1. Introduction
Artificial intelligence (AI) systems are increasingly used to support and automate decision-making in organizations, including in public consumer banking. These systems are expected not only to increase efficiency and reduce costs but also to deliver fair, transparent and responsible outcomes, especially in domains where decisions significantly impact citizens’ everyday lives. As AI becomes more deeply embedded in organizational processes, it challenges traditional normative frameworks and raises urgent questions about how ethical values can be integrated into technical systems. One such value, benevolence, has received little attention in current debates, despite its fundamental role in human trust-building and moral reasoning.
Benevolence refers to the intention to do good, to promote the well-being of others and to act in a way that reflects care, empathy and fairness (Baier, Reference Baier1986; Mayer, Davis & Schoorman, Reference Mayer, Davis and Schoorman1995). In organizational contexts, benevolence has been studied as a key component of trust in leadership (Dirks & Ferrin, Reference Dirks and Ferrin2002), service delivery (Lankton, McKnight & Tripp, Reference Lankton, McKnight and Tripp2015) and stakeholder relations (Hosmer, Reference Hosmer1995). As decisions become increasingly data-driven and automated, the question arises whether and how such a social and relational value can be translated into machine-supported decision-making.
The concept of Artificial Benevolence seeks to address this gap by exploring how benevolence can be conceptualized, operationalized and governed within AI systems. Unlike related principles such as fairness or explainability, artificial benevolence emphasizes the intent behind decisions and an active orientation toward the welfare of others. It focuses not only on avoiding harm or bias but also on promoting positive outcomes and caring responses. Our approach goes beyond existing discussions by positioning benevolence as a distinct governance and design objective, one that requires technical, organizational and legal translation.
This issue is particularly critical in public consumer banking. Public banks carry both economic and social responsibilities. Their decisions, such as in credit scoring, access to services, or handling hardship, often affect housing, mobility, entrepreneurship and financial inclusion. As they increasingly deploy AI systems, an essential question arises: How can these technologies align with the public sector’s normative duty to act benevolently?
Moreover, public banks operate within a strict legal and regulatory framework. Deploying AI in this context raises complex legal questions ranging from data protection and anti-discrimination law to constitutional matters, such as the equality principle and the state’s duty to ensure basic welfare. Here, Artificial Benevolence becomes as much a legal-institutional challenge as a matter of design.
This article contributes to the growing debate at the intersection of AI ethics, organizational design and public governance by investigating Artificial Benevolence in the decision-making of public consumer banking. Combining insights from Information Systems and Law, we examine both the conceptual foundations of benevolence and the legal implications of embedding benevolence into AI-driven decision-making. Specifically, the paper makes three contributions: First, it conceptualizes artificial benevolence as a distinct normative goal in AI governance, differentiating it from related concepts. Second, it develops a socio-technical understanding of how benevolence can be operationalized through system design, decision logic and organizational routines. Third, it analyzes the legal challenges that arise when public banks seek to embed benevolence into algorithmic decision-making, with a particular focus on constitutional and regulatory obligations.
The paper is structured as follows. Section 2 lays the conceptual foundation by introducing benevolence as a normative and strategic concept and outlining its potential translation into AI systems. Section 3 discusses the specific challenges and opportunities of embedding artificial benevolence in AI-supported decision-making within public consumer banking. Section 4 analyzes the legal implications of such practices, focusing on constitutional, regulatory and institutional aspects. Finally, Section 5 offers conclusions and outlines avenues for future research and policy development.
2. Conceptual foundations
2.1. Benevolence as a normative, relational and strategic concept
Benevolence, commonly defined as the disposition to do good, to care for others and to act in their best interest, is a central value in moral philosophy, human relationships and organizational trust (Baier, Reference Baier1986; Mayer et al., Reference Mayer, Davis and Schoorman1995). It entails more than fairness or non-maleficence; it implies a proactive orientation toward the well-being of others.
In trust theory, benevolence is one of three key antecedents of perceived trustworthiness, alongside ability and integrity (Mayer et al., Reference Mayer, Davis and Schoorman1995). It signals that an actor is not merely competent and principled but also genuinely concerned with the interests of others. In organizational settings, benevolence has been studied in leadership (Dirks & Ferrin, Reference Dirks and Ferrin2002), service delivery (Lankton et al., Reference Lankton, McKnight and Tripp2015) and inter-organizational relationships (Zaheer et al., Reference Zaheer, McEvily and Perrone1998).
Recent work in organization theory expands this understanding by conceptualizing benevolence not only as a moral disposition but also as a strategic organizational orientation. Beveridge and Höllerer (Reference Beveridge and Höllerer2023), for instance, define organizational benevolence as a collective and sustained commitment by organizations to promote the welfare of external stakeholders. They show how such a disposition can be embedded in organizational identity, routines, and communication – often serving both ethical and reputational purposes. In this sense, benevolence becomes a governable and instrumental quality, not limited to individual actors or isolated acts of kindness.
