Abstract
The emergence of agentic AI systems that can pursue complex goals through multi-step planning and adaptive execution presents new challenges for attributing moral responsibility for the products developed by human-agentic AI teams. While the “responsibility gap” literature examines blame attribution for AI outcomes and the “moral crumple zone” concept identifies how humans absorb misattributed responsibility, neither framework adequately addresses the temporal distribution of moral responsibility across sequential agentic workflows in human-AI teams. This article introduces the responsibility cascade, a phenomenon whereby moral responsibility diffuses, dilutes, and redistributes across the sequential steps of an agentic AI workflow, contrary to traditional moral attribution frameworks. Drawing on philosophical accounts of guidance control, collective agency, and organizational responsibility, we argue that agentic AI creates chains of delegated agencies that fundamentally transform how responsibility propagates through human-AI collaborations. We propose a Sequential Responsibility Attribution Framework (SRAF) that distinguishes four modes of responsibility distribution in agentic workflows: initiation responsibility, checkpoint responsibility, delegation responsibility, and outcome responsibility. This framework provides diagnostic value by identifying responsibility gaps and normative guidance for designing agentic systems that preserve meaningful human control and avoid moral crumple zones. Beyond its philosophical contribution, the framework has direct implications for AI governance, informing how regulators, organizations, and system designers can preserve accountability as agentic AI becomes prevalent in consequential domains including healthcare, finance, software development, and critical infrastructure.
1 Introduction
1.1 Agentic AI
Artificial intelligence is undergoing a fundamental shift in its architecture. Unlike earlier AI systems that responded reactively to user prompts with separate outputs, modern agentic AI systems operate with much greater independence, pursuing complex goals through multi-step planning, tool use, and adaptive execution (Hughes et al., 2025). These systems do not merely generate content or predictions; they break down high-level objectives into smaller tasks, make intermediate decisions that influence future actions, interact with external environments via APIs and tools, and adjust their strategies based on feedback (Abou Ali et al., 2025).
These capabilities are now a reality. Today’s AI systems—including Claude, GPT-4, and Gemini—offer agentic modes that can browse the web, run code, manage files, and complete multi-step tasks with minimal human input. Businesses increasingly depend on AI agents to manage important workflows: venture capital firms use agentic systems for due diligence research, software teams deploy AI coding assistants that suggest, implement, and test changes across entire codebases, and healthcare providers pilot AI agents to coordinate patient care across multiple providers and systems. The European Union’s AI Act, the United States’ NIST AI Risk Management Framework, and new governance standards worldwide are all addressing how to ensure human oversight of these systems. Understanding how moral responsibility is distributed across agentic workflows is thus not merely a philosophical exercise but a pressing governance challenge with immediate practical implications.
This transition from reactive to agentic architectures introduces a fundamentally new philosophical issue for attributing moral responsibility. When a human initiates an agentic workflow—instructing an AI system to “research investment opportunities and prepare a recommendation,” “build and deploy a data pipeline,” or “manage this project’s resource allocation”—they set in motion a sequence of AI actions, decisions, and environmental interactions that may extend far beyond the scope of traditional frameworks for moral responsibility.
Consider a typical case: A project manager instructs an agentic AI system to optimize team resource allocation for a product development initiative. The system analyzes workload data, proposes task reassignments, adjusts project timelines, and communicates directly with management systems—with human approval only at designated milestone checkpoints. When the project then fails due to resource misallocation, how should moral responsibility be distributed? The human set the goal but did not control intermediate decisions. The AI made consequential choices but lacks moral agency. Checkpoint approvers authorized specific actions but lacked full insight into the system’s reasoning. The organization approved deployment, yet no individual had sufficient control over the outcome. This scenario exemplifies what we call the responsibility cascade—the progressive diffusion of moral responsibility across the sequential steps of an agentic workflow that resists traditional attribution frameworks.
1.2 Existing Frameworks
Contemporary philosophical engagement with AI and moral responsibility has generated valuable frameworks that, nonetheless, prove insufficient for the context of agentic AI.
