taking jobs, fatal autonomous-vehicle crashes, racial bias in criminal sentencing, the ugly polarization of the 2018 election—tech giants are working hard to assure us of their good intentions surrounding AI. But some of their public relations campaigns are creating the surface illusion of positive change without the verifiable reality. Call it “‘machinewashing’” (Obradovich et al., 2019).

Machinewashing is closely related to or used interchangeably with competing concepts that emerged in parallel: AI washing, (AI) ethics washing, ethical white washing, ethics bluewashing, and ethics theater (see Table 2). Given the evolving status of the discussion, research into the ethical issues raised by machinewashing is still in an early stage. Against this background, we reason by analogy to bring conceptual clarity into the machinewashing domain, bridging it with the established greenwashing domain (Cornelissen & Durand, 2014; Ketokivi et al., 2017; Vaughan, 2014). Thus, the guiding research questions of this article are: What are the antecedents, outcomes, idiosyncrasies, and underlying practices of machinewashing, where the source domain of greenwashing can inform the target domain of machinewashing (Ketokivi et al., 2017; Nersessian, 2008)? Furthermore, we ask: What are the similarities and differences, where the analogy does not fit, and where machinewashing goes beyond greenwashing given the disruptive momentum of AI? In other words, we analyze salient properties and relations that allow for transferring knowledge from the structural analogous greenwashing field. In addition, we stress the limitations of such a transfer in light of diverging theory assumptions that reveal unique characteristics of machinewashing and its distinct conceptual boundaries (Alvesson & Sandberg, 2011). Consequently, we summarize a conceptual model of machinewashing (see Fig. 1; Table 3) as a guiding framework that helps to organize and describe the various understandings and implications of machinewashing.

Although a scientific concept definition of machinewashing may help better understand its underlying practices and trigger focused (empirical) research, few attempts have been made to provide a concept definition (Nersessian, 2008). Therefore, machinewashing is defined as a strategy that organizations adopt to engage in misleading behavior (communication and/or action) about ethical Artificial Intelligence (AI)/algorithmic systems. Machinewashing involves misleading information about ethical AI communicated or omitted via words, visuals, or the underlying algorithm of AI itself. Furthermore, and going beyond greenwashing, machinewashing may be used for symbolic actions such as (covert) lobbying and prevention of stricter regulation.

Given the relevance and strength of the link between the greenwashing and machinewashing domain, we outline a research program that builds on the analogy-derived model as a foundation (Ketokivi et al., 2017). As greenwashing occurred in a range of literatures, we show how the analogy links to multiple domains and opens the way for empirical study of the diverse machinewashing practices (Table 4 and 5). Thus, we provide evidence of the analogy's structural soundness and factual validity, encouraging future research and interdisciplinary dialog about machinewashing with a set of possible research questions on the macro, meso, and micro-level (Ketokivi et al., 2017). We emphasize that the suggested theory domains should not be seen as an exhaustive list but as an initial starting point to advance our understandings of machinewashing and broaden the scope of the research program (Alvesson & Sandberg, 2011; Ketokivi et al., 2017). Finally, we discuss broader avenues for future research and practical implications.

Analogies are central to management and organizational theories (Cornelissen & Durand, 2014; Ketokivi et al., 2017; Swedberg 2014; Vaughan, 2014). Many important theories that evolved in previous decades build on analogies to introduce new ideas, provide explanations and comprehensive insights into complex subjects, and trigger scientific discourse (Astley & Zammuto, 1992; Cornelissen, 2005; Williamson, 1971). An analogy typically links a source and a target domain through structural mapping, thereby utilizing the knowledge of the source to draw implications in the target domain (Ketokivi et al., 2017). In the following, we reason by analogy to connect the machinewashing domain with the source domain of greenwashing. To do so, we invoke a structural mapping that connects and transfers elements from the greenwashing debate to corresponding elements of the machinewashing discourse to ultimately arrive at a new and theoretically informed model of machinewashing (Gentner & Smith, 2012; Ketokivi et al., 2017; Nersessian, 2008; Vaughan, 2014). It is essential to be aware of the potential limits of such a knowledge transfer, given the distinct characteristics of the machinewashing phenomenon. A helpful way to approach this challenge is Alvesson and Sandberg’s (2011) problematization strategy to identify and contest core assumptions underlying a given domain.¹ Thus, we use the problematization approach as an important means to stress the boundaries of the analogy, bringing forth salient elements of the machinewashing phenomenon.