From this perspective, benevolence can be seen as a dual-purpose practice: It reflects genuine concern for others while also generating potential benefits for the organization, such as customer loyalty, trust in digital systems and differentiation in increasingly automated service environments (Hurley, Gillespie, Ferrin & Dietz, Reference Hurley, Gillespie, Ferrin and Dietz2013). Particularly in public consumer banking, where legitimacy and citizen orientation are paramount, benevolent action, whether human or algorithmic, can help organizations align short-term responsiveness with long-term societal value.
This dual nature of benevolence, as both a relational norm and a strategic capability, is essential for understanding its potential translation into AI-supported systems. It suggests that Artificial Benevolence should not be understood only as an ethical imperative but also as an organizational choice that affects how systems are designed, governed and justified to the public.
2.2. From human to artificial benevolence
The question of whether benevolence can be “translated” into AI systems is both philosophical and practical. From a design perspective, Artificial Benevolence refers to the idea that AI systems can be designed to act not just efficiently or fairly, but with a deliberate orientation toward promoting human well-being. To address this, we must understand how benevolence functions as a human value and how it might be translated into system-level logic and behavior. Artificial benevolence does not imply that machines develop moral agency or emotional empathy. Rather, it refers to the capacity of AI systems to act in ways that are systematically oriented toward the well-being of others, as encoded through design, rules and institutional embedding.
In human interactions, benevolence is typically communicated through intention, empathy and moral reasoning. In organizations, these values are institutionalized via leadership, culture, routines and social norms. Translating such values into AI systems requires rethinking how intention and care can be embedded in logic-driven processes. While machines do not possess intentional states, designers and institutions can embed intentionality in the form of encoded objectives, weighted outcomes and procedural safeguards. This parallels approaches in value-sensitive design (Friedman & Hendry, Reference Friedman and Hendry2019), which aim to ensure that technologies reflect and respect human values from the outset.
However, value translation is non-trivial. Human benevolence is often situational, relational and emotionally attuned (Coeckelbergh, Reference Coeckelbergh2010; Floridi & Sanders, Reference Floridi and Sanders2004). AI systems, by contrast, rely on abstract representations and generalized patterns, lacking the contextual sensitivity and affective grounding that inform human moral reasoning (Moor, Reference Moor2006). This creates risks of oversimplification or misalignment, especially when systems operate autonomously in sensitive domains like public banking.
2.3. Defining artificial benevolence
We define artificial benevolence as the deliberate orientation of an AI system toward promoting the well-being of individuals and communities affected by its decisions. It does not require the system to have moral intentions but does require it to produce outcomes consistent with benevolent aims. Artificial benevolence thus constitutes a distinct design and governance goal: to configure systems in ways that consistently promote the good of others, especially in asymmetrical, high-stakes, or ethically sensitive contexts. While related to existing concepts such as trustworthy AI or AI ethics, artificial benevolence offers a distinct focus. Trustworthy AI generally refers to systems that are lawful, ethical and robust (European Commission, 2020), encompassing a wide range of principles such as fairness, transparency and accountability. In contrast, artificial benevolence is explicitly oriented toward the active promotion of human and societal well-being, emphasizing care and responsiveness in decision outcomes rather than procedural compliance alone. AI ethics provides overarching normative guidance but often lacks actionable pathways to implementation (Morley, Floridi, Kinsey & Elhalal, Reference Morley, Floridi, Kinsey and Elhalal2020). Artificial benevolence addresses this gap by translating ethical intentions into operational design choices, decision logics and institutional routines that are sensitive to social needs and contextual values (Chandrasiri, Bandara, Rosemann, Ostern & Voss, Reference Chandrasiri, Bandara, Rosemann, Ostern and Voss2025; Dignum, Reference Dignum2019).
Artificial benevolence can manifest across various interrelated layers of socio-technical systems. At the design level, benevolence must be embedded through explicit objectives, ethical constraints and participatory development processes that foreground user welfare (Friedman & Hendry, Reference Friedman and Hendry2019; Van Wynsberghe, Reference Van Wynsberghe2013). In the field of Information Systems, design science research emphasizes the importance of incorporating stakeholder values directly into artifact construction (Gregor & Hevner, Reference Gregor and Hevner2013). Building on this, Rosemann, Ostern, Voss and Bandara (Reference Rosemann, Ostern, Voss and Bandara2023a) argue that benevolence should not be understood as a one-time design feature but as a routinized and data-informed practice that evolves through interaction between technical artifacts and human actors.
Beyond design, the decision logic within AI systems can instantiate benevolence through prioritization schemes that emphasize fairness, harm reduction or social need. For instance, in public consumer banking, loan approval algorithms could be calibrated not only for creditworthiness but also for social vulnerability or community-level impacts (Binns, Reference Binns2018). Such configurations ensure that benevolent intent is reflected in algorithmic outcomes. This aligns with recent conceptualizations of mutualistic benevolence, which highlight how AI-supported processes can facilitate value co-creation between organizations and their customers, balancing short-term customer well-being with long-term institutional legitimacy (Rosemann, Bandara, Ostern & Voss, Reference Rosemann, Bandara, Ostern and Voss2023b).