Responsibility Gap. Matthias (2004) introduced the concept of the “responsibility gap” to describe situations where autonomous learning systems cause harm, but no human agent can be legitimately held culpable. Because learning systems can develop behaviors that their designers could not predict, traditional frameworks assigning responsibility to manufacturers or operators are inadequate. Santoni de Sio and Mecacci (2021) later elaborated four distinct responsibility gaps: culpability gaps, moral accountability gaps, public accountability gaps, and active responsibility gaps. These analyses provide essential foundations but assume a relatively direct causal chain from AI action to harmful outcome. In contrast, agentic workflows involve multiple causal nodes across extended temporal sequences, each potentially creating a gap in a different way.
Moral Crumple Zone. Elish (2019) introduced the concept of the “moral crumple zone” to describe how humans in highly automated systems may—either accidentally or intentionally—bear the brunt of moral and legal responsibility when the overall system malfunctions, even if they had limited control over the system’s behavior. Just as the crumple zone in a car absorbs the force of impact to protect passengers, the human operator absorbs responsibility to preserve the integrity of the technological system. Ayad and Plaks (2025) agree that moral wrongs committed by AI agents in close-knit human-agentic AI teams are judged more leniently than those of their human counterparts. This finding illuminates an important dynamic of responsibility misattribution. It focuses primarily on single-point attributions of failure in automated systems rather than on the temporal diffusion of responsibility across multi-step agentic workflows.
Problem of Many Hands. Thompson (1980) characterized the “problem of many hands” as the difficulty in assigning moral responsibility when many different individuals contribute in various ways to collective outcomes. Van de Poel et al. (2012) refined this idea, defining the problem as occurring when there is “a gap in a responsibility distribution in a collective setting that is morally problematic.” This framework addresses distributed responsibility in collective action but assumes human agents at each node. Agentic AI introduces hybrid responsibility chains that combine human decision points with AI-mediated actions in ways that neither purely individual nor purely collective frameworks adequately capture.
A related challenge arises from what might be called the problem of many things: the recognition that all technologies mediate human practices and shape outcomes in ways that complicate responsibility attribution (Verbeek, 2011). On this view, even conventional software artifacts bear traces of upstream decisions—when an engineering team incorporates pre-developed libraries, it imports potential bugs, design assumptions, and constraints that mediate the form and function of the resulting system. The problem of many things thus serves as a background condition for all three frameworks discussed above: responsibility gaps, moral crumple zones, and many-hands problems each arise, in part, because technologies are never neutral intermediaries but active mediators of human action. However, there is a crucial distinction between technology as mediator and technology as agent. In the mediating role described by Verbeek (2011), technologies shape and channel human action while humans remain the primary locus of decision-making. Agentic AI systems go beyond mediation: they do not merely shape the conditions under which humans act but instead take on delegated decision-making roles, pursuing goals through autonomous multi-step planning and execution. This shift from mediating to agentic technology is precisely what generates the distinctive responsibility challenges addressed in this article. While the problem of many things reminds us that all technological systems contribute to responsibility diffusion, agentic AI amplifies this diffusion qualitatively—not merely by mediating human action, but by inserting AI-driven decision nodes into causal chains that were previously occupied exclusively by human agents.
Figure 1 illustrates the theoretical positioning of these frameworks. As the figure shows, existing accounts address either single-point human-AI interactions (responsibility gap, moral crumple zone) or distributed human-only collectives (problem of many hands). The responsibility cascade concept addresses the upper-right quadrant: sequential workflows involving human-AI hybrid decision chains—a configuration increasingly common in agentic AI deployments but not adequately captured by prior frameworks. The horizontal axis represents causal structure, from single-point events to sequential, distributed processes. The vertical axis represents agent composition, from purely human collectives to human-AI hybrid systems. The responsibility cascade concept (highlighted) occupies a novel position addressing sequential workflows involving human-AI collaboration—a configuration not adequately addressed by existing frameworks. Arrows indicate intellectual lineage: the responsibility cascade synthesizes insights from the responsibility gap (Matthias, 2004), moral crumple zone (Elish, 2019), and problem of many hands (Thompson, 1980) while addressing the distinctive features of agentic AI workflows.
