Many ML algorithms have been accused of suffering from some degree of epistemological “opacity.” The accusation concerns the lack of transparency of, and the ensuing difficulty to grasp, the logic behind specific algorithmic outputs. Opacity may take various forms and come in different degrees. We focus on three: some ML algorithms may be opaque for technical or legal reasons, or because of users’ illiteracy.
The form of opacity affecting ML algorithms such as ANNs is often indicated by the “black box” metaphor. As Pasquale (
2015, p. 3) describes it, the inputs and outputs of the algorithmic system are accessible, but it is hard to inspect its inner workings to know with certitude how one becomes the other. For Burrell (
2016, p. 5), the interaction between the high dimensionality of data and the technical complexity of algorithmic codes generates opacity. As discussed later, in the domain of public institutions, the ensuing partiality of knowledge may be problematic for the attribution of human responsibilities for algorithmic outputs and decision-making. There are also domains where technical opacity is not necessarily problematic. For instance,
Alpha Go ZERO, a ML application designed to play Go, makes extensive use of neural networks and deep learning, training itself through self-play in order to identify the best moves and their winning percentages (Silver et al.,
2017). However, no particular explanation of the decision-making process or justification of the selected moves seem necessary in the context of an activity whose purpose is merely winning a game (Pégny & Ibnouhsein,
2018, p. 13).
The most frequent sources of legal opacity are Intellectual Property (IP) instruments such as patents, copyrights, trademarks, and trade secrets (Meyers,
2019). For instance, Citron and Pascal (
2014) have studied automated scoring systems for credit adjudication. They describe how IP instruments provide the legal basis to routinely deny requests for information about individual decisions. IP instruments have thus become a way to avoid audits of the underlying predictive algorithms by any actor outside the scoring entity (Citron & Pascal,
2014, p. 10). In addition, proprietary laws often cover the corpus of data used in ML algorithms training. Such laws are a source of opacity insofar as they impede a close inspection of the accuracy of the data backing decisional processes. For Citron and Pascal (
2014, p. 28), while confidentiality of proprietary algorithms with respect to the public at large could be maintained, emerging legal frameworks should facilitate the disclosure of source codes to trusted neutral experts. They should also allow for audit trails to assess the inferences and correlations of specific processes. Some commentators have also noted how IP clauses have gradually changed their traditional function from protecting against a competitor’s misappropriation to obstructing accountability (Katyal,
2019, p. 1246).
Illiteracy opacity indicates a generalized lack of technical knowledge to understand how AI decision-making systems work. This form of opacity is only partly due to the inherent technical complexity of algorithmic decision-making (Burrell,
2016; Danaher,
2016). It is by and large a circumstantial condition. Some recent surveys show an adult population “remarkably ill-informed about AI” (DeCario & Etzioni,
2021), including university students with no computer science background (Sulmont et al.,
2019). Lay persons struggle to distinguish what AI applications can do in real file as against fiction, and tend to conflate AI with robotics (e.g., in such movies as
Wall-e or
Her, which depict robots endowed with human intelligence; see Long & Margeko,
2020). Common misconceptions include conflating human thinking with computer processing; underestimating the role of humans in AI systems design and implementation; or overestimating the power of AI to solve societal problems (Sulmont et al.,
2019).
The low levels of AI literacy are not beyond remedy, of course. Moreover, in the context of public institutional action, there is a limit to what technical skills ordinary public officials need in order to appropriately interact with AI tools and integrate them into the exercise of their functions. Surely, public officials may not be expected to write and read codes. However, they also surely need higher order competencies such as understanding basic AI concepts and techniques (Long & Margeko,
2020). They should also be capable of critical evaluation, by identifying ethical concerns, as well as AI capabilities and limits in order to evaluate when it is appropriate to resort to AI, to what purpose, and in what measure. Such competencies are largely underdeveloped in the present context where AI has not yet been globally incorporated into schools’ curricula or officeholders’ training programs, and the main sources of information is popular media. While the number of countries offering public officeholders AI training programs is increasing,
8 the use of AI tools in the public sector is already happening, and the knowledge gap within public institutions is quite concrete.