However, embedding benevolence in the technical core is insufficient without broader organizational integration. Institutions must establish mechanisms that reinforce benevolence as a normative and operational commitment. This includes formal structures such as human-in-the-loop oversight, ethical auditing and procedural accountability (Dignum, Reference Dignum2019; Morley et al., Reference Morley, Floridi, Kinsey and Elhalal2020), as well as more subtle forms of governance that align internal values with external stakeholder expectations.
Moreover, benevolence in AI systems must be adaptive. Organizations must continuously assess whether AI-supported processes are producing intended benevolent effects or causing unintended harm. Techniques from continuous learning, human feedback loops and algorithmic impact assessments are essential to this goal (Raji, Smart & White et al., Reference Raji, Smart, White, Mitchell, Gebru, Hutchinson, Smith-Loud, Theron and Barnes2020; Selbst, Boyd, Friedler, Venkatasubramanian & Vertesi, Reference Selbst, Boyd, Friedler, Venkatasubramanian and Vertesi2019).
Taken together, these perspectives suggest that artificial benevolence is not a static or isolated feature but a dynamic, embedded and processual property of socio-technical systems. It must be cultivated across design, decision logic, institutional routines and continuous adaptation. Rather than relying solely on rule-based ethics or abstract principles, benevolent AI requires an organizational infrastructure that supports repeatable, legitimate and responsive practices.
3. AI-supported decision-making in public consumer banking
Public consumer banking institutions, such as publicly owned savings banks, operate at the intersection of market logic, public mandates and social responsibility. These institutions increasingly adopt AI systems to support or automate high-stakes decisions related to credit scoring, loan approvals, overdraft management and customer support (Veale, Van Kleek & Binns, Reference Veale, Van Kleek and Binns2018; Zarsky, Reference Zarsky2016). In doing so, they aim to improve efficiency, reduce bias and enhance service accessibility. However, AI deployment in this context raises unique normative and institutional challenges, particularly regarding the alignment of algorithmic processes with the institution’s intended mission.
Public banks are typically governed by legal and political obligations to uphold fairness, transparency and inclusion (Bozeman & Johnson, Reference Bozeman and Johnson2015). Decisions made by these institutions can significantly affect social mobility, housing access and financial inclusion, especially for vulnerable populations. Thus, embedding artificial benevolence into AI-supported decision-making is not merely an ethical ideal, but central to fulfilling the bank’s public mandate. This requires not only ethical system design but also the institutional integration of benevolent governance structures and accountability frameworks (Dignum, Reference Dignum2019; Lepri, Oliver, Letouzé, Pentland & Vinck, Reference Lepri, Oliver, Letouzé, Pentland and Vinck2018).
AI-supported systems can standardize benevolent practices, for instance, by prioritizing hardship-sensitive risk assessments or offering protective features for financially distressed customers (Veale et al., Reference Veale, Van Kleek and Binns2018). Yet, these systems also risk obscuring decision processes and limiting the role of human judgment, which is crucial in cases requiring empathy or contextual discretion (Citron & Pasquale, Reference Citron and Pasquale2014). Research suggests that public trust in AI depends not only on technical accuracy but also on perceived legitimacy and normative alignment (Morley et al., Reference Morley, Floridi, Kinsey and Elhalal2020).
To institutionalize artificial benevolence, public banks can foster participatory design mechanisms, embed human-in-the-loop processes for ethical oversight and implement auditing procedures to assess the ongoing social impact of AI systems (Morley et al., Reference Morley, Floridi, Kinsey and Elhalal2020; Raji et al., Reference Raji, Smart, White, Mitchell, Gebru, Hutchinson, Smith-Loud, Theron and Barnes2020). These structures could ensure that AI-driven decisions are not only efficient but also just, inclusive and responsive to the bank’s broader societal role. Moreover, unlike in purely commercial settings, these practices must be aligned with relevant legal mandates and statutory provisions. Questions of discrimination, the safeguarding of fundamental rights and further organizational constraints on algorithmic decision-making will be examined in the ensuing section.
4. Legal challenges
Given the sheer volume of regulatory requirements, the following analysis can merely outline the legal challenges associated with the topic of artificial benevolence in public consumer banking. In addition to a concise overview, the legal issues will be examined with regard to general constitutional principles and applicable elements of European secondary law.
Initially, the fundamental preliminaries surrounding the role of artificial benevolence in the area of public savings banks will be presented in Section 4.1. In addition to explaining the legal status of savings banks (Section 4.1.1), this will particularly cover their obligation to serve a public purpose (Section 4.1.2), while simultaneously being classified as market players (Section 4.1.3). Subsequently, the relevant concrete implications for the design of artificial benevolence will be illustrated in Section 4.2. This concerns the issues of discrimination (Section 4.2.1), administrative self-commitment (Section 4.2.2), trust (Section 4.2.3) and transparency (Section 4.1.4). Thereafter, the part closes with an explanation of the new regulatory framework resulting from the enactment of the EU AI Act (Section 4.3).
4.1. Legal preliminaries
4.1.1. Legal status of public savings banks
As already described, public consumer banks are progressively employing AI systems, just as their private competitors do. The concomitant transition of at least parts of the decision-making process from human staff to AI systems tests the existing banking regulatory framework in terms of the legitimacy and validity concerning the decisions made.