1.3 Thesis and Contribution
Agentic AI systems create responsibility cascades—temporal structures in which moral responsibility for outcomes becomes increasingly diluted across sequential workflow steps, leaving new gaps that existing frameworks cannot fully identify or address. The article makes three main contributions. First, it introduces the concept of the responsibility cascade as a distinct phenomenon that requires philosophical analysis, setting it apart from static responsibility gaps and single-point moral crumple zones. Second, it develops a Sequential Responsibility Attribution Framework (SRAF) that identifies four modes of responsibility distribution in agentic workflows—initiation responsibility, checkpoint responsibility, delegation responsibility, and outcome responsibility—and analyzes specific mechanisms by which responsibility cascades occur (the full development and diagnostic application of SRAF is presented in Sect. 4). Third, it offers normative guidance for designing responsibility-preserving agentic systems that maintain meaningful human control without creating moral crumple zones. This article thus explores both a philosophical question—how moral responsibility is distributed across sequential agentic workflows—and a governance question—how systems, organizations, and regulatory frameworks should be designed to ensure meaningful accountability. While the philosophical analysis provides the conceptual foundation, the governance implications translate that analysis into actionable insights for policymakers, organizational leaders, and system designers navigating the rapid deployment of agentic AI.
2 Conceptual Foundations
2.1 Agency, Control, and Moral Responsibility
The philosophical literature on moral responsibility has long recognized that responsibility attributions require some form of control over outcomes. Among the most influential accounts of this relationship, Fischer and Ravizza (1998) developed a comprehensive theory of moral responsibility grounded in the concept of reasons-responsiveness, which has become a central reference point in contemporary debates on free will and moral accountability. Their framework, later synthesized and updated by Talbert (2026), offers a pragmatically useful account because it shifts the focus from metaphysical questions about determinism to the practical conditions under which agents can be held responsible. We adopt their framework here not to endorse a particular position in the free will debate, but because its emphasis on control mechanisms is especially well suited to analyzing the distributed and sequential character of human-AI collaboration. Fischer and Ravizza (1998) and, more recently,Talbert (2026) distinguish between regulative control—the ability to do otherwise—and guidance control—acting on one’s own reasons-responsive mechanism. They argue that moral responsibility requires guidance control rather than regulative control: an agent is morally responsible for an action insofar as the mechanism that produces the action is the agent’s own and is appropriately responsive to reasons.
In practical terms, this distinction shifts the focus from whether an agent “could have done otherwise” to whether the agent acted on their own reasoning in a way that could respond to relevant considerations. A project manager who delegates a task to an AI system exercises guidance control at the time of delegation. However, as the AI executes the task through multiple steps, the question becomes whether subsequent outcomes can still be attributed to the manager’s guidance control, or whether the AI’s intermediate decisions have created new causal pathways that the manager neither controlled nor could have anticipated. This framework proves illuminating for the agentic AI context. When humans initiate an agentic workflow, they exercise guidance control at the outset—defining goals based on their own reasons-responsive deliberation. As the workflow advances through sequential AI-mediated steps, however, the human’s guidance control gradually diminishes. The mechanisms responsible for subsequent actions increasingly become those of the AI system rather than the human’s. Fischer and Ravizza (1998) and, more recently, Talbert (2026) emphasize that responsibility requires ownership of the mechanism on which one acts. In agentic workflows, a key question arises: Can a human “own” the outcomes of an AI mechanism that they initiated but did not control through intermediate steps? Traditional ownership principles, developed for human action and human-to-human delegation, need significant revision for hybrid human-AI action chains.
2.2 Meaningful Human Control
Research on human-AI teaming has established meaningful human control (MHC) as a key requirement for deploying autonomous systems in morally significant areas (Santoni de Sio & van den Hoven, 2018).Van Diggelen et al. (2018) suggest that MHC involves: (1) human operators making informed, conscious decisions about technology deployment; (2) operators having enough information to evaluate and verify system actions in context; (3) thorough testing before deployment and proper training for operators; and (4) clear links between technology actions and humans who are aware of their moral responsibility.
