Responsibility Gaps, LLMs & Organisations: Many Agents, Many Levels, and Many Interactions

Abstract

In this article, we propose a business ethics-inspired approach to address the distribution dimension of responsibility gaps introduced by general-purpose AI models, particularly large language models (LLMs). We argue that the pervasive deployment of LLMs exacerbates the long-standing problem of “many hands” in business ethics, which concerns the challenge of allocating moral responsibility for collective outcomes. In response to this issue, we introduce the “many-agents-many-levels-many-interactions” approach, labelled M3, which addresses responsibility gaps in LLM deployment by considering the complex web of interactions among diverse types of agents operating across multiple levels of action. The M3 approach demonstrates that responsibility distribution is not merely a function of agents’ roles or causal proximity, but primarily of the range and depth of their interactions. Contrary to reductionist views that suggest such complexity inevitably diffuses responsibility to the point of its disappearance, we argue that these interactions provide normative grounds for safeguarding the attribution of responsibility to agents. Central to the M3 approach is identifying agents who serve as nodes of interaction and therefore emerge as key loci of responsibility due to their capacity to influence others across different levels. We position LLM-developing organisations as an example of such agents. As nodes of interactions, LLM-developing organisations exert substantial influence over other agents and should be attributed broader responsibility for harmful outcomes of LLMs. The M3 approach thus offers a normative and practical tool for bridging potential gaps in the distribution of responsibility for LLM deployment.

Introduction

The growing deployment of highly autonomous AI systems in domains central to human life—such as medical diagnoses, education, elder care, and transportation—carries profound societal and political implications (Westerlund, 2020) and underscores the urgent need for concerted efforts in interdisciplinary research and practice to mitigate both current and potential risks associated with AI implementation (Russell et al., 2015; Tóth et al., 2022). Some of the most pressing risks range from digital harms—such as copyright infringement, political propaganda, and algorithmic bias—to physical harms, including injuries and fatalities caused by automated vehicles. Recent advancements in general-purpose AI models—such as large-scale language models (LLMs) including ChatGPT, Llama, and Gemini—offer significant potential for widespread real-world applications that impact numerous facets of human life (Bommasani et al., 2021). However, these AI models also bring new and more diffuse risks of discrimination, misinformation, deception, and environmental as well as socioeconomic harms (Hagendorff, 2024; Weidinger et al., 2022). In response, scholars have called for more interdisciplinary research to examine the factors that influence the locus of moral responsibility for AI deployment, with particular attention to the role of organisations (Lenk, 2017; Tóth et al., 2022).

However, the proper ascription of moral responsibility for AI deployment presents significant challenges. Scholars point to the emergence of “responsibility gaps” that the use of highly autonomous machine learning systems seems to create (Baum et al., 2022; Matthias, 2004; Nyholm, 2018), making it difficult to attribute blame to either the humans deploying such systems or to the systems themselves. Even when suitable candidates for the attribution of responsibility are identified, the problem of how to distribute this responsibility remains (De Jong, 2020), given the inherently collective nature of AI deployment. Generative AI technologies, such as LLMs, add new complications to the attribution and distribution of moral responsibility, such as blame–credit asymmetry in the evaluation of outcomes (Porsdam Mann et al., 2023). Although scholars have recognized the urgent need to develop robust methods for accurately attributing and distributing moral responsibility for harms resulting from AI deployment (Coeckelbergh, 2020), the collective dimension of such responsibility ascriptions remains insufficiently addressed, with more questions than answers.

In this article, we propose a business ethics-inspired approach to address the distribution dimension of responsibility gaps arising from the deployment of LLMs. We argue that the pervasive development of LLMs exacerbates the problem of “many hands,” previously examined in business ethics research in relation to the distribution of moral responsibility for collective outcomes. To address this problem, we propose a “many-agents-many-levels-many-interactions” (M3) approach that connects responsibility distribution gaps in LLM deployment with the prior treatment of the many hands problem in business ethics research. Our approach highlights the need to consider at least three collective dimensions when distributing moral responsibility in the context of LLM deployment. First, the multiple agents involved—the “many hands”—complicating the identification of those who bear moral responsibility. Second, the multiple levels of moral responsibility associated with these agents, such as micro, meso, and macro levels, as well as the connections between them. Third, the multiple interactions among loosely connected stakeholders, including individual and corporate users, developers, industry associations, governmental bodies, and public interest organizations, leading to a deeper entanglement of the agents contributing to LLM outcomes. Accordingly, addressing responsibility gaps in LLM deployment requires more than identifying all relevant agents at all the levels involved; it also necessitates an understanding of the many interactions taking place among agents and across levels. Moral responsibility for LLM deployment must therefore be conceptualized in terms of the interrelated dynamics of many agents, many levels, and many interactions.

The M3 approach proposed in this article highlights the existence of an intricate web of responsibility ascriptions surrounding the moral decisions, actions, and outcomes associated with LLM deployment and emphasizes that within this web, the distribution of responsibility is influenced by the extent of interactions among agents operating at various levels. We argue that these interactions do not inevitably result in a diffusion of responsibility to the point where we face unbridgeable gaps in the distribution of responsibility. Instead, we contend that agents engaged in a broader array of interactions tend to accumulate greater power and influence over others and should, accordingly, bear greater moral responsibility for harmful outcomes of LLM deployment. Our proposal thus offers a non-distributive solution to responsibility distribution gaps generated by LLM deployment: instead of distributing responsibility so widely that it becomes diluted or meaningless, we propose focusing responsibility ascriptions on those agents who possess greater capacity to influence the actions and decision of others, given the extensive range of interactions they are involved in.

The main result of the M3 approach is that the distribution dimension of responsibility gaps can be addressed by focusing on agents who serve as nodes of interactions and who emerge as key loci of responsibility based on their power to influence other agents operating at various levels. We identify LLM-developing organisations as a prime example of such agents, given their central position across all three dimensions of the M3 approach. As nodes of interactions, these organisations, such as OpenAI, Meta, and Google, are often in the position to exert the greatest power and influence over other agents and should therefore be attributed broader responsibility for harmful outcomes of LLM deployment. There exist sufficient normative grounds to close remaining gaps in the distribution of responsibility for LLM outcomes by ascribing residual responsibility to these LLM-developing organisations. While our findings specifically concern LLMs as a distinct category of generative AI, they are also applicable to other forms of highly autonomous AI systems, provided these systems share key characteristics such as broad applicability (i.e., general-purpose AI), ease of access (i.e., freely available and open to the general public, rather than restricted to expert communities), and deep societal effects through shaping of human practices on a global scale (e.g., LLMs have profound effects on education, which literally impacts worldwide population).