Predominantly publicly owned consumer banks are generally classified as an element of public administration in the domain of statehood, but have acquired a certain level of autonomy due to their status as independent public legal entities (Marois, Reference Marois2021). According to Weber, such independence is regularly embedded in order to facilitate citizens’ access to the optimal degree of efficient, bespoke administration. However, the self-determination rights of regional entities reserve a protected space for formal and material decisions that are autonomous from centralized control (Weber, Reference Weber2019).
The implementation of a benevolence orientation may therefore be autonomous within the scope of aforementioned institutional self-government. Conversely, any regulation governing implementation must assure this space thoroughly. Insofar as this requires a decision of relevance to fundamental rights concerning the relationship between the state and its citizens, it must be defined positively in accordance with the principles of primacy of and reservation by law (Huhn, Reference Huhn2016; Weber, Reference Weber2019).
4.1.2. Public purpose
Besides public ownership, a key feature of a public bank is its focus on a public purpose (Decker, Reference Decker2016; Omarova, Reference Omarova2024). It is this public purpose that justifies action by a state entity (Luhmann, Reference Luhmann1968). In line with a welfarism approach, the end to be attained through algorithmic benevolence is the maximization of equal prosperity (Adler, Reference Adler2000). As in the case of public consumer banks, securing adequate monetary liquidity for the majority of the population represents the enhancement of well-being which is in the public interest (Marois, Reference Marois2021; Omarova, Reference Omarova2024). The objective of public savings banks is insofar directed precisely at stimulating the economy by providing financial access to broad and financially disadvantaged sections of the population (Butzbach, Reference Butzbach, von Mettenheim and Del Tedesco Lins2008; Decker, Reference Decker2016). The system architecture of public savings banks is hence geared toward institutional benevolence from the outset. In this regard, artificial benevolence as a process-related concept is closely associated, but not always congruent with forementioned finale objectives and should not be equated with a position such as so-called AI for Social Good (Iazzolino & Stremlau, Reference Iazzolino and Stremlau2024).
Furthermore, according to Omarova, public banks generally tend to be more risk-averse than their private competitors, given that they do not need to maximize profits. Consequently, public banks also promote overall macroeconomic stability (Omarova, Reference Omarova2024). In this respect, they facilitate not only a social function but also a stabilizing function for the financial system (Mann, Reference Mann2023). Germanys municipal public savings banks, for example, have historically emerged as guarantors of economic growth and resilience for the German economy due to their role in providing services benefiting society as a whole (Omarova, Reference Omarova2024). To that extent, public banks also serve as benchmarks for their private competitors (Decker, Reference Decker2016).
Insofar as the legally relevant actions of an entity must be assessed normatively and are invariably so presupposed by human will, decisions made by an automatic system must, a fortiori, satisfy the legal requirements for human actions (Herbosch, Reference Herbosch2023). If the public purpose of public banking has been executed by human staff up to now, this purpose must also be implemented in forthcoming benevolent decision-making practices of artificial systems. The digital shift of a state-controlled process can simultaneously constitute an advantage and a disadvantage for the legal subject concerned (Smuha, Reference Smuha and Smuha2025). On the one hand, in terms of a fundamental rights perspective, the datafication of the subject, given the omnipresent relationship of subordination characterized by an asymmetry of power, threatens to dehumanise the subject. Yet on the other hand, the potentiation, harmonisation and thus formal regulation of the fulfilment of the purpose may promote its well-being (Acemoglu, Reference Acemoglu2021; Smuha, Reference Smuha and Smuha2025).
4.1.3. Conflict between economic efficiency and public purpose
A financially unstable customer base endangers the stable repayment of given loans, the so-called recovery rate (Hamdan et al., Reference Hamdan, Landscheidt, Fischer-Appelt, Goletzko, Jochum, Hamdan, Köster, Vogel, Weis and Zimmer2024). Illiquid and financially strained customers are therefore generally unattractive to banks, meaning that private competitors will not enter into a lending relationship with them due to the absence of a contractual obligation to do so. However, by disassociating from the axiom of profit maximization, public banks, unlike their private competitors, can also grant financially less attractive customers more favorable loan terms through cross-financing (Menand & Ricks, Reference Menand and Ricks2024; Omarova, Reference Omarova2024). Nonetheless, public savings banks continue to operate functionally as regular players in the free market, subject to its laws of demand and price formation (Meulen et al., Reference Meulen, Christiaans, Wilke and Wohlmann2024). This provokes a fundamental conflict of objectives between fulfilling their statutorily defined public purpose and the economic interest in reliable repayment and its implications for a stable economic system (Busch & Franceschi, Reference Busch and Franceschi2021).
The aforementioned conflict has intensified amid the backdrop of the Brussels concordat and the resulting abolition of guarantor liability, alongside increasing pressure from private competitors (Säcker, Ganske & Knauff, Reference Säcker, Ganske and Knauff2022). While public consumer banks are legally bound to offer a sufficient level of service even to unprofitable customers, private competitors can more easily secure particularly profitable customers by focusing on service and adjusting their pricing structure.