Van den Bosch et al. (2025) empirically studied MHC in human-AI teams across different collaboration designs, identifying three measurable components: subjective control (the human’s sense of control), normative control (alignment with ethical standards), and moral control (acceptance of responsibility for outcomes). Their results show that when AI systems make autonomous decisions without human intervention at decision points, human operators feel less responsible, even if the outcomes align with their moral values. These findings indicate that the design of human-AI collaboration directly affects the attribution of responsibility. Reducing human oversight in agentic workflows can boost efficiency but may undermine clear moral accountability.
Lange et al. (2025) suggest that accountability is crucial to human-AI agent relationships and propose a framework that conditions AI agent engagement on appropriate user behavior. This approach incorporates specific design strategies such as distancing, disengaging, and discouraging certain interactions to ensure the AI aligns with both user and societal interests. Venkatesh and Jayavardhan (2025) discovered that as AI systems become more autonomous, organizations need to prioritize transparency to build and sustain trust, accountability, and human oversight. To address security risks like data misuse, cyber threats, and system manipulation, implementing strong ethical frameworks and technical safeguards is essential. However, their study’s respondents believed that bias, job displacement, and human dignity were less critical and could be managed through proper governance and re-skilling programs. Salehi et al. (2025) arrive at a similar conclusion; their proposed frameworks for emerging multi-agent radiological workflows emphasize the need to responsibly preserve diagnostic innovation while ensuring medical transparency and accountability. However, the fact that MHC exists does not guarantee that it is exercised. Van Der Waa et al. (2021) found that humans under time pressure might trust the agent too much and ignore its explanations designed to prevent such over-trust. These findings suggest that the design of human-AI collaboration patterns has direct implications for responsibility attribution. Minimizing human oversight in agentic workflows can enhance efficiency at the cost of clear moral accountability.
The idea of meaningful human control has gained attention beyond academic philosophy. The European Union’s AI Act requires “human oversight” for high-risk AI systems, mandating that such systems be designed to allow human overseers to “correctly interpret” system outputs and to “decide, in any particular situation, not to use the system or to intervene on the operation of the system” (Article 14). The United States’ NIST AI Risk Management Framework also stresses human oversight as a core governance element, advocating for “meaningful human involvement” proportional to AI system risk levels. However, these regulatory frameworks provide limited guidance on how oversight should be structured within sequential workflows, or how responsibility should be assigned when oversight falls short despite good-faith compliance with oversight requirements. The responsibility cascade concept specifically addresses this gap by offering a framework for understanding why certain oversight structures succeed or fail in establishing genuine accountability.
2.3 Delegation and Hybrid Agency Chains
The relationship between a human and an agentic AI system during workflow execution can be described as one of delegation. The human grants execution authority to the AI while (ostensibly) maintaining supervisory responsibility. However, delegation to AI systems differs in important ways from delegation to human subordinates.
When a principal delegates to a human agent, the agent brings their own moral agency to the task. The agent can exercise moral judgment at decision points, recognize and respond to morally salient features of situations, and bear independent moral responsibility for their contributions. The principal’s responsibility is mediated by, but not exhausted by, their choice of agent and task specifications.
When a principal delegates to an AI system, no such moral agency is involved. The AI cannot exercise genuine moral judgment, cannot recognize moral salience as such, and cannot bear moral responsibility. The entire responsibility for outcomes must be shared among human actors—but the structure of delegation may make it unclear how this responsibility should be divided.
List (2021) has examined parallels between artificial agents and collective agents, arguing that under certain conditions, both might qualify as responsible moral agents. However, as List notes, current AI systems do not meet the criteria for moral agency—they lack the capacity for moral understanding, a consistent identity that grounds ownership of actions, and responsiveness to moral reasons that responsibility requires. For this analysis, we assume that AI systems cannot bear moral responsibility themselves, meaning that all responsibility for agentic workflow outcomes must be assigned to human actors.