Responsibility Gaps, Many Hands, and LLMs

There seem to be multiple ways of understanding responsibility gaps opened by emerging technologies such as machine learning systems—referred to by Tigard (2021) as the “techno-responsibility gap.”1 In the initial conceptualisation, Matthias (2004) noted that the rise in autonomy of AI systems driven by machine learning technology creates a gap in responsibility attribution. There is no one to blame for the outcomes of “learning automata” because neither the human programmers nor the AI systems have the necessary control to be held responsible for them. More recently, Santoni de Sio and Mecacci (2021) identified four distinct responsibility gaps: culpability, moral accountability, public accountability, and active (forward-looking) responsibility. Additional gaps have been proposed by other scholars, including gaps in retribution (Danaher, 2016) and vulnerability (Vallor & Vierkant, 2024). Many researchers argue that the actions of learning automata can be connected to the humans involved in AI deployment, thus bridging responsibility gaps (Baum et al., 2022; Kiener, 2022).

However, identifying individuals as adequate targets of responsibility attribution might not resolve the problem: “even when the key human players involved have been identified, responsibility-gaps and retribution-gaps are not yet plugged” (De Jong, 2020, p. 734). Beyond the question of “who, if anyone, is to blame for harm caused by an artificial agent” (Hindriks & Veluwenkamp, 2023: 3), an important part of responsibility gaps in AI deployment seems to be apportioning responsibility for the outcomes generated by the AI system (Nyholm, 2020; Santoni de Sio & Mecacci, 2021). This may be difficult not only because we lack the ontological grounds for placing blame on the right agents (as Matthias (2004) suggested)—the responsibility attribution gap—but also because it is epistemologically difficult to correctly determine how blameworthy the agents are (Königs, 2022)—the responsibility distribution gap. Both dimensions—identifying agents who meet the relevant threshold for being ascribed responsibility and determining the extent of their responsibility—are thus integral to understanding responsibility gaps introduced by AI systems. While the former dimension highlights the potential impossibility of identifying any agent who can meaningfully bear responsibility, the latter emphasizes the potential impossibility of apportioning responsibility to agents, once they have been identified. Therefore, resolving the ontological issue of who is responsible addresses only the attribution dimension of responsibility gaps. The epistemological issue of how much responsibility each agent bears—the distribution dimension—remains unresolved. This leads directly to the classic problem of many hands.

How many agents are responsible and to what extent for the outcomes generated by AI? This question encapsulates the problem of responsibility distribution, or the “problem of many hands”—a problem that is primarily epistemological in nature, as it stems from a lack of knowledge and clarity (Van de Poel et al., 2012). Because it involves multiple parties with multiple tasks assigned, the deployment of AI systems has an inherent collective dimension (Coeckelbergh, 2020). Beyond the involvement of many developers, users, and regulators (Taddeo & Floridi, 2018), the distribution of responsibility in AI deployment is further complicated by the issue of ‘many things’—the many complex technologies that causally contribute to AI outcomes (Coeckelbergh, 2020). To more accurately capture the distribution dimension of responsibility gaps in AI deployment, it is therefore appropriate to frame the issue as one of collective responsibility. This perspective underscores that responsibility distribution gaps in AI deployment are, at their core, a matter of collective action, which inherently overlaps with the problem of many hands.

Prior Treatment of the Many Hands Problem

Although responsibility gaps in AI deployment have recently become a topic of great interest, they intersect with broader discourses on collective action across disciplines, most notably in business ethics (Bovens, 1998; French, 1995; Kaptein & Wempe, 2002) and engineering ethics (Van de Poel et al., 2012). Given the involvement of multiple agents, tasks, and technologies in AI deployment, the attribution of responsibility for AI outcomes must be understood as a matter of collective responsibility. Accordingly, addressing responsibility distribution gaps in AI contexts requires a thorough exploration of their connection to the problem of many hands in collective settings, as extensively examined in the business ethics literature.

The problem of many hands can be understood “in terms of a morally problematic gap in the distribution of responsibility among the members of a collective” (Van de Poel et al., 2012, p. 50). The problem arises from the fact that organisations “consist of staff, departments and divisions, each with their own functional responsibilities, which should be geared collectively, and which should be furnished with the necessary assets for giving expression to the collective responsibilities” (Kaptein, 1998, p. 116). There are four main aspects of how the issue of many hands has been conceptualised within organisations.

First, the long chain of responsibility within complex organisations makes it more difficult to identify all those who are directly responsible. The larger the organization, the fewer people can be held responsible, as Van Gunsteren (1974) observes. Bovens (1998) formulates this distribution of responsibilities as a paradox, noting that individual responsibility for actions decreases more than proportionally as the number of individuals among whom responsibility is distributed increases.

Second, because responsibility is shared among many individuals, “no one feels personally and directly accountable” (Badaracco, 1992, p. 71), and responsibility becomes diffused: “no individual feels the need to take responsibility, so in the end no one does, and unethical conduct is more likely” (Treviño & Nelson, 1995, p. 161). This challenge is further complicated by the issue of “many fingers” (Kaptein & Wempe, 2002, p. 169), a subset of the many hands problem, which arises when individual and, consequently, organisational responsibilities become obscured or neglected due to the many tasks assigned to each employee.

Third, another aspect of the many hands problem in organisations is related to the relationship between individual tasks and collective output. The division of responsibility “means that organizational members essentially do their jobs with blinders on—they see only what’s directly ahead of them and no one sees (or takes responsibility) for the whole picture” (Treviño & Nelson, 1995, p. 162).