Finally, cultural misconceptions about unrealistic levels of objectivity and accuracy of AI tools are often reinforced by communication strategies of tech corporations (Ajunwa,
2020, p. 1688; Boyd & Crawford,
2012). This phenomenon results in instances of “automation bias,” which challenge not only lay citizens but also institutional structures. For example, Katyal (
2019) and Wexler (
2018) warn about the lack of scrutiny of courts regarding algorithmic processes in the context of criminal justice. Burrell (
2016) and Danaher (
2016) recall the urgency of granting institutional access at all levels to competent independent experts who can advise on the performance of AI tools, as well as the development of public educational efforts for AI literacy.
Algorithmic opacity and office accountability
The three forms of opacity variably characterizing ML algorithms may raise serious normative challenges for the integration of this kind of automation into anticorruption within the framework of a public institutional ethics of office accountability. The various degrees of algorithmic opacity implicated in different AI tools used in different institutional processes challenge the accountability of public institutional action. As seen in the “
Anticorruption within an institutional ethics of office accountability” section, an institutional ethics of office accountability is the normative premise of a working public institution. Such a public institutional ethics demands the officeholders’ engagement in a constant and vigilant communicative effort to justify to each other the rationale of their use of their power of office. Such an internal mobilization, as seen, integrates legalistic approaches to anticorruption insofar as it can engage officeholders to sustain public institutional action, and react to such institutional dysfunctions as corruption over time.
Within the framework of a public institutional ethics of office accountability, a necessary condition for anticorruption is that the various components of officeholders’ deliberations about their uses of their power of office are, as much as possible, accessible and intelligible to them.
9 Insofar as, more or less significant portions of these deliberations depend on automated information processes that present some degree of at least one (possibly more) of the forms of opacity we have identified, that necessary condition for anticorruption as a matter of office accountability is variably but importantly challenged.
The history of using ML algorithms for preventing welfare frauds provides telling illustrations of how algorithmic opacity—in its diverse degrees and forms—may challenge office accountability. Such challenges emerge in their most destructive form when coupled with algorithmic malfunctions. One such malfunctions has dramatically manifested itself when the Michigan Unemployment Agency hired three private companies to develop an automated AI system, MiDAS, for detecting and adjudicating alleged frauds in unemployment benefits. Before implementing the system, public officers used to conduct interviews with claimants. The interviews aimed to explain questions and dispel doubts; officers enjoyed a large margin of discretion in asking questions and determining the presence of some fraud and how it should be remedied (e.g., by returning undue benefits, see Elyounes,
2021).
Besides reducing the costs deriving from the adjudication of benefits, MiDAS operationalizes one of the tenets of the regulatory approach to anticorruption (see the “
ML algorithms and the limits of an exclusively legalistic approach to anticorruption” section). MiDASI may be seen as a tool to enhance the institutional process, by reducing the space for human error and the distortions due to officers’ (ab)use of their discretion. However, between 2013 and 2015, the MiDAS algorithm incorrectly flagged over 34,000 people for fraud, causing massive loss of benefits, bankruptcies, homelessness, and even suicide (Hao,
2020). This situation resulted in a spike of court appeals and two class-action lawsuits. Investigations revealed that around 90% of the MiDAS fraud determinations were inaccurate (Calo & Citron,
2021, p. 828). The technical sources of such inaccuracies were as diverse as the misreading of scanned information; corrupt or inaccurate data mining, and errors in the “income spreading formula (Calo & Citron,
2021, p. 828). MiDAS also mistakenly flagged as fraud any discrepancies between the information provided by claimants and other federal, state, and employer records, with no margin of appreciation for unintentional errors (Felton,
2015; Wykstra,
2020). Of course, some mistakes could have occurred also by using other tools; and human errors in this department are frequent too. But the technical opacity of MiDAS made it difficult for the officers in charge to identify the mistakes and correct them. Ultimately, by progressively losing control of the institutional process, the officers’ capacity to account for its malfunctions diminished too.