This creates what I call a delegated agency chain: a series of actions where human decisions alternate with AI-mediated execution phases, forming a structure in which responsibility must somehow flow from human initiators through AI intermediaries to the final outcomes.
3 Responsibility Cascade
3.1 Temporality of Agentic Workflows
Agentic AI workflows have a unique temporal structure that sets them apart from both single AI decisions and extended human deliberative processes. A typical agentic workflow progresses through several phases:
Initiation Phase: A human defines a high-level goal or task, providing parameters and constraints. Humans exercise relatively strong guidance control at this stage, although their influence over downstream outcomes is prospective rather than actual.
Planning Phase: The AI system breaks down the goal into subtasks, develops an execution plan, and may present the plan to the human for approval. The human might have checkpoint control but limited ability to assess all implications of the plan.
Execution Phase: The AI system carries out subtasks sequentially, potentially interacting with external systems, gathering information, and making intermediate decisions. Human oversight during these phases can vary from continuous supervision to no oversight at all.
Checkpoint Phase: At specific points, the human may be asked to approve, modify, or stop the workflow. The human controls at these moments, but often with incomplete information about previous phases and limited foresight into upcoming ones.
Completion Phase: The workflow concludes with an outcome, which could be an artifact, a recommendation, a series of environmental changes, or a failure.
This framework creates conditions for a responsibility cascade: the human’s control, while present during initiation and checkpoints, is gradually mediated by AI decisions that restrict the set of possibilities and establish path dependencies that the human might not fully understand.
3.2 Responsibility Dilution
Four distinct mechanisms contribute to responsibility dilution in agentic workflows:
Epistemic Diffusion. Moral responsibility traditionally requires epistemic conditions. The agent must have known, or should have known, the nature and consequences of their action. As an agentic workflow progresses, the human’s epistemic access to what the AI is doing and why gradually diminishes. The human initiator specifies a goal but may not understand how the AI interprets it. The checkpoint approver sees the proposed actions but may not fully grasp their implications. At each step, the epistemic conditions for robust responsibility attribution become more difficult to meet. Consider a concrete example: when a financial analyst instructs an AI agent to “identify acquisition targets matching our strategic criteria,” the analyst understands the stated criteria but cannot observe how the AI interprets concepts like “strategic fit” or “market positioning” when filtering thousands of potential targets. By the time the analyst reviews the shortlist, the epistemic conditions for full responsibility—knowing what was excluded and why—cannot be satisfied.
Control Attenuation. Fischer and Ravizza (1998) emphasize that responsible action requires that the relevant mechanisms be the agent’s own. In agentic workflows, the human’s mechanism (their reason-responsive deliberation) initiates the workflow, but the AI’s mechanisms govern subsequent phases. Human control over outcomes diminishes as the workflow proceeds, raising questions about ownership of the process that produces outcomes.
Causal Opacity. As the causal chain from human initiation to outcome lengthens, the link between any specific human decision and the outcome becomes more indirect. When harm occurs, multiple human decision points (such as initiation, checkpoint approvals, and deployment authorization) are causally involved, but none close enough to be primarily responsible.
Temporal Dispersion. Human decisions that shape an agentic workflow occur at different times and may be made by various people in different roles. The person who authorized AI deployment might be different from the person who started this workflow, and both may differ from those who approved checkpoints. Responsibility then becomes spread across time and individuals without clear principles for assigning accountability. In enterprise settings, temporal dispersion often aligns with organizational hierarchies: the Chief Technology Officer who approved agentic AI adoption operates at a different level and timeframe than the analyst who initiated a specific workflow, who differs from the compliance officer who approved a checkpoint three days later. Responsibility chains thus span not only time but also organizational structure, creating attribution challenges that neither individual nor organizational frameworks can easily resolve.
3.3 Cascade Patterns
These mechanisms give rise to four characteristic responsibility cascade patterns:
Dilution Cascade. As workflow length increases, responsibility at each node decreases, potentially summing to less than full responsibility for the outcome. Consider a workflow with five sequential phases, each involving some human decision. If responsibility at each node is assessed at 20% (given limited control and epistemic access), the total assigned responsibility is 100%. However, if responsibility at each node is evaluated at 15% (even with more limited control), the total assigned responsibility is only 75%, leaving a genuine gap.