Machinewashing should not be considered a mere extension of greenwashing. At its core, machinewashing relates to broader AI issues, which warrant treating machinewashing as a distinct phenomenon rather than a digital version of greenwashing (see Table 2). Thus, the presented conceptual analogy is informed by the following peculiarities that characterize machinewashing concerning these broader AI issues (Ketokivi et al., 2017; Nersessian, 2008).

In the sections above, we reasoned by analogy to structurally map the machinewashing phenomenon and delineate its unique conceptual boundaries in a new model (Gentner & Smith, 2012; Ketokivi et al., 2017; Nersessian, 2008; Vaughan, 2014). Given the novelty of the phenomenon, few attempts have been made to provide a conceptual definition of machinewashing (see Table 2). Yet, a scientific concept definition may help to better understand the diverse machinewashing practices and trigger a focused research program (Nersessian, 2008). Thus, a definition can assist future attempts to operationalize and engage in the empirical study. This is crucial when it comes to identifying and measuring machinewashing and its outcomes. Further, from a practical perspective, a conceptual definition can assist corporations in avoiding misleading about their AI ethics activities. Consequently, we believe there is a need for a comprehensive theory-informed machinewashing definition. Therefore, the following definition derives from the structural analogy and the idiosyncrasies of the outlined machinewashing model (Ketokivi et al., 2017; Nersessian, 2008).

Definition: machinewashing is defined as a strategy that organizations adopt to engage in misleading behavior (communication and/or action) about ethical Artificial Intelligence (AI) / algorithmic systems. Machinewashing involves misleading information about ethical AI communicated or omitted via words, visuals, or the underlying algorithm of AI itself. Furthermore, and going beyond greenwashing, machinewashing may be used for symbolic actions such as (covert) lobbying and prevention of stricter regulation.

Machinewashing as a novel and multifaceted phenomenon requires the close attention of practitioners, regulators, and the broader public, making it a highly interesting and timely subject of study for business ethics and related fields (Bietti, 2020; Johnson, 2015; Martin et al., 2019). Given the relevance and strength of the link between the greenwashing and machinewashing domains established above, we argue that the analogy-derived machinewashing model can constitute the basis for future research to study the diverse machinewashing practices (Fig. 1; Table 4). Thus, in the following, we provide evidence of the structural soundness and factual validity of the analogy by demonstrating how machinewashing: (1) links to existing theories underpinning the greenwashing debate, and (2) lends itself to multiple research questions for empirical study (Ketokivi et al., 2017). The below-described theories are grouped according to macro, meso, and micro levels, indicating the respective focus of the theories (Cornelissen & Durand, 2014; Jeurissen, 1997). In light of Alvesson and Sandberg’s (2013) problematization strategy, we emphasize the need to challenge the suggested theories concerning their underlying core assumptions and the distinct characteristics of machinewashing (see Idiosyncrasies of Machinewashing). Consequently, the proposed theories should not be seen as an exhaustive list but as an initial starting point to expand the analogy and broaden the scope of the machinewashing research (Table 5).

Future machinewashing research may focus on macro theories dealing with organizations and the broader environments they are embedded in. This includes studying machinewashing concerning macro-level institutions, such as markets and governments, and more general cultural traditions of a given context (Cornelissen & Durand, 2014; Jeurissen, 1997). To adapt and shape this institutional environment, organizations may adopt machinewashing practices (Scott, 2014). In this regard, the following macro-level theories offer fruitful avenues to shed light on machinewashing practices: legitimacy theory, corporate political activity, and resource-dependence theory.

Legitimacy refers to the extent to which an organization operates according to a given set of social rules, such as formal and informal ones (Long & Driscoll, 2008; Suchman, 1995; Suddaby et al., 2017). Legitimacy theory follows the idea that organizations strive for social approval of their conduct in a given setting while avoiding disapproval to ensure their business continuity. Thus, to maintain what has been termed “social license to operate,” organizations engage in a constant legitimization process (Melé & Armengou, 2016). Legitimacy theory can help shed light on machinewashing practices that aim at various types of legitimacy—pragmatic, moral, and cognitive (Bietti, 2020; Greenwood et al., 2011). For instance, previous research associates greenwashing with an organization’s failure to attain pragmatic legitimacy (Seele & Gatti, 2017). Analog, this raises the question: how does machinewashing relate to different types of legitimacy—pragmatic, moral, and cognitive (Long & Driscoll, 2008)? Previous legitimacy research also associates corporate ethics codes with strategic self-interest, which may lead to symbolic isomorphism (Long & Driscoll, 2008). Similarly, today’s AI ethics codes may serve the organizational self-interest and gradually converge, such that a hypocritical approach to ethics may become the new social norm in the institutional context (Jobin et al., 2019; Long & Driscoll, 2008). Thus, legitimacy theory can also illuminate mimetic pressure from the institutional environment regarding how mimetic pressure in the market might affect a firm’s use of machinewashing. This also raises the question: which kind of pressure (mimetic or normative) is more influential in adopting machinewashing practices?