What is more, a number of plaintiffs proved that the agency never properly notified them of the fraud allegations in a way that could give them a reasonable chance to defend themselves (Egan & Roberts,
2021; See also 2015 official memo from Shaefer and Gray to the U.S. Department of Labor). This problem seems to apply to many automated welfare-fraud detection systems. Notifications tend to reach citizens only at the end of the process; “they don't usually give them any information about how to actually understand what happened, why a decision was made, what the evidence is against them, what the rationale was, what the criteria were, and how to fix things if they’re wrong” (Wykstra,
2020, p. 9). The opacity of the notification process is in stark contrast with the tenets of office accountability which, on the contrary, were reflected in the interview-based system originally deployed.
A measure of legal opacity made the situation worse. Interviews with leading attorneys of class-action lawsuits describe how defendants “fought hard” in court—sometimes shielding behind intellectual property reasons—to avoid sharing key information about MiDAS’s innerworkings (Elyounes,
2021, pp. 495, 492). The hearings before administrative law judges often showcased some significant illiteracy opacity too, for example, facing agency staff unable to provide evidence to support MiDAS’s fraud accusations, including the misrepresentation of some claimants’ earnings (De la Garza,
2020, p. 4; see also the official memo from Shaefer and Gray to the U.S. Department of Labor,
2015, p. 3). The situation deteriorated to the point of having the state legislature pass a law requiring the Michigan Unemployment Agency to return to manual fraud determinations (De la Garza,
2020, p. 3).
10
The discussion of MiDAS’s shortcomings suggests that the integration of this type of ML technology into anticorruption must proceed with some precautions. Consider
Prozorro and the Brazilian application predicting the probabilities of a civil servant being corrupt (see the “
The integration of ML algorithms into anticorruption strategies” section). These are both “proprietary” applications; legal opacity may thus cover their underlying ML techniques. In the case of
Prozorro, corruption risk indicators are not constant, but they evolve through exposure to new data inputs and correlations (Petheram et al.,
2019, p. 18). Although this enhanced “plasticity” may discourage corrupt tenders from tricking the system, it adds to the technical opacity of this application as it increases the difficulties of interpreting and explaining the logic behind the process. And, of course, the challenges of illiteracy opacity are intuitively relevant in the face of the complexity of public institutional action (especially in cases of systemic corruption) and generalized lack of technical training for public officeholders.
To mitigate the opacity-related difficulties of integrating ML algorithms into the diagnosis of public institutional action, XAI techniques offer potentially appealing props for enhancing the accountability of institutional processes. For example, Loi et al. (
2021, p. 262) have recently proposed that proof be given that the same algorithm be consistently applied throughout decisions (“consistency transparency”), thus enhancing the justification of decisions via XAI models based on “design publicity” (see, also, Kroll et al.,
2017). More generally, post-hoc explanatory techniques can help to understand the “why” of algorithmic outputs, thus making institutional processes drawing on ML algorithms better intelligible, inter alia, to the officeholders relying on them for the exercise of their institutional functions. However, we must also acknowledge that some of the problems encountered with MiDAS would most likely still be present. XAI explanatory techniques would remain unable to detect algorithmic miscalculations or incorrect data input. Moreover, even very accurate XAI ML applications could lack normative informativeness, for example, by implementing an unfair treatment of citizens. Because algorithmic fairness is a disputed notion (Tsamados et al.,
2022), public officeholders’ role in assessing and answering for the compatibility of their institutional ethical standards with the fairness metric encoded in ML systems seems irreplaceable.