Fourth, the difficulty of ascribing moral responsibility to one or multiple agents is linked to the relatively independent nature of the contributions of various agents, which manifests in two directions. On the one hand, there are situations of causal overdetermination, where multiple contributions, each sufficient on its own to generate the harmful outcome, lead to a scenario where “Since the result is overdetermined, no one agent could have prevented it by acting differently, so it seems each agent is off the hook” (Baum et al., 2022, p. 11). On the other hand, there are situations of causal underdetermination, where several contributions, each morally neutral or harmless in isolation, combine to generate a harmful outcome. In these cases, “no agent by herself could have prevented the result, so again it is difficult to ascribe moral responsibility to anyone” (Baum et al., 2022, p. 11).

Given that the many hands problem arises not only within collective settings but also more broadly in relation to collective outcomes, we draw attention to the analysis of the many hands problem by Van de Poel et al. (2012, p. 60) in the context of climate change, which illustrates that “it is sometimes very difficult, if not impossible, to hold anyone responsible for a collective harm”. This observation leads Van de Poel et al. to shift from Bovens' (1998, p. 47) framing of the problem—where a collective is responsible for an undesirable outcome, yet no individual within the collective is responsible—to a new definition. They propose a definition of many hands as “… a gap in a responsibility distribution in a collective setting that is morally problematic” (Van de Poel et al., 2012, p. 63). While we find this definition valuable, we suggest a slight modification that considers the fact that many hands problems occur not only within collective settings but also in relation to collective outcomes to which multiple agents—potentially both individual and collective—contribute. Therefore, we propose extending the original definition to include the gap in responsibility distribution leading to a morally problematic collective outcome.2 This results in the following revised definition:

A problem of many hands occurs if there is a gap in a responsibility distribution in a collective setting or for a collective outcome that is morally problematic.

This extended definition more accurately captures responsibility gaps in AI deployment, which, like climate change, is not confined to a collective setting (e.g., one organisation) but rather involves multiple societal agents, including individuals and collectives such as developing organisations, computer scientists, users, and regulators. Moreover, the definition aligns with the sense of responsibility that one usually has in mind when discussing responsibility gaps. This sense is backward-looking responsibility for past actions or outcomes, primarily in the sense of responsibility-as-blameworthiness (Van de Poel et al., 2012). In this context, “an agent can reasonably be held responsible-as-blameworthy if and only if certain conditions are fulfilled” (Van de Poel et al., 2012, pp. 52–53). While it is beyond the scope of this article to fully elaborate these conditions, scholars typically highlight freedom and epistemic conditions (Coeckelbergh, 2020; Fischer & Ravizza, 1993; Hakli & Mäkelä, 2019; Sison & Redín, 2023), with further nuances such as causation, deliberation, or coercion (Constantinescu et al., 2022).

The extended definition of the many hands problem, following Van de Poel et al., also applies to forward-looking responsibility, especially as this dimension is inherently connected to backward-looking responsibility: failure to fulfil current or future role-responsibility renders an agent blameworthy in a retrospective evaluation (Bovens, 1998). The way in which these responsibility gaps are evaluated as morally problematic may also be influenced by particular perspectives over responsibility, such as virtue-ethics, deontology or consequentialism (Van de Poel & Royakkers, 2023).

Large Language Models Exacerbate the Problem of Many Hands

The widespread use of LLMs enabled by transformer neural network architecture (Devlin et al., 2019; Vaswani et al., 2017) and currently powering chatbots, such as ChatGPT, Claude and Mistral, introduces new challenges in terms of responsibility distribution gaps in AI deployment. Unlike their antecedents developed for neural machine translation, transformer-based models in natural language generation “can be trained on unlabelled text, which potentially makes all text available on the internet training data for these models” (Gubelmann, 2024, p. 6). Their large scale is related to increases in both the number of parameters and the amount of training data (Bender et al., 2021), which have turned them into general-purpose language models, used across disciplines as broad as medicine diagnosis, computer programming, essay writing or music composition. Multimodal large language models generate diverse forms of content—including images, video, audio, and text—across a wide range of disciplines. As a result, the effects of LLMs at the societal level are more profound, ramified, and unpredictable than those of other AI systems that serve a more specific purpose, such as identifying images containing bicycles and pedestrians, thus broadening the potential risks.

The extensive taxonomy developed by Weidinger et al. (2022) outlines several risks associated with LLM development. These risks include (a) discrimination, hate speech, and exclusion, which contribute to social stereotypes, unfair discrimination, offensive language, and exclusionary norms; (b) information hazards, such as privacy compromised by leaks of sensitive information; (c) misinformation harms, including the dissemination of false or misleading information and the material harm resulting from the dissemination of poor-quality information in fields like medicine or law; (d) malicious uses, such as the reduced cost and increased effectiveness of disinformation; (e) human‒computer interaction harms, such as the promotion of harmful stereotypes as well as amplified risks of user nudging, deception, or manipulation due to human-like interaction; and (f) environmental and socioeconomic harms arising from the operation of LLMs, with include risks such as increased inequality and negative impacts on job quality or unequal access to benefits due to hardware, software, and skill constraints.

Importantly, LLMs expand exponentially the number of agents involved in their deployment—the longer the chain of agent responsibility involved in the entire cycle of AI systems is, the less responsibility is assigned to each agent. This amplifies situations of causal underdetermination for LLM deployment: multiple agents contribute, with each contribution below a harmful threshold but with their combined actions collectively producing a harmful outcome. The infamous example of the lawyer who supported his case with false facts provided by ChatGPT illustrates this situation, “prompting a judge to weigh sanctions as the legal community grapples with one of the first cases of AI “hallucinations” making it to court” (Bohannon, 2023). The risks of LLMs seem particularly noteworthy in terms of content toxicity (Dwivedi et al., 2023) and hallucinations (Bender et al., 2021) when used to generate advice in high-stakes domains, such as health, safety, legal or financial matters (Hagendorff, 2024).

This longer chain of agent responsibility exacerbates situations where those involved in LLM deployment perform their tasks with ‘blinders on,’ failing to see the broader consequences of their actions and focussing narrowly on their specific roles. For example, computer scientists working to enhance the efficacy of a language model may struggle to recognize that an unintended side effect is the educational deskilling of entire generations of students, who, whether due to laziness or an overreliance on the cognitive abilities of chatbots, outsource their essay assignments to these systems. This, in turn, leads to “potential homogenization of writing styles, the erosion of semantic capital, or the stifling of individual expression” (Hagendorff, 2024, p. 9). In parallel, responsibility for the harm generated by LLMs is evaded, with blame shifting from model-developing computer scientists to data scientists, and then to users and regulators. Additionally, LLMs raise issues about blame–credit asymmetry when moral responsibility is ascribed to users. Specifically, individuals using LLMs to generate good outcomes are evaluated as less praiseworthy in terms of moral responsibility than those who generate negative outcomes with the same amount of human effort (Porsdam Mann et al., 2023).