Future research may also benefit from the theoretical insights of corporate political activity (CPA) research, including lobbying and public affairs. As shown above, machinewashing is strongly associated with legislative, judicial, and academic lobbying (Bietti, 2020; Floridi, 2019; Ochigame, 2019; Rességuier & Rodrigues, 2020). Here, the CPA and public affairs lenses may help study machinewashing practices of organizations that are actively shaping their institutional environments. As discussed by den Hond et al. (2014), corporations may simultaneously play on two chessboards, the self-regulatory chessboard and the CPA chessboard, to achieve overall favorable outcomes in the non-market sphere. Technology organizations may similarly play on the AI ethics chessboard while engaging in the policymaking chess game to lobby against governmental regulations (Floridi, 2019; Yeung et al., 2020). However, organizations risk being exposed (machinewashing) as players on both chessboards, which may adversely impact their reputation and legitimacy if these two strategies are not aligned (den Hond et al., 2014; Rehbein et al., 2018). Interesting future research questions may revolve around the conceptual framing of two chessboards' logic, including institutionalization and paradoxes resulting from misalignment. Questions may include: Does machinewashing distract society from questioning the limits of current AI ethics programs and from pushing governments to adopt stricter regulations? How are lobbying expenditures against more stringent AI regulations related to organizational spending on AI ethics programs (Zuboff, 2021)? Which role do societal watchdogs or internal whistleblowers play in exposing a misalignment of the two chessboards? In addition, the CPA theory lens is beneficial in investigating the rising academic lobbying: to what extent does organizational funding of public research favor corporate interests? Does the lobbying of academic institutions undermine the independence of academia?

Resource-dependence theory (RDT) argues that firms' long-term survival and growth hinges on access to critical resources. Future research may build on this theoretical angle when studying machinewashing as a nonmarket organizational strategy used to “reduce environmental interdependence and uncertainty” (Hillman et al., 2009, p. 1404). Here, AI ethics and other related firm activities can be observed as mechanisms to reduce external dependence for critical resources and create a favorable non-market environment to secure essential resources from external stakeholders (Mellahi et al., 2016; Shirodkar et al., 2018). From this perspective, the flow of digital (behavioral) data may be perceived as a crucial resource that machinewashing practices help to secure (Zuboff, 2019). Thus, interesting questions may relate to how resource pressure impacts the adoption and use of machinewashing? How is AI used as a resource itself to engage in machinewashing and influence external stakeholders? Previous research also highlights the value of RDT when understanding boards (Hillman et al., 2009). In particular, RDT may help to shed light on the way in which organizations may use AI ethics boards to limit dependence or gain certain resources, as the case of Google’s ethics panel composition has shown (Metzinger, 2019).

Organizational-focused theories can be utilized to better understand organizational antecedents and outcomes of machinewashing. Thus, organizational-focused theories are on an intermediate level between macro and micro-theories, opening the door to a better understanding of internal conditions under which organizations engage in machinewashing practices (Cornelissen & Durand, 2014). In this regard, future research may draw on insights from organizational institutionalism, instrumental/deliberative CSR, and signaling theory.

Organizational institutionalism. A fruitful avenue of future research is the decoupling concept as discussed in organizational institutionalism (Greenwood et al., 2018). Decoupling focuses on the potential gap between firms’ policies and practices or, more precisely, means and ends (Bromley & Powell, 2012). Decoupling is relevant for machinewashing research, particularly to observe formal AI policies adopted by technology corporations concerning their daily business conduct. CSR and greenwashing research highlight the need to study talk action dynamics and, in light of aspirational CSR talk (Christensen et al., 2020a, 2020b; Glozer & Morsing, 2020). Firms’ resources dedicated to ethical AI may have limited or no relation to their intended AI goals. Thus, research questions along these lines include to what extent are organization AI ethics principles aligned with day-to-day practices? How do internal procedures and the goal to preserve organizational efficiency relate to the adoption of machinewashing practices? What role can ethics boards play in ensuring that ethics guidelines and codes are translated to daily practice? Another interesting question to explore is to what extent are specific machinewashing practices (already) institutionalized within a given organization?