The integration of artificial benevolence must not threaten the stability function arising from the democratically legitimised structure of the public bank (Hamdan et al., Reference Hamdan, Landscheidt, Fischer-Appelt, Goletzko, Jochum, Hamdan, Köster, Vogel, Weis and Zimmer2024). This is also reflected in secondary Union law: With regard to Art. 18 para. 6 Directive (EU) 2023/2225 of 18 October 2023, the freedom to grant benevolence under the law is limited where the results of the creditworthiness assessment indicate that the consumer is unable to fulfil their contractual duty (Busch & Franceschi, Reference Busch and Franceschi2021). Furthermore, the operationality of a socially sustainable banking industry itself demands economic stability; and such a condition may not be attained by overemphasizing a single social sustainability objective (Ekkenga & Winner, Reference Ekkenga and Winner2024). The implementation of benevolence must therefore always be considered in the context of other economic, social and environmental objectives.
Provided that these challenges are overcome, a consistent distribution practice – expansion of financial capacity through the granting of credit certainly represents a distribution decision – offers great potential for a socially constituted state. Especially with regard to cost-effective robo-advisory services and knowledge management, AI holds the potential to enable state consumer banking to achieve an even more competitive balance between economic efficiency and welfare objectives (Molosiwa & Molosiwa, Reference Molosiwa and Molosiwa2025).
4.1.4. Public consumer banks and regulation
This raises the question of the extent to which savings banks can respond to new regulatory requirements at an institutional level. Considering its preeminent regulatory standards, the finance sector can be expected to have enhanced expertise in managing additional regulation with regard to the possible implementation of a benevolence orientation (Kilian & Ebel, Reference Kilian, Jäck and Ebel2025). Larger credit institutions are endowed with advantages concerning any regulatory efforts due to the greater scalability of their work (Mackenzie-Gray Scott & Abrusci, Reference Mackenzie-Gray Scott and Abrusci2023). The compliance advantages of more sizeable banks over smaller competitors linked to greater scale proved evident, for example, in relation to the Commission Implementing Regulation (EU) 2015/2388 of 17 December 2015 (PSD2) (Kilian & Ebel, Reference Kilian, Jäck and Ebel2025).
Furthermore, banking networks located in only one jurisdiction obtain a regulatory benefit over cross-border banking institutions owing to the lack of transjurisdictional regulatory disparity (Hacker, Reference Hacker2024). Domestic public consumer banking associations profit in this respect from intensified regulatory pressure to implement algorithmic benevolence relative to their less-sized competitors and those operating in a transnational business field. Given the compulsion to adapt at a later date due to overarching EU obligations, it is worthwhile to embrace a strategy of early national implementation of these legal standards in favor of a competitive domestic banking industry (Ekkenga & Winner, Reference Ekkenga and Winner2024).
4.2. Legal implications for the architecture of benevolent decision-making
Insofar as the previous section outlined the necessary prerequisites for implementing artificial benevolence in the savings bank sector, the following section will present the specific legal implications for institutional implementation of artificial benevolence in public retail banking.
4.2.1. Equality and discrimination
On a regular basis, discriminatory practices influence the assessment of whether a consumer bank enters into a contractual relationship with a potential customer (Molosiwa & Molosiwa, Reference Molosiwa and Molosiwa2025; Stahle, Reference Stahle2023). Especially in the context of private competitors, socially marginalized population groups are particularly vulnerable to exploitative business practices such as disproportionately high interest rates (Stahle, Reference Stahle2023). Naturally, the bank’s interest in minimizing default risk regularly conflicts with the individual interest in gaining a loan with the lowest possible interest rate. The question of whether or not a consumer construction loan is granted will commonly have a significant impact on the subjective lifestyle choices of the borrower through decades to come. However, it is also undeniable that not every occurrence of unequal treatment can be dismissed on the basis of benevolence. It is in the very nature of a company operating in a free market economy that it can only provide services within the limits of a certain level of profitability.
Benevolence intersects at least in part with algorithmic fairness, i.e., a normative equilibrium geared toward inclusion, equality and the absence of unjustified unequal treatment of the data subjects concerned (Corrêa, Garsia & Elbi, Reference Corrêa, Garsia and Elbi2025). If a group of characteristic bearers is over- or underrepresented in the output compared to an equally large group of non-characteristic bearers, then a relevant demographic disparity exists, impacting both benevolence and fairness. The technical causes of such bias lie in the adoption of unrepresentative data or its unequal inclusion in the data set (Grozdanovski & De Cooman, Reference Grozdanovski and De Cooman2025; Genovesi et al., Reference Genovesi, Mönig, Schmitz, Poretschkin, Akila, Kahdan, Kleiner, Krieger and Zimmermann2024).
From a legal-doctrinal perspective, a capriciously disparate decision-making practice infringes upon the general principle of equality and the prohibition of arbitrariness (Weber, Reference Weber2019). In this regard, secondary law prohibiting discrimination is also regularly applicable to public consumer banks (Hamdan et al., Reference Hamdan, Landscheidt, Fischer-Appelt, Goletzko, Jochum, Hamdan, Köster, Vogel, Weis and Zimmer2024).
