About two decades ago, “The Gap of Responsibility” as a problem was first introduced by Matthias to refer to an ethical challenge of AI. Due to their ability to learn, AI technologies are able to function beyond the control and anticipation of designers and users. Matthias argues that in this case, the necessary and sufficient conditions of responsibility would not be satisfied either for designers or for users, and that is the responsibility gap. The problem of the responsibility gap has invited many researchers to explore solutions. There is significant divergence among current accounts of the responsibility gap, which mainly results from their basic theoretical assumptions. In this paper, we categorize the responses to three approaches: dissolving the problem, filling the gap by attributing responsibility to AI, and filling the gap by attributing responsibility to human agents. By analyzing the social and practical aspects of the concept of responsibility, we defend a compound, contextual, and gradual notion of responsibility. On this basis, we claim that the first and second approaches are not successful, due to being grounded on poor concepts of responsibility. Hence, the third approach is more plausible. Moreover, since in the familiar cases of responsibility gap we fill the gap through revising social roles, in the case of AI-based responsibility gap we can follow a similar solution too, as suggested in some works of the third approach. Finally, we discuss some of remained challenges for the third approach.
Keywords: Responsibility, The Responsibility Gap, AI Ethics, AI Agency, Social Roles
Introduction
In his paper “The Responsibility Gap: Ascribing Responsibility for the Actions of Learning Automata” Matthias introduces the concept of the responsibility gap. He starts with a definition of responsibility and its necessary and sufficient conditions and proceeds to show how the development of AI can pose challenges for attribution of responsibility. Matthias asserts that we can fairly attribute responsibility to someone only when she has control over her behaviors and their consequences in a proper sense. Moreover, one should know certain relevant facts about her actions, and freely choose one way of acting among many. In usual cases of employing machines, if the operator uses the machine according to the manufacturer's manual, the manufacturer is responsible for the consequences, and otherwise, the operator is (Matthias, 2004: 175). But for the learning machines, this is not the case. Matthias examines four kinds of AI design, including symbolic systems, connectionism, genetic algorithms. and autonomous agents. For each case, he demonstrates that due to their learning capacities, the conditions of responsibility are not satisfied, neither for manufacturers nor for operators. In fact, he claims that these machines are so opaque and complex that no human meets the necessary conditions for being responsible for the consequences of their functions. And this is the gap of responsibility.
Materials & methods:
In the last two decades, many works have been dedicated to the AI responsibility gap and plausible answers for it. There is significant divergence among these accounts, which mainly comes from their basic theoretical assumptions. In this paper, we categorize the responses to three main approaches: dissolving the problem, filling the gap by attributing responsibility to AI, and filling the gap by attributing responsibility to human agents. By analyzing the social and practical aspects of the concept of responsibility, we defend a compound, contextual, and gradual notion of responsibility. On this basis, we defend the third approach. Moreover, we argue that in the case of the AI responsibility gap, like other responsibility gaps in the history of our social life, revising social roles is the basic pattern for overcoming the gap.

Discussion & Result:
- Dissolving the Problem
Some authors don’t empathize with Matthias about the essence and importance of the responsibility gap. Kiener, for example, argues that the real problem about AI and responsibility isn’t the gap of responsibility. Instead, it is the abundance of responsibility. Since every person who causally participates in hurting someone and has a moral obligation not to do so is responsible, these two conditions are satisfied by many (Kiener, 2025: 368).
On the other hand, Danaher and Munch et al. accept Matthias’ articulation of the responsibility gap, but assert that not only isn’t the gap evil, but it is even good. The main idea is that AI can perform tasks, thereby taking the burden of responsibility off humans’ shoulders. This in turn decreases the psychological pressures of responsibility and provides more joy and comfort for human agents, which is morally good (Danaher, 2022; Munch et al., 2023).
- Attributing Responsibility to AI
A main line of responses to Matthias’ challenge is to suggest reasons for filling the gap by attributing responsibility to AI. Here we have two approaches: property-oriented and relational. In the first approach, scholars introduce some properties (like autonomy, agency, etc.) as the basic conditions of responsibility, and then try to show that AI can actually or possibly meet these conditions (Haselager, 2005; Rodogno, 2016; List & Pettit, 2011; Floridi, 2016). The relational approach, instead, considers the status of AI in the network of our social relations. If we perceive an AI technology as if it is a responsible agent, and thereby make sense of its action, then we have good reasons to attribute responsibility to it (Sullins, 2006).
- Attributing Responsibility to Humans
For many scholars, attribution of responsibility to AI sounds implausible. On one hand, they make arguments to reject the responsibility of AI, and on the other hand, they develop accounts for preserving human responsibility in the age of AI technologies. Sparrow suggests Human-in-the-loop as a way of preventing the responsibility gap. He argues that the final action of AI should always be done only after the confirmation of a human agent (Sparrow, 2007). Also, Human-on-the-loop emphasizes that a human agent can only supervise or veto the function of AI (Scharre, 2018). But due to different factors like automation bias and over-trust of humans in AI, it seems that in both approaches the human control of the output wouldn’t be effective and real. As a response to these problems, meaningful human control (MHC) was suggested as a solution concentrating on designing the socio-technical network in a way that meaningful human control is preserved. Two conditions of tracking (designing the system for being reason-responsive and tracking humans’ moral intents) and tracing (designing the system such that there is always a human who understands the system’s function and its consequences) are assumed to serve this goal (Santoni de Sio & van den Hoven, 2018; Santoni de Sio and Mecacci, 2021, 1063–1076).
- Revising Social Roles as the Basic Solution
To find the best approach, it is proper to have a closer look at the concept of responsibility. As a social construct, this concept serves to reinforce people’s will for promoting moral values by putting some pressure on them. To do this task in the best way, we need a rich concept of responsibility which, despite monist accounts that equate it with one single notion (i.e., answerability) or pluralist accounts that consider it as having multiple distinct types (i.e., accountability, culpability, etc.), assumes responsibility as a compound notion which is constituted by attributability, accountability and answerability. A compound, gradual, and contextual notion of responsibility aligns with our daily use of the word and also makes it adequate for performing its role in morality. This thick notion discredits the first and second approaches toward the responsibility gap, which respectively dissolve it and attribute responsibility to AI. Both approaches are based on a thin notion.
On the other hand, the responsibility gap has familiar examples in the history of our social life. Humans have been hurt by various natural forces, where prima facie no human agent was responsible for that. But social roles, as dynamic parts of social structure, make it possible to provide a human with the requirements of responsibility, i.e., the needed knowledge and power, to take responsibility and prevent those harms as much as possible. Hence, in the case of the AI responsibility gap, the practical and successful solution should follow this basic pattern too: revising social roles and their corresponding norms in order to prevent harms which are currently out of our control.
Conclusion:
Contemplating the role of responsibility in our morality and our ordinary use of the term leads us to a compound, gradual, and contextual account of responsibility. This rich concept limits meaningful attribution of responsibility only to human agents. Hence, to overcome the gap, we should find ways for preserving human responsibility in AI systems. A familiar way of doing so is revising social roles, which includes defining new tasks, providing new resources of knowledge and power, and making responsibility attribution reasonable.