Instrumental and deliberative CSR. Communication about ethical AI is often labeled as untrustworthy, aspirational, and not least as machinewashing (Benkler, 2019; B. Mittelstadt, 2019; Truby, 2020). Similar concerns have previously been raised about CSR as a mere marketing or public relations tool, where a substantial gap between symbolic and aspirational managerial ‘talk’ and the actual upholding of social responsibility standards prevails (Christensen et al., 2017; Haack et al., 2012; Winkler et al., 2020). In response to such critique, recent research observes how communication can move from an instrumental to a deliberative mode (B. D. Mittelstadt et al., 2015; Palazzo & Scherer, 2006; Scholz et al., 2019; Winkler et al., 2020). Seele and Lock (2015) outline a pathway about how idealistic principles can be translated into practice via discourse ideals. Further, Winkler et al. (2020) develop a framework of agnostic rhetoric that may facilitate the enactment of aspirational talk. In a similar vein, Christensen et al. (2017) emphasize a license to critique, outlining that ongoing communication (involving criticism and contestation) about sustainability standards is necessary to develop, adjust, and fine-tune standards as functional governance tools. Consequently, such deliberative approaches may also represent fruitful avenues for future machinewashing research, triggering important questions: Can a deliberative approach to AI ethics offset the lack of societal watchdogs and assist in transferring principles into practice? Who controls the current AI ethics discourse, and why? How are power and information asymmetries in the AI ethics discourse related to machinewashing practices? How can symbolic practices (e.g., ethics working groups and multi-stakeholder partnerships) be turned into meaningful and constructive discussions on ethical AI?

Signaling theory focuses on the behavior of a sender (e.g., a corporation) and a receiver (e.g., the customers, the public) in a communication process that is characterized by diverging information access (Connelly et al., 2011). Thus, signaling theory leans towards micro-levels, dealing with how the organization “signals” information to receivers, who choose the way of interpreting the communicated signal (Connelly et al., 2011). In greenwashing literature, the disclosure of corporate signals is studied as a form of deliberative communication of positive information that lacks observable attributes (Bowen, 2014). From this perspective, the corporate engagement in symbolic practices of ‘ethical talk’ about AI can be an effective signal toward society, underlining the corporate commitment to ethical values (Walker & Wan, 2012; Wu et al., 2020). In this regard, signaling theory stresses that corporations might signal false claims “if initial financial payoffs outweigh future losses once the truth becomes known” (Whelan & Demangeot, 2015, p. 1). Thus, signaling theory opens the door for observing machinewashing from an economic self-interest perspective and instrumental rationale. Analogous to greenwashing research, one may ask whether machinewashing pays off for an organization (Berrone et al., 2017)? Is machinewashing used to change perceptions about organizational AI ethics performance? Does evidence exist that the market values machinewashing (Du, 2014)? Does greater transparency about AI ethics programs, such as the disclosure of algorithmic code, mitigate information asymmetries between organizations and their stakeholders?

Whereas the previous paragraphs outline machinewashing issues connected to organizations and the environments in which they are embedded, the study of machinewashing on a micro-level is likewise important (Lyon & Montgomery, 2015). Micro-level theories focus on the individual within and around organizations (Cornelissen & Durand, 2014). Micro-level theories are essential, as machinewashing practices can influence or be shaped by prominent individuals within an organization. In addition, as a phenomenon that lies in the beholder's eye, it is crucial to study how external stakeholders evaluate machinewashing. Thus, how individuals perceive various machinewashing practices. Previous greenwashing research builds on two prominent theories in this regard: agency and attribution theory.

Agency theory may provide a fruitful avenue for future research to analyze internal antecedents of machinewashing. Individual actors may play a crucial role in designing, implanting, and executing AI ethics programs. Managers may engage in machinewashing activities and thereby follow personal interests (as opposed to shareholder interests), creating costs and inefficiencies for the principle (the organization) and broader society (Bosse & Phillips, 2016; Hadani & Schuler, 2013; Petrenko et al., 2016). An interesting question in this regard is how do managers instrumentalize AI ethics or build on information asymmetries to pursue their own career goals? Mark Zuckerberg’s Facebook role has recently received critical attention (Abdalla & Abdalla, 2020; Zuboff, 2021). Thus, agency theory raises questions, such as: how does machinewashing impact agency cost? How can organizations control and verify that an agent acts in the principal's interest, not engaging in machinewashing practices? Further, it may be interesting to apply agency theory regarding employee’s relation to machinewashing practices. The introductory example of Timnit Gebru and Margaret Mitchell may be an interesting starting point to study how machinewashing practices may impact organizational identification, employee engagement, and performance (Johnson, 2021; Vincent, 2021). Thus, further questions raised by agency theory include: how do AI ethics programs relate to individual employees? What is the impact of machinewashing on employee performance, well-being, and satisfaction? How can individual members of the organization be included in creating AI ethics codes and guidelines? How can individuals (continuously) challenge and be challenged by the code of their respective organizations?