Additionally, court proceedings following automated fraud detection have shown some hesitancy of the human overseers to question AI-based processes (Charette,
2018). Such a hesitancy, partly due to low levels of AI literacy among officeholders, should be factored in by public institutions adopting explainable AI tools to mitigate opacity. XAI techniques may indeed exacerbate cognitive bias amongst users, including the “illusion of explanatory depth” triggered by an automated explanation in conditions of alleged process transparency (Bertrand et al.,
2022).
One last remark derives from some recent studies observing how AI programmers tend to design explanatory agents in view of their own competences without necessarily weighing them for the application’s intended users (Miller,
2018, p. 4). This condition makes XAI tools unable to offer interactive explanations tailored to the cognitive heterogeneity of users and their goals (Mittelstadt et al.,
2019; Murdoch et al.,
2019; Watson,
2022). Therefore, the officeholders’ direct engagement within the framework of a public institutional ethics of office accountability seems necessary to offer normative guidance for assessing the conditions under which automated explanations are compatible with answerability practices in the context of anticorruption efforts.
Surely, the challenges we have just presented are particularly menacing for office accountability were ML applications to replace officeholders in resisting corruption. But public officeholders may more parsimoniously use ML applications to enhance their capacity to detect accountability deficits. For example, ML algorithms may flag those officeholders who, by the nature of their role, are most in need of their colleagues’ support to ensure that their use of their power of office does not undermine institutional action. It remains nevertheless crucial that it is the officeholders’ primary responsibility to investigate such deficits.
How ML algorithms may weaken officeholders’ engagement
Throughout our discussion, we have emphasized the integration of anticorruption within a generalized effort to sustain public institutional action through an institutional ethics of office accountability. Alongside legal efforts to curb corruption through regulative and retributive measures often requiring the intervention of an external authority (see the “
ML algorithms and the limits of an exclusively legalistic approach to anticorruption” section), this ethical dimension of anticorruption works primarily from within an institution. It is premised on the engagement of public officeholders into a direct and continuous effort to check, sustain, and correct each other’s work in the context of their interrelated institutional action. This engagement is a call for public officeholders to take on responsibility for the working of their institution and answering to each other for the dysfunctions that may nevertheless occur.
The call to action for public officeholders’ engagement in anticorruption strategies can be unpacked into two elements. First, there is a commitment to enhancing the officeholders’ mobilization and their capacity to ask and offer each other an account of their conduct. Second, this essential (although not exclusive) driving force of anticorruption comes from within a public institution and may not be (exclusively) outsourced. The analysis of the prospects of integrating ML algorithms into the fight against corruption should consider the contribution that this technology can make as concerns these elements too.
Regarding the first element, we see a risk of corrosion of the officeholders’ engagement from decisional processes because of a growing resort to ML algorithms in public institutional action. As the automatization of critical decisions increases, officeholders are decreasingly stimulated or, in fact, capable to be vigilant, both with respect to their own conduct and to that of the other officeholders. This tendency is certainly proportional to the extent and the kind of reliance of institutional processes on ML techniques. Also, it is incrementally substantial in keeping with the level of automatization and the role of AI within it. In the case of ML applications, the tendency is also an indirect consequence of algorithmic opacity: The less officeholders are authorized to investigate (legal opacity), implicated in (technical opacity), and capable of understanding (illiteracy opacity) the grounds of the decision-making processes within their own institution, the less they can call each other to account for their conduct and answer for the failure of their institutional action.
The process of officeholders’ disengagement relates to a number of background factors. These include, for example, the successful branding campaigns of tech corporations promising the accuracy and efficiency of their AI-driven applications. But think also of campaigns of development organizations fostering the modernization of bureaucratic structures through the automation of public procedures. These factors risk leading to an overestimation of ML technological innovations’ positive contribution to public institutional action. Some social psychology studies suggest that such an overestimation risks making officeholders susceptible of unwarrantedly deferring to algorithmic advice, even in the face of contradictory information from other sources (automation bias, see Alon-Barkat & Busuioc,
2020, p. 24). These institutional and psychological factors contribute to sustaining a climate potentially inimical to a critical engagement not only with the practical limitations of what such innovations may or may not achieve, but also, and most importantly for us, with the ethical challenges that such innovations may raise.