As shown by Gubelmann (2024), despite impressive abilities to autonomously optimise their parameters during training, these LLMs cannot be properly regarded as moral agents and therefore bear no responsibility for their outcomes. This is related to multiple factors. Some concern technical limitations, given LLMs’ merely statistical underlying process, as they “do not output text directly—rather, they produce a probability distribution over different utterances from which samples can be drawn” (Weidinger et al., 2022, p. 215). Other factors are related to broader concerns. Similar to previous AI systems with high autonomy, LLMs are not persons capable of engaging in a moral relationship (Sparrow, 2021); nor are they autonomous enough to be granted moral agency (Hakli & Mäkelä, 2019); nor do they meet standard criteria for ascriptions of moral responsibility, such as the freedom and epistemic criteria (Constantinescu et al., 2022). Relatedly, the very epistemology of generative AI is potentially a source of responsibility gaps. LLMs operate as “black boxes,” with opaque internal processes that produce inherently unpredictable outputs—a defining characteristic of these systems. This finally creates confusion regarding the agents who bear responsibility, as LLMs are inevitable bound to remain unpredictable to a certain degree. In turn, LLMs generate a troubling effect whereby human-generated and artificially-generated content becomes indistinguishable (Strasser, 2024). Given the ease of accessing LLM chatbots with free accounts or no accounts at all, there is an unprecedented possibility of generating text and images that easily pass as human-made, with perverse effects in fields as diverse as education, media, and healthcare.

Many-Agents-Many-Levels-Many-Interactions (M3) Approach

We propose a “many-agents-many-levels-many-interactions” approach, labelled M3, to address responsibility gaps in LLM deployment by considering how the distribution of responsibility is influenced by the range of interactions among agents placed at various levels of action. We argue that these interactions do not lead to a diffusion of responsibility to the point where we face unbridgeable gaps in the distribution of responsibility. Instead, agents involved in more interactions accumulate greater power and influence over other agents and bear greater responsibility for harmful outcomes of LLM deployment. Our proposal thus offers a non-distributive solution to responsibility distribution gaps generated by LLM deployment. Specifically, we suggest that responsibility for LLM outcomes does not necessarily become lost because it is distributed between the multiple agents involved in LLM deployment. Instead, we can address responsibility distribution gaps dynamically, by focusing on the interactions that occur between agents and that evolve over time, generating cumulative and synergetic—as opposed to distributive—effects. These effects amplify over time to the point that agents engaged in more interactions bear enhanced responsibility given their power to influence other agents with whom they interact.

Many Agents Bearing Moral Responsibility

Part of the responsibility gap in AI deployment, particularly LLMs, comes from the difficulty of distributing moral responsibility for AI-generated outcomes to the multiple agents involved. The current discourse on the responsibility of programmers, developers, and designers for outcomes of AI deployment, as articulated in previous research (Baum et al., 2022; Kiener, 2022), lacks adequate consideration of the organisations employing these professionals, as well as the power and corresponding responsibilities associated with organisations developing AI systems. To fully account for the collective nature of AI deployment (Coeckelbergh, 2020), it is not sufficient to consider the collective dimension of the multiple individual “hands”. We must also consider the collective agents involved, such as the organisations involved in the full cycle of LLM deployment, including those that gather datasets; develop and train models; implement these models in digital or physical systems (e.g., chatbots, AI robots, LLM-powered avatars and digital twins); market, use or maintain AI systems; oversee, regulate or evaluate these systems’ outcomes; or, finally, terminate these systems.

Therefore, we find it important to include among the many hands that bear moral responsibility for LLM deployment both individuals (e.g., users, designers, engineers, computer scientists) and structured collectives, such as organisations (e.g., developing companies, manufacturing companies, corporate users, regulatory bodies). Business ethics research has long argued that organisations should be regarded as moral agents per se and be ascribed collective moral responsibility for their outcomes, in addition to the individual responsibility of their members (see Constantinescu, 2024; Kaptein & Wempe, 2002; Phillips, 1995 for an overview of arguments).

However, identifying humans and organisations as appropriate moral agents addresses only one dimension of the responsibility distribution gap. Additionally, we need to consider the relationships between agents in collective contexts. As Nissenbaum (1996: 29) points out, “Where a mishap is the work of “many hands,” it may not be obvious who is to blame because frequently its most salient and immediate causal antecedents do not converge with its locus of decision-making.” Moral responsibility can be ascribed to decisions, actions, and outcomes (Baum et al., 2022). When one is morally responsible for an outcome, this typically involves being morally responsible for the decisions leading to the actions, the execution of the actions leading to the outcomes, or both. Nonetheless, there are situations in which one bears moral responsibility for an action without being morally responsible for the full range of decisions leading to that action. For instance, a researcher who fine-tunes an LLM on biased data as specified in a research protocol bears responsibility for the technical action but not necessarily for the decision to use problematic data if they were not involved in dataset selection. Similarly, a content moderator who implements filtering rules for LLM outputs bears responsibility for properly executing those rules but not necessarily for the higher-level decisions about which content categories should be restricted.

In other cases, one may bear moral responsibility for an outcome without being morally responsible for the full range of decisions and correlated actions leading to that outcome. For example, the CEO of an AI company bears responsibility for harmful outcomes caused by their LLM products, even if they did not personally make all the technical decisions or perform the model training actions. Similarly, LLM developers bear some responsibility for job displacement outcomes, even if they were not involved in the business decisions to implement automation using their models. An AI safety lead bears responsibility for harmful outputs of an LLM, even if they did not design the architecture that enabled those capabilities. Similarly, researchers who develop capabilities that enable more convincing deepfakes bear some responsibility for subsequent misuse, even if they did not participate in or approve of the harmful applications.