As a consequence arises the question of whether legal risks can be mitigated through technical means. Genovesi et al. state that the data set could be revised according to a metric of conditional demographic parity. Such conditional demographic parity expresses the aim of excluding other factors from the survey that are merely correlated with the output but are neutral. For budgetary constraints, existing ISO or IEC standards based on a metric of conditional demographic parity should be used (Genovesi et al., Reference Genovesi, Mönig, Schmitz, Poretschkin, Akila, Kahdan, Kleiner, Krieger and Zimmermann2024).
4.2.2. Decision-making culture and legal self-commitment
Benevolence must also be assessed with regard to the development of a legally relevant institutional decision-making culture. For Acemoglu and Robinson, decision-making cultures develop horizontally in institutions. In conjunction with the institution and the economic processes involved, the configurations of decision-making culture form a social equilibrium. The configuration of that decision-making culture is by no means immutable, but can sometimes change rapidly (Acemoglu & Robinson, Reference Acemoglu and Robinson2024).
The underlying, more resilient decision-making culture foundation must be differentiated from the large number of context-dependent interrelationships, social norms and political justifications. The transmission of aforementioned decision-making culture through generations may also be accompanied by a tendency to stick to existing decision-making patterns (Acemoglu & Robinson, Reference Acemoglu and Robinson2024). In the public sector, so-called derivative participation entitlements can arise on the basis of a historically formed service architecture. Such a participation right is a manifestation of the status positivus of a welfare state (Weber, Reference Weber2019).
4.2.3. Design promoting trust and acceptance
Trust describes openness to an expected but uncertain course of events (Prifti et al., Reference Prifti, Krijger, Thuis, Stamhuis, Kuhlmann, de Gregorio, Fertmann, Ofterdinger and Sefkow2023). In this regard trust serves as a particularly relevant resource in consumer banking. For instance, Indian consumers highlighted trust in their bank as the most important factor for their customer satisfaction (Bharti, Prasad, Sudha & Kumari, Reference Bharti, Prasad, Sudha and Kumari2023). The element of trust also shapes the question of digitalizing the operations carried out by banks. For example, customers regularly opt human services rather than algorithmic ones in particularly sensitive, subjectively charged consume issues (Freisinger & Mendelsohn, Reference Freisinger, Mendelsohn, Kuhlmann, de Gregorio, Fertmann, Ofterdinger and Sefkow2023).
As explained above, benevolence is an integral prerequisite for trust in an AI system (Liu & Moore, Reference Liu and Moore2025). According to Acemoglu and Lensmann, a successful introduction of a system can only take place in line with socially developed acceptance of the system. If a system is initially used incorrectly, such misuse may lead to long-term resistance to its use (Acemoglu & Lensman, Reference Acemoglu and Lensman2023). Further factors affecting user acceptance include a perceived increase in subjective benefits and improved usability for consumers (Bharti et al., Reference Bharti, Prasad, Sudha and Kumari2023). The legal establishment of benevolence implementation must therefore ensure notable added value and ease of use for consumers and foster trust in the digital decision-making process.
In this regard, benevolence already holds a relevant function in the specific design of consumer banking advice. For example, banks must ensure that any robo-advisors used in the context of investment advice act in the best interests of the customer, Art. 64 para. 2 Commission Delegated Regulation (EU) 2017/565 of 25 April 2016. According to Art. 54 para. 12 Commission Delegated Regulation (EU) 2017/565 of 25 April 2016, the customer’s social circumstances must be reflected in the recommendation as an essential criterion (Ekkenga & Winner, Reference Ekkenga and Winner2024). Benevolence criteria are also integrated into the determination of the criteria for sustainable investment within the meaning of Recital 22 of Commission Delegated Regulation (EU) 2022/1288 of 6 April 2022.
4.2.4. Transparency
Appropriate legal oversight of public allocation decisions also demands a transparent control architecture (Martini, Reference Martini2008; Martini, Botta, Nink & Kolain, Reference Martini, Botta, Nink and Kolain2020). According to Harel and Pearl, transparency should be segmented into transparency in the narrower sense, the ability to participate, and, last but not least, contestability. The democratic legitimacy of a decision therefore must be established through the possibility of public discourse. However, only “intrinsically public decisions,” i.e. decisions regarding activities that can only be carried out by a public entity, would require legitimizing transparency (Harel & Perl, Reference Harel and Perl2024, p. 59).
The aforementioned claim is barely plausible. This is apparent from the original protective purpose of the transparency requirement. Kuhlmann states that within the subordinate relationship between the state and its citizens, a monopoly on information widens an existing power imbalance. When, in the context of private economic action, the focus lies on the subjective responsibility of the individual and their active position as a consumer, sufficient access to information is also essential (Prifti et al., Reference Prifti, Krijger, Thuis, Stamhuis, Kuhlmann, de Gregorio, Fertmann, Ofterdinger and Sefkow2023).
In view of the consumer protection orientation underlying the relevant consumer protection standards of financial services law, the information provided must also be presented in an effortlessly accessible manner (Lechevalier & Potel-Saville, Reference Lechevalier and Potel-Saville2025; Rennig, Reference Rennig, Kuhlmann, de Gregorio, Fertmann, Ofterdinger and Sefkow2023). Adaptive ML systems, which are already challenging to comprehend, are even more opaque for individual consumers, thereby hindering effective control of the decision outcome (Harel & Perl, Reference Harel and Perl2024). A non-royal distribution decision that can therefore be replaced by private bodies is thus just as subject to control under democratic rule of law.