11
The increasing resort to ML risks conveying a certain presumption of distrust towards public officeholders, whose work seems in need of some “extra-human check.” The propagation of such a sentiment may weaken officeholders’ institutional commitment and, therefore, pave the way for further corruption. Resorting to ML instruments may even serve as a “moral buffer,” motivating officeholders to relinquish their responsibility to the algorithms altogether (Busuioc,
2020, p. 832 ). An illustration comes from the alarm triggered by some civil lawyers in the United States about witnesses representing the state during court hearings, who are uncapable of justifying allegedly flawed decisions based on AI systems. Hao (
2020) mentions such officeholders’ responses as: “well, the computer did it—it’s not me,” while civil lawyers struggle to find effective litigation strategies and wonder with frustration: “Oh, am I going to cross-examine an algorithm?” This tendency pushes in a direction contrary to office accountability insofar as it muddles with the officeholders’ claims to the authorship of institutional action. This tendency is particularly problematic from an anticorruption point of view to the extent that, as argued in “
Anticorruption within an institutional ethics of office accountability” section, the fight against corruption requires engaging officeholders to answer for institutional action and dysfunctions.
The question of the authorship of decisions informed by the outputs of ML algorithms is a thorny one. Hundreds of individuals can be behind an algorithmic output, very often representing private institutions and their interests. There is the leadership of the tech corporation that produced the ML application; the entity commercializing the large digital datasets; the professionals assessing the quality of the data and preparing it for training the algorithm; the programmers who participate in the design of the algorithm; and, of course, the ML algorithm itself, which once developed has the potential to “learn” from new data inputs and change its decision-making rules. These manifold contributions are not only technical. They involve value choices such as deciding over tradeoffs between predictive accuracy and the interpretability of outcomes; decisions about inductive risk management
12; and the choice of fairness metrics. Technical and value laden choices are embedded into the technology effectively institutionalizing them (Crawford,
2016; Gillespie,
2014), while the outcomes produced have the potential to affect the life prospects of millions of people. Pondering these elements in deliberation has been the traditional domain of officeholders’ discretion, and a pre-condition of their taking responsibility for institutional action and dysfunctions. Insofar as the integration of ML algorithms interferes with this logic, it must be accompanied with extra caution and an awareness of the institutional transformations that may ensue.
The plurality of actors implicated in the design, development, and implementation of ML applications is also significant as concerns the second element that qualifies the integration of anticorruption within a public institutional ethics of office accountability. The compresence of many external actors and their implication in giving direction to public institutional action risks limiting the margins for officeholders’ initiative to take corrective actions from the inside a public institution. A concrete form such a risk may take is that the integration of ML algorithms becomes the Trojan horse of the privatization of anticorruption efforts. Insofar as private actors hold the monopoly over automated technologies, the integration of such technologies into the working of public institutions through, inter alia, anticorruption applications may reinforce the dependence of public institutional action on third parties, thus further disengaging officeholders. Differently put, this monopoly of private actors over automated technologies opens the doors for computer programmers and CEOs of tech corporations to become de facto public institutional agents, thus triggering the privatization of many sectors of the state’s action, including anticorruption.
13
The risk of privatizing public institutional action is indirect and contingent on the design and use of ML algorithms in the public sector. The status quo may be changed, and there are some examples of ML applications conceived in a way that reduces this kind of risk. This is the case of the
Dozorro application we encountered in the “
The integration of ML algorithms into anticorruption strategies” section. The design process of
Dozorro included a very active participation from government agencies and civil society organizations. This case suggests that the justification for integrating AI within anticorruption in public institutions is possible but conditional. It is conditional upon finding ways to pursue this innovation that do not have the effect of discharging public officeholders from their direct commitment to countering corruption by critically engaging with each other’s uses of their power of office. We hope that the analysis in this section has offered an initial qualified contribution to this quest.