Because LLMs deepen the problem of causal underdetermination, they make it more difficult to connect, for instance, effects such as undetectable forms of academic fraud through chatbots with the multiple agents standing behind this type of outcome. This calls for an approach that considers not only the many agents (“hands”) involved in AI deployment but also the many levels of moral responsibility.

Many Levels of Moral Responsibility

To date, most discussions concerning moral responsibility for highly autonomous AI deployment focus on the individual level of moral responsibility, such as among individual users, designers, engineers, developers, and programmers (Baum et al., 2022, Coeckelbergh, 2020; Johnson & Verdicchio, 2019; Loh & Loh, 2017). Very few address the collective level, which includes organisations, national and international regulatory bodies, NGOs and civil society, and transnational bodies (Hellström, 2013; Wright & Schultz, 2018). Attempts to correlate all these levels are rather the exception (such as Lenk, 2017; Tóth et al., 2022). For example, Lenk (2017: 223) suggested that the effects of technology are best captured through a “hierarchical model that adequately and differentially puts the responsibilities on the various levels”, acknowledging that “individual responsibility and corporate responsibility do not as such have the same meaning; they cannot simply be mutually reduced to one another”. Notably, Tóth et al. (2022) proposed an AI robot accountability approach based on “accountability clusters”, which includes multiple actors and institutions, positioned on micro–meso–macro levels, together with connections within and across levels, as well as among actors, to guide the deployment of AI systems in business settings.

Business ethics research has long discussed the relevance of various levels of responsibility ascriptions for harmful outcomes of corporate activity, as a response to the problem of many hands and to the potential responsibility voids generated by corporate collective action. After an initial focus on the “personal characteristics of individual transgressors” (Kaptein, 2011: 844), research started to pay attention to “the characteristics of the organizational context within which unethical behavior occurs” (idem) in a shift from a “bad apples” approach to a “bad barrels” approach (Treviño & Youngblood, 1990). The individual and organisational levels of responsibility were subsequently complemented with governmental and international regulatory levels of responsibility for outcomes of corporate activity, thus adding the “bad cellars” and “bad orchards” (Hibel & Penn, 2020; Muzio et al., 2016) approaches to the discussion.

Drawing on these findings related to business ethics and AI ethics, we suggest that a four-level account is the best fit to address the distribution of moral responsibility for the deployment of general-purpose AI models with deep societal implication, particularly LLMs. The microlevel of LLM deployment is the equivalent of the “apples” approach in business ethics, which focuses on individual responsibility; the mesolevel is the equivalent of the “barrels” approach, which focuses on organisational responsibility; the macrolevel mirrors the “cellars” approach, which focuses on governmental responsibility; and, finally, the supralevel is the equivalent of the “orchards” approach, which focuses on international regulatory body responsibility.

Many Interactions Grounding Moral Responsibility

A major additional difficulty that arises when moral responsibility is distributed to those involved in LLM deployment is related to the “complex ramifications of impacts, consequences and side effects” (Lenk, 2017: 219), which are further complicated by “complex systems interactions and dynamic changes easily transgressing linear thinking and traditional causal disciplinary knowledge” (idem). To account for the deep entanglement of multiple agents, which include individual developers and users, collective organisations developing and implementing LLM models, and collective bodies overseeing LLM deployment, we need to consider nonlinear interrelations and interactions that incorporate the complexity of individual, collective, and corporate contributions. In doing so, we must acknowledge the inadequacy of strategies that place moral responsibility solely on single agents or on whole systems as such (Lenk, 2017).

In practice, the use of LLMs can often become less cooperative and more conflictual, with competing interests emerging at the micro, meso, macro, and supra levels. For instance, individual users may seek to exploit LLMs for dishonest purposes, such as cheating on their homework; LLM-developing companies may exploit the psychological vulnerabilities of their users in order to push their model for increased profit; the government may utilize LLMs for ideological propaganda; and international organizations might establish policies that favour some parts of the world to the detriment of others, thereby exacerbating epistemic injustice.

In addition to acknowledging the micro–meso–macro–supra levels of responsibility in LLM deployment, it is important to acknowledge that these levels are deeply interconnected and that the responsibility placed at one level often impacts that at other levels. This requires moving away from the reductionist assumption that typically underlies the treatment of responsibility gaps in AI ethics and the many hands in business ethics, whereby moral responsibility is distributed among agents and becomes diffused to the point where no one meets the threshold for attribution of responsibility. We need a non-distributive solution for responsibility distribution gaps introduced by highly autonomous AI systems and complicated by LLMs. This solution considers the multiple collective dimensions of responsibility, including the agents involved, their level of involvement, and their interactions, all of which are necessary for a comprehensive responsibility assessment. Rather than assuming that responsibility for LLM outcomes is progressively diluted as it is distributed, we propose that we must consider how agents’ responsibility may actually increase due to the influence they exert on the moral responsibility of other agents involved in generating LLM outcomes.

In lieu of the reductionist assumption, we need to work with a dynamic, interactional view of agents and levels of responsibility, and in this way identify how the responsibility of one agent in LLM deployment influences that of other agents. The interactions among agents across levels often generate cumulative and synergetic effects (Lenk, 2017). Business ethics research has previously emphasised how various connections among individual members of organisations lead to emerging organisational processes (Steen et al., 2006), which are correlated with the way individuals bear moral responsibility for the unethical behaviours of others by virtue of their participation in a shared-value culture (Dempsey, 2015). Scholars have discussed the amplifying effects of individual actions within organisations through role modelling, both as downwards spiralling that multiplies wrongdoing (Den Nieuwenboer & Kaptein, 2008) and as upwards spiralling that generates contagious repetition of virtuous behaviour (Cameron & Winn, 2012; Tsachouridi & Nikandrou, 2019). Both types of spirals become self-reinforcing in time, generating increased moral responsibility for all the agents involved, potentially in a mutually enhancing manner—whereby the interactive dynamics between individual actions and corporate structures can increase the degree of responsibility each party bears for ethical failures or achievements (Constantinescu & Kaptein, 2015).