The implementation of benevolence or the adoption of a data set that reflects a decision-making practice characterized by benevolence must therefore be executed sufficiently transparently for consumers. Existing information asymmetries can be reduced, for example, by displaying the intended applications (Ebner, Reference Ebner, Kuhlmann, de Gregorio, Fertmann, Ofterdinger and Sefkow2023). Furthermore, the risk of decision-making practices shifting to non-transparent informal areas must be prevented by the regulatory approach (Geminn, Reference Geminn2023; Wischmeyer, Reference Wischmeyer, Kuhlmann, de Gregorio, Fertmann, Ofterdinger and Sefkow2023).
Nonetheless, assuring the operationality of public services, of which state consumer banks are a part, is in the public interest, which can justify interference with informational self-determination arising from the creation of sectoral profiles (Botta, Reference Botta, Kuhlmann, de Gregorio, Fertmann, Ofterdinger and Sefkow2023). The required measures to promote transparency are determined by the severity of the interference in each individual case (Botta, Reference Botta, Kuhlmann, de Gregorio, Fertmann, Ofterdinger and Sefkow2023; Hornung, Reference Hornung2024; Raji, Reference Raji2022).
The significance of adequate transparency with regard to benevolent business practices becomes particularly clear when considering algorithmic creditworthiness assessments. In order to mitigate a deterioration in credit risk and thus enable the lender to operate profitably in the first place, the lender must be able to assess the borrower’s creditworthiness or make use of collateral (Busch & Franceschi, Reference Busch and Franceschi2021). The results of an creditworthiness assessment will regularly influence the terms and conditions of the actual loan in the sense of an algorithmic prejudice (Genovesi et al., Reference Genovesi, Mönig, Schmitz, Poretschkin, Akila, Kahdan, Kleiner, Krieger and Zimmermann2024).
Insofar, the key metrics for a creditworthiness assessment are to minimize the default rate and maximize the return (Harel & Perl, Reference Harel and Perl2024). The data processed by external credit agencies or the financial institution may for example include current repayment obligations, payment defaults or the number of current credit card agreements (Genovesi et al., Reference Genovesi, Mönig, Schmitz, Poretschkin, Akila, Kahdan, Kleiner, Krieger and Zimmermann2024). However, with regard to data minimization according to the meaning of Art. 5 para. 1 lit. c Regulation (EU) 2016/679 of 27 April 2016, the data entered must still have some direct connection to the intended purpose of assessing the potential lenders creditworthiness (Genovesi et al., Reference Genovesi, Mönig, Schmitz, Poretschkin, Akila, Kahdan, Kleiner, Krieger and Zimmermann2024).
A lender can usually only root its forecast in observable information. The prospect of an algorithmic credit assessment may therefore encourage consumers to change their consumption behavior. This would harbor the risk of obscuring genuine factors affecting creditworthiness. In this respect, a lender has no interest in disclosing the coefficients underlying the assessment; in particular, the intent is to counteract gaming, i.e. the deliberate influencing of digital decision-making practices (Langenbucher, Reference Langenbucher and Smuha2025).
On the reciprocal side, within the scope of complex neural networks, it will be difficult to prove indirect discrimination by the financial institution. Langenbucher therefore argues, that a use of algorithm-based creditworthiness assessments may, in particular, enable discriminatory decision-making practices to be masked, given the potential for correlations. However, for economic reasons alone, lenders have an intrinsic interest in avoiding false negativity, that is only a mere correlation without genuine causality (Langenbucher, Reference Langenbucher and Smuha2025). In the case of smaller consumer loans, a bias will be more recognizable as unlike with more extensive lending, usually no collateral is used (Genovesi et al., Reference Genovesi, Mönig, Schmitz, Poretschkin, Akila, Kahdan, Kleiner, Krieger and Zimmermann2024). If a consumer bank cannot guarantee whether the external credit agency satisfies its own requirements for substantive consideration of benevolence, it must therefore perform a sufficient human review of the findings.
4.3. Requirements of the EU AI act
For Smuha and Yeung, the requirements of the AI Act (Regulation (EU) 2024/1689 of 13 June 2024) on the implementation of a decision-making practice of benevolence correspond to the approach of regulated self-regulation that characterizes current EU law. With regard to the concrete implementation of the normative imperative, regulated self-regulation appears to be less effective overall than proactively designed full monitoring (Smuha & Yeung, Reference Smuha, Yeung and Smuha2025). In view of our interpretation of benevolence as a continuing commitment, the implementation of recurring monitoring obligations e.g. the measures under Art. 14 para. 3 must be positively highlighted. On another positive note, the emerging normative requirements of the AI Act are partly integrated into existing vertical regulatory networks (Kilian & Ebel, Reference Kilian, Jäck and Ebel2025).