Another theory that may prove helpful for future research is attribution theory (Harvey et al., 2014; Lange & Washburn, 2012). Attribution theory may help to capture external evaluations of machinewashing on an individual basis. This may include studying individuals' perceptions of machinewashing practices and the conclusions they draw from them. Thus, important avenues for future research may consist of how observers make sense of machinewashing communication and whether this may relate to aspects such as purchase and investment intention or product and organizational loyalty (Ginder et al., 2019; Pizzetti et al., 2019). In addition, attributional processes may help explain individuals’ emotions and behavior when facing potential deceptive AI claims. This raises several questions: What micro-level attribution processes occur when consumers perceive AI ethics programs to be misleading? How do consumers and the wider public perceive deceptive practices, such as using intellectual property rights law to avoid disclosing algorithmic code or obscuring responsibility for unintended outcomes of semi-automated systems by blaming the human in the loop?

To further explore and theorize about machinewashing empirical inquiry is paramount, as current research remains largely conceptual (Benkler, 2019; Jobin et al., 2019; Wagner, 2018). As shown above, future research may utilize the theories from the three focus levels as a basis for empirical study. The suggested research questions (see Table 5) can be seen as initial ideas to apply these theories to different organizations, different organizational and institutional contexts, and investigate individual behavior. The complexity of machinewashing practices demands a holistic way of proceeding that draws on different methodological approaches, such as quantitative, qualitative, and combined methods. Further, to gain profound insights into machinewashing, it is necessary to collect data on multiple levels (individuals, organizations, and aggregated institutional contexts) and study processes and how they develop over time. The introductory case of Timnit Gebru exemplifies the involvement of multiple levels (from individual to aggregated institutions) and the need to study machinewashing practices from within and outside the organization and how it unfolds over time (Johnson, 2021; Vincent, 2021).

Future research may also benefit from further operationalizing the outlined machinewashing concept. This may include developing ways to measure stakeholder perceptions of machinewashing as a phenomenon that lies in the eye of the beholder (Lyon & Montgomery, 2015). In this regard, one possible pathway for future research is developing a scale that allows assessing perceptions of machinewashing. Scale development is a multistage process that helps create a validated scientific measurement instrument (DeVellis, 2017; Hinkin, 1995; Jian et al., 2014). In addition, scales can reveal insights into variables such as machinewashing that are not readily observable (DeVellis, 2017). Thus, a scale can benefit future research to generate and test hypotheses empirically, such as the perceived legitimacy of corporate AI soft policies.

With the widespread implementation of AI, organizations have started implementing AI ethics programs to meet societal and regulatory concerns and counter potential adverse outcomes of AI. However, critical voices increasingly question the reality behind corporate AI ethics, labeling it as machinewashing (Johnson, 2021; Ochigame, 2019). This article set out to offer conceptual clarity into the emerging machinewashing debate, discussing its resemblance and difference to greenwashing and providing a research program based on multiple theory lenses present in existing greenwashing literature. We have reasoned by analogy, invoking as structural mapping of greenwashing as the source and machinewashing as the target domain to allow for a knowledge transfer, thus, creating a theory-informed model of machinewashing (Ketokivi et al., 2017; Nersessian, 2008). Given the distinct idiosyncrasies of machinewashing, we have cautioned against a blind transfer of theoretical assumptions from greenwashing to machinewashing (Alvesson & Sandberg, 2013). Thus, we have emphasized the need for adaptations and critical reflection about greenwashing theories against distinct machinewashing characteristics. Further, we strongly encourage future research to draw on theoretical insights beyond the previously mentioned theories to allow for a rethinking and the generation of novel insights beyond the traditional “research box” or community (Alvesson & Sandberg, 2014). Overall, with our conceptual machinewashing model, the definition, and the theory-driven research program, we strive to trigger future research and cross-disciplinary dialog about machinewashing.