For example, in the case of LLMs, we might have virtually simultaneous situations of computer scientists who do not sufficiently tweak a model, data scientists who do not make explicit the limits of datasets, LLM-developing organisations that do not observe correlations, regulators that do not limit harm to users, supra-territorial bodies that do not correlate social effects, individual users who do not report issues from their interaction with chatbots and so on. The way in which each agent currently exercises their own role responsibility impacts other agents’ ability to exert moral responsibility as an obligation in the future. When neither organisations nor users nor regulators take the necessary steps to ensure safe deployment of LLMs, each negligent behaviour influences others to behave negligently, thus generating an overall pattern of neglect associated with each agent’s role. Unethical behaviour by LLM-developing organisations is facilitated by negligent—or lack of—regulatory action and, in turn, leads to unethical use by individuals. Current failure to assume prospective responsibility results in being ascribed retrospective moral responsibility and blame in the future.

To account for the distribution-responsibility gap, it becomes relevant to consider a multidirectional concept of moral responsibility, whereby agents exert a bidirectional or multidirectional influence on the other agents: from designers to programmers, from users to LLM-developing organisations, from regulators to marketers, and so on. Each agent is backward-looking responsible not only for their own behaviour but also forward-looking responsible for stimulating and enabling others to behave in a certain way. The more the agents stimulate and support unethical practices by others, the more they fail to meet their current responsibility and become blameworthy over time, in a mutually enhancing manner (Constantinescu & Kaptein, 2015).

Centrality as a node in the web of interactions surrounding LLM deployment is a factor that increases the degree of responsibility attached to an agent. Some agents may be engaged in more interactions than others, thus potentially bearing greater influence over other agents at multiple levels. This further suggests that agents engaged in more interactions play the role of nodes in the web of LLM deployment and may thus be ascribed more moral responsibility. Theoretically, as the positions of the nodes change constantly, there is, in principle, no predetermined degree of responsibility attached to positions. In practice, however, it is often the case that some agents play the role of a node more than others do, thus exerting more influence on the others.

LLM-Developing Organisations as Nodes of Interaction

We suggest that organisations developing general-purpose AI models, such as LLMs, play a key role as nodes of influence on other agents involved in LLM deployment, given their centrality in interactions across all micro, meso, macro, and supra levels in our approach. First, through their internal structures, these organisations influence individual members’ ethical behaviour (micro level) in the production, implementation, and use of LLMs. A pertinent example is the resignation of former employees from organisations such as OpenAI due to their disagreements with corporate policy (Perrigo & Henshall, 2024). Second, organisations developing LLMs engage in multiple interactions with other similar organisations—for example, the release of a new LLM creates competitive pressure on others to develop comparable or more advanced LLMs (meso level). A case in point is the release of DeepSeek, which has reportedly pressured OpenAI to advance its ChatGPT model further in order to justify its costs and pricing structure (Schiffer, 2025). Third, these organisations significantly impact entire industry alliances at the regional or national level, as well as regulatory bodies that need to monitor, prevent, and punish unethical deployment of AI systems (macro level). Examples include bodies such as IEEE, the OECD or UNESCO releasing principles and guidelines for ethical development of AI, as well as the more recent European Commission’s AI Act. Fourth, these organisations set the agenda for global transnational bodies in terms of oversight, compliance, guidelines, and other mechanisms (supra level). Given the current lack of such supra-territorial worldwide bodies regulating general-purpose AI models, LLM-developing organisations possess vastly more power than any other agent in the LLM deployment process.

LLM-developing organisations occupy a pivotal role, as they are uniquely positioned to understand, foresee, evaluate, and prevent the risks and harms associated with LLM deployment. Compared to other agents, these organisations possess significantly greater knowledge regarding LLM outputs. As Weidinger et al. (2022) explain, this advantage stems from several factors: (a) the rapid transition from research to real-world application hinders third-parties’ ability to anticipate and mitigate risks; (b) this challenge is further compounded by the high level of technical expertise and computational resources required to train and fine-tune LLMs; and (c) limited access to raw LLMs constrains the broader research community to conduct comprehensive risk assessments and implement early mitigation strategies. Moreover, organisations involved in training and developing LLM applications are particularly well-placed to address issues such as discrimination and bias in training data more effectively during the model training phase than during subsequent downstream product development (Weidinger et al., 2022). This positions LLM-developing organisations at the origin of the causal chain of LLM outcomes and confers upon them a distinct power position relative to other agents involved in the AI deployment, a position that normatively warrants the ascription of more moral responsibility.

Organisations developing LLMs for broad audiences—such as OpenAI, Microsoft, Google, and Anthropic—open the causation chain for a complex range of harmful uses of these systems. General-purpose LLMs can be openly accessed free of charge, without relevant rules set internally, without external supervision, and by people who do not face the risk of being denied future use of these systems when they purposely use them to inflict harm. LLM-developing organisations thus create the possibility for harmful behaviours that would not be possible without these technologies, particularly in terms of the difficulty in distinguishing between artificially generated and human-made visual or textual content. This places LLM-developing organisations in a relevant causal relationship with outcomes generated by users, even if these outcomes were unintended by the organisations. Such a causal relationship provides the normative grounds for ascribing moral responsibility and blame to these organisations. While moral responsibility is not reducible to causation and involves other elements, such as acting with knowledge, toward a certain goal, and free from unbearable coercion (Constantinescu et al., 2022), being in the position to initiate a causal chain of events is a normative foundation for bearing responsibility for the outcome of that chain (Popa, 2021), regardless of its length.

For instance, the difficulty in distinguishing between chatbot-generated and student-generated essays in academia has led to a complex restructuring of evaluation methods, including handwritten examinations. Externalising such costs onto third parties, such as students and universities, unjustifiably absolves LLM-developing organisations of moral responsibility. Measures to internalise these externalities could include a requirement that organisations developing LLMs also invest in offering the relevant software to detect artificially generated content that would mitigate malicious uses, such as deepfakes or plagiarism. Additionally, making their models openly investigable regarding the sources used to generate artificial content would help mitigate effects such as copyright infringement.