At the same time, the exegesis of norms requires a sufficient empirical, normative and technical foundation against the background of the specific field of application (Metikos, Reference Metikos2023). The ambiguous wording and inadequate available options for verifying compliance with the regulation spark legal uncertainty. Risk aversion and the vagueness of the normative text generate additional personnel and material costs for the economic entities affected (Kilian & Ebel, Reference Kilian, Jäck and Ebel2025).
Benevolence is emphasized in the text, with Recital 1 referring directly to the requirement of human-centricity. Recital 2 explains the particular relevance of trustworthy AI concerning the protection of natural persons. Recital 22 refers directly to the social and environmental well-being of data subjects. Art. 5 para. lit. c focuses in particular on the potential harm to disadvantaged social groups in the category of prohibited systems. Furthermore, the risk management system required for high-risk AI systems takes particular account of threats to the legal interests of data subjects, Art. 9 para. 2 lit. a.
According to Langenbucher, algorithmic credit checks raise concerns about the emergence of an a-priori commitment to the recommendation. In the legislative process, the Commission of the European Union therefore identified creditworthiness assessments as a particularly precarious domain for the application of AI, leading to Art. 27 para. 1 s. 1 EU AI Act. However, atypical candidates may benefit from algorithm-based creditworthiness assessments compared to analogue assessments under certain circumstances due to the systems mere focus on correlation in data (Langenbucher, Reference Langenbucher and Smuha2025).
The relevance of benevolence criteria, particularly in connection with the financial sector, is particularly clear with regard to determining the necessity of a fundamental rights impact assessment under forementioned Art. 27. Yet, due to the current normative anchoring of the assessment, there is a potential risk that it will erode into a mere bureaucratic obstacle without any actual protective added value (Smuha & Yeung, Reference Smuha, Yeung and Smuha2025). A more promising approach would be to enshrine a model of multi-stage impact assessment that includes a mandatory quantitative questionnaire and a quantitative assessment matrix (Bertaina et al., Reference Bertaina, Biganzoli, Desiante, Fontanella, Inverardi, Penco and Cosentini2025).
5. Conclusion and outlook
This article has examined the concept of artificial benevolence and its potential role in the AI-supported decision-making processes of public consumer banks. Our interdisciplinary analysis highlights that embedding benevolence into algorithmic systems is not merely a design task but requires institutional commitment, regulatory guidance and a clear articulation of public purpose. Public banks, with their dual role of economic actor and public institution, represent a unique and particularly sensitive context for this endeavor. Particularly, we suggest incorporating the process-related understanding of artificial benevolence into the normative arrangement of current target positions, as is typical for public service providers, in a progressive manner by means of pluralisme ordonné (Delmas-Marty, Reference Delmas-Marty2006).
We have shown that benevolence can be operationalized as a socio-technical practice that spans the design, governance and oversight of AI systems. However, translating this value into practice is constrained by legal, organizational and technological factors. From a legal perspective, challenges arise in reconciling the principles of equity, equality and administrative transparency with the structural logic of automated decision-making. While the emerging European regulatory framework, especially the AI Act, offers some guidance, it remains ambiguous in its normative expectations and operational requirements.
Looking ahead, further research is needed to deepen our understanding of how benevolence can be balanced with economic efficiency, particularly in the face of structural pressures such as competition, resource constraints and digital transformation. We also see potential in exploring how institutional routines and decision-making cultures can evolve to support benevolent practices over time. Finally, the development of impact assessment tools that capture not only risks but also the potential for social good will be essential to move beyond compliance toward meaningful, value-based innovation in public sector AI.
Acknowledgements
The authors gratefully acknowledge the support of the Flow Factory, a research and innovation lab jointly established by the University of Münster and the Sparkassen-Finanzgruppe, which provided an inspiring environment for the development of this work.
Funding statement
The authors declare none.
Competing interests
The authors declare none.
Dr. Marleen Voß is an Assistant Professor at the University of Münster and a member of the European Research Center for Information Systems (ERCIS). Her research focuses on artificial intelligence, digital transformation and business process management, with particular emphasis on trust, benevolence and human–AI interaction. She has published in leading Information Systems outlets and studies how organizations adopt generative AI in knowledge management and decision-making.
Marleen is Managing Director of the Flow Factory, a research and innovation lab that advances AI-driven transformation in the financial sector through cross-sector collaboration. She has an international research profile shaped by her work with the Centre for Future Enterprise at QUT.
Laurenz Döring is a Research Associate and Doctoral Candidate at the University of Münster, affiliated with the Freiherr-vom-Stein-Institute. His research focuses on public law, particularly German and European constitutional, administrative and regulatory law, as well as interdisciplinary intersections with technology and society.
His dissertation examines the legal constraints of algorithmic systems in public consumer banking. He analyzes relevant secondary law and the constitutional implications concerning public purpose, fundamental rights and institutional supervision, grounding the doctrinal analysis in information technology and legal sociology.
Laurenz studied law at Ludwig-Maximilian-University Munich. He conducted a research visit at the Centre for Socio-Legal Studies, University of Oxford, published in public law and received several distinctions, including winning the 2024 INUR Essay Competition and the Best Memorandum Award at the SELS Copenhagen Moot Court. He has also presented constitutional law and legal history at international conferences at the University of Vienna and Georgetown University.
Open access