Three further nuances concerning power asymmetries and moral responsibility need to be distinguished. First, not all LLM-developing organisations possess the same power to influence others. Big tech—and start-ups or NGOs partnering with big tech—have considerably more power than smaller organisations developing LLMs, as they have more access to hardware and datasets that enable fine tuning and greater model accuracy. Second, not all individuals working in LLM-developing organisations bear the same degree of moral responsibility. Business ethics studies emphasise the key role of individuals in higher hierarchical positions and levels of decision-making within organisations (Kaptein & Wempe, 2002). There is an interplay between individual and corporate action that underlies moral responsibility dynamics (Constantinescu & Kaptein, 2015). This makes a case for looking into multiple agents’ levels and interactions inside organizations that deploy LLMs. Third, there is a power asymmetry between organisations that are key loci of moral responsibility for LLM development and those that are in a position to hold these organisations responsible for the outcomes of their LLMs. Given the lack of proper transnational legislation and regulatory oversight, individuals, academics, and NGOs alone cannot counterbalance the societal risks and negative effects of LLMs released to the general public, despite the acknowledgement of the need to hold organisations more accountable (Theodorou & Dignum, 2020).

To what extent does the M3 approach, with its focus on organisations as nodes of interactions and, consequently, as a key locus of responsibility, succeed in addressing the problem of the deficit between decision-makers and responsibility-bearers? First, LLM-developing organisations become responsibility-bearers for the diffuse and difficult-to-pin effects of their general-purpose AI models. As LLMs become more accurate, they evolve into powerful decisionmakers—or at least powerful foundations for human decision-makers. A pertinent example is the recent scandal involving the UK technology secretary using ChatGPT for policy advise (Stokel-Walker, 2025). While unintended by developing organisations, these types of LLM use cases still need to be positioned within the broader framework of corporate responsibility. In contrast to clear instances of LLM misuse, such as academic fraud or the proliferation of fake news in political propaganda—where users should primarily be held responsible—more ambiguous cases require shifting some responsibility to the LLM-developing organizations. Use cases such as medical diagnosis or public policy, where LLMs are used to ground decision-making, should be traced back to those agents that have more capability to foresee the potential dynamics of negative societal effects, given their involvement in more interactions across more levels of actions compared to other agents. Such agents are LLM-developing organisations. Unintended effects do not absolve these organizations from attribution of blame; failing to recognize the potential negative effects, when one has the capabilities to do so, is akin to negligent or even reckless behaviour, which is blameworthy.

However, in practical terms, holding LLM-developing organizations liable for a wide range of negative societal outcomes might raise concerns about economic viability, given the boundless user count and use case diversity. While this article does not intend to provide a detailed practical solution to the fundamental deficit between decision-makers and responsibility-bearers, but rather to establish normative grounds that could guide such solutions, we note here that our efforts to clarify the distribution of responsibility represent an important first step in addressing this deficit. By arguing that the more an agent is involved in multiple interactions across multiple levels of LLM deployment, the more it should be considered as a responsibility-bearer for negative outcomes of LLMs, we offer a tool to reduce this deficit.

Finally, positioning LLM-developing organizations as nodes of interaction and as key loci of responsibility does not exclude other agents as potential nodes of responsibility. As discussed above, the positions of the nodes are dynamic and subject to change based on the interactions involved. Given the global scope of LLM development, with competing interests between organisations across continents and the potential for politization of LLMs—exemplified by entities such as DeepSeek—the role of international regulatory bodies becomes increasingly crucial. In fact, supra-territorial bodies, when and if supra-territorial bodies are established, may and likely should become nodes of interaction and responsibility as well.

Implications for Policy-Making and Managerial Practice

The findings presented in this article bear direct implications for both policy-making and managerial practice. Future regulatory frameworks must consider the positions held by different categories of organisations as nodes of interaction. The greater the power and influence an entity exerts over others, the greater that entity’s degree of responsibility for the outcomes of AI deployment is. For instance, the key role played by LLM-developing organisations and their power to influence the moral responsibility of other agents across AI deployment offers arguments for the introduction of organisation-focused regulation (Ferretti, 2022; Kokshagina et al., 2023). This provides normative grounds for imposing greater role responsibility on LLM-developing organisations for future preventive actions.

This point aligns with regulations at the European level, where the AI Act, fully effective from August 2026, establishes distinct obligations for providers of general-purpose AI models (GPAIMs), which include LLMs (AI Act, 2024, Chapter V, Annexes XI-XII). These provisions address transparency regarding model capabilities and limitations, as well as the content used for training. Additionally, the AI Act lays out further obligations for providers of GPAIMs that present systemic risks, determined by criteria such as model capabilities (parameters, training computation, benchmarks), market reach (at least 10,000 registered business users in the EU), and input/output modalities (AI Act, 2024, Annex XIII). Providers of GPAIMs that pose systemic risks must also address issues such as risk assessment and mitigation, incident reporting, cybersecurity protection, post-market monitoring, and value chain cooperation.

Importantly, the European AI Act distinguishes between “high-risk AI systems” and “general-purpose AI models with systemic risk” as two separate regulatory categories requiring different oversight approaches. High-risk AI systems are defined by their potential to harm individuals’ health, safety, or fundamental rights in specific deployment contexts (AI Act, 2024, Article 7; Annex III). The Act applies a use-case approach to high-risk regulation, focusing on how and where AI is deployed rather than the capabilities of the underlying model. In contrast, systemic risks apply specifically to general-purpose AI models with high-impact capabilities or significant market reach (AI Act, 2024, Articles 51 and 55, Annex XIII). These risks extend beyond individual harms to include potential large-scale societal impacts, such as actual or foreseeable negative effects related to major accidents, disruptions to critical sectors, serious consequences for public health and safety, negative effects on democratic processes, public security, economic security, and the dissemination of illegal, false, or discriminatory content. The regulatory approaches for high-risk AI systems and general-purpose AI models with systemic risk differ significantly in terms of compliance mechanisms. While high-risk systems follow conformity assessment procedures, providers of general-purpose AI models with systemic risks may demonstrate compliance through codes of practice, harmonized standards, or alternative adequate means (AI Act, 2024, Article 56).

While the AI Act is the first legal framework for AI, with limited applicability to the EU market, its effects could extend further. However, two weaknesses highlighted by our many hands-many-levels-many-interactions approach should be noted. First, the problem of distributing responsibility for general-purpose AI models, such as LLMs, is only partially addressed through the systemic risks-based approach. Well-known LLMs, such as ChatGPT, Claude, or Gemini, typically fit the category of GPAIMs with systemic risks. However, this category imposes fewer obligations on organisations developing these models compared to those related to high-risk AI systems. Specifically, providers of GPAIMs with systemic risks can simply demonstrate compliance with the AI Act by adhering to the General-Purpose AI Code of Practice, or by conforming to harmonized standards or alternative means. This offers providers of GPAIMs with systemic risks the option to demonstrate compliance using their own preferred means. Given the broad societal impact of LLMs, as well as the position of LLM-developing organisations as nodes of interaction, a more promising approach under the AI Act would be to more closely align GPAIMs with systemic risks with high-risk AI systems and impose stricter provisions on LLM-developing organisations.

Second, the AI Act defines providers as “a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge” (AI Act, 2024, Article 3(3)). Our understanding of LLM-developing organisations as providers of general-purpose AI models is more specific, encompassing only the organisations that develop and market these models. These organizations are highlighted as key nodes of interaction and key locus of responsibility for LLM outcomes in the current article. The second part of the AI Act’s definition, referring to entities that deploy GPAIMs under their own name or trademark, pertain more to organisations that deploy models under their name. A more nuanced approach to different types of organisations within the category of GPAIMs providers could be relevant when distributing responsibility for outcomes of LLMs.

Beyond the AI Act, the role of soft regulation complementing hard regulation needs to be considered. This includes the development of standards for ethical AI deployment, such as the IEEE P70xx ethics standards, and certification programs for intelligent and autonomous systems developed by the IEEE Standards Association (IEEE, 2019). These standards could complement or become part of hard regulation (Theodorou & Dignum, 2020). Furthermore, they could offer regulators relevant tools to address the complex societal side-effects of LLMs and encourage LLM-developing organisations to internalise their moral and economic externalities. For instance, LLM-developing organisations should not only be backward-looking morally responsible for the data they feed into their models but also forward-looking morally responsible for how they allow the users to feed the model in a problematic way–even when this is not directly harmful. Regulation should emphasize future role-responsibility for LLM-developing organisations, which includes more oversight and a curatorial approach to model development.

At the managerial level within organizations, the collective dimension of responsibility for LLM outputs requires consideration of the many hands both inside and outside organizations. In this regard, the M3 approach is relevant for the many corporate agents interacting across various levels of LLM deployment, as well as for the many individual agents interacting across various levels inside organisations developing, deploying, using or regulating LLMs. Individual moral responsibility complements organisational moral responsibility. For individuals to effectively take on future responsibilities, they need to clearly understand what is expected of them in their upcoming roles. Certain individuals within organisations bear greater responsibility due to their position as nodes of interactions across departments, teams and divisions, with more influence over others, especially in light of factors such as hierarchical position and decision-making power.

When it comes to individuals who are a key locus of responsibility inside LLM-developing organisations that are themselves a key locus of responsibility, it seems that CEOs and other executives act as internal nodes of interactions and responsibility, given their power within the organisation and in relation to agents outside the organisation. Shareholders, in addition to CEOs and other executives, bear significant responsibility for influencing a company’s direction (Kaptein, 2025). As the legal owners of the enterprise, shareholders have a moral obligation to ensure their company operates ethically in its interactions with all stakeholders.

Finally, the issue of ensuring AI literacy for individual users of LLMs across society stands at the intersection of regulation and LLM-developing organisations. Future role responsibility, whether imposed by regulators or voluntarily undertaken by organisations, should include enhancing general literacy around LLMs, particularly at various educational levels. For example, LLM-developing organisations could collaborate with educational institutions or NGOs to provide free, tailored courses to students in secondary and tertiary education. As currently formulated, the AI Act only requires that providers and deployers of AI systems ensure “a sufficient level of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf” (AI Act, 2024, Article 4). These provisions should be expanded to include raising AI literacy among non-professional end-users of general-purpose AI models.

Conclusion

In this article, we propose a potential solution to the distribution dimension of responsibility gaps arising from the deployment of general-purpose AI models, particularly LLMs, by examining three elements: (1) the agents who bear moral responsibility, (2) the levels at which moral responsibility is ascribed, and (3) the interactions among agents and across levels, which serve as the normative basis for attributing responsibility. The main contribution of our approach lies in introducing the role of interactions among agents situated at various levels of LLM deployment as a normative grounding for moral responsibility ascriptions for LLM outcomes. Agents engaged in a greater range of interactions exert more influence over others and accumulate more causal power in relation to the outcomes of LLM deployment. We emphasise the importance of giving particular attention to collective agents, such as organisations, which play a pivotal role in the development and deployment of LLMs. An important outcome is that LLM-developing organisations occupy a nodal position within networks of interactions, which normatively grounds the attribution of greater responsibility—and, where appropriate, blame—for harmful LLM outcomes in comparison to any other agent involved in the LLM deployment process.

Our article contributes to the ongoing discussion of Responsible AI by advancing the M3 approach, which provides a framework for not only ascribing retrospective moral responsibility for harmful practices in LLM deployment, but also for encouraging agents to a proactive stance toward responsibility. When agents recognize that they will be held blameworthy, they are more likely to act responsibly (Johnson & Verdicchio, 2019). LLM-developing organisations must proactively reflect on their role and take further steps to understand, define, and assume corporate moral responsibility for AI outcomes. Our approach thus directly addresses concerns raised by Santoni de Sio and Mecacci (2021: 1063), who warn that responsibility culpability gaps are problematic insofar as “the more persons designing, regulating, and operating the system can legitimately (and possibly systematically) avoid blame for their wrong behaviour, the less these agents will be incentivised to prevent these wrong behaviours.” By offering a nuanced and multi-level account of the scope of individual and organisational responsibility, focused on the interactions between agents across levels, the M3 approach strengthens the moral and practical justification for LLM-developing organisations to adopt preventive strategies aimed at mitigating future harm.

Author Contributions

Both authors contributed to the study conception and design. Both authors read and approved the final manuscript.

Funding

Mihaela Constantinescu declares that her research has been funded by the European Union (ERC, avataResponsibility, 101117761). Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them.

Data Availability

Data sharing is not applicable to this article as no datasets were generated or analyzed during the current study.

Declarations

Competing Interests

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.