Checkpoint Accumulation. Each checkpoint adds a new responsible party, but none bears primary responsibility. Checkpoint approvers share responsibility for what they approved, but each can point to limited information and time pressure. Responsibility accumulates horizontally, with no node clearly accountable.
Crumple Cascade. Building on Elish’s (2019) concept, responsibility cascades to the most proximate human actor, regardless of their actual control. The checkpoint approver nearest to the harmful outcome bears disproportionate responsibility, which is most evident in the causal chain, even if earlier initiators and deployers were more causally and morally significant.
Delegation Vacuum. The organization authorized the agentic system, but no individual human exercised sufficient control at any point to bear traditional responsibility. Organizational responsibility exists but is difficult to translate into personal accountability.
4 Sequential Responsibility Attribution Framework (SRAF)
4.1 Four Modes of Responsibility
To address responsibility cascades, we propose distinguishing four modes of responsibility distribution that may be operative at different points in an agentic workflow:
Mode 1: Initiation Responsibility. This mode relates to the human who specifies the goal and starts the agentic workflow. Initiation responsibility is similar to principal responsibility in principal-agent relationships—the initiator is responsible for setting the system in motion with specific objectives and parameters. The scope of this responsibility covers the foreseeable consequences of reasonable goal setting. The initiator is accountable for harms that a reasonable person in their position should have anticipated might result from pursuing the goal through agentic means. However, as workflow complexity increases, initiation responsibility decreases—the more indirect the path from goal to outcome, the less responsible the initiator is for that outcome.
Mode 2: Checkpoint Responsibility. This involves humans who approve, modify, or stop AI-proposed actions at intermediate stages. Checkpoint responsibility is a distinct form created by human-in-the-loop designs in agentic systems. Checkpoint approvers bear responsibility proportional to their knowledge, access, and decision-making power at that point. This creates an epistemic burden: approvers must have enough information to make meaningful judgments; otherwise, their approval does not generate strong responsibility. Responsibility at checkpoints also depends on time constraints: when real-world constraints limit deliberation time, the responsibility for approval is reduced.
Mode 3: Delegation Responsibility. This relates to the organizational and systemic decisions to deploy agentic AI for specific tasks. Delegation responsibility involves meta-level accountability for establishing conditions in which responsibility cascades can occur. It is divided into design responsibility (accountability for how the system was designed to distribute human oversight) and deployment responsibility (accountability for whether the system was used in appropriate contexts). Those who design poor checkpoint systems or deploy agentic systems in domains where responsibility gaps exist are responsible for any resulting harms.
Mode 4: Outcome Responsibility. This refers to the traditional backward-looking responsibility for the results of actions. In agentic workflows, outcome responsibility should be shared across all previous modes rather than being concentrated in one place. The main challenge is the aggregation problem: how should responsibility be combined across multiple modes and actors? We suggest that outcome responsibility should be allocated based on proportional contribution: each actor’s share equals their causal role in the outcome, weighted by their access to knowledge and control at their intervention point.
4.2 Diagnostic Application
SRAF offers a method to analyze responsibility distribution in specific agentic workflows:
Step 1: Map Workflow Structure. Identify all phases, initiation, planning, execution, and checkpoints, and all decision points with involved humans.
Step 2: Identify Responsibility Modes. For each decision point, determine which responsibility modes apply. One person can be involved in multiple modes (e.g., someone who both started the workflow and approved a checkpoint).
Step 3: Assess Knowledge and Control. For each node, evaluate the level of knowledge (did the human know or should they have known relevant facts?) and control (was the action driven by the human or AI?).
Step 4: Diagnose Cascade Patterns. Identify which cascade patterns, if any, are present in the workflow. Are there dilution effects? Checkpoint accumulation? Crumple dynamics? Delegation vacuums?
Step 5: Identify Gaps. Determine whether total attributable responsibility falls short of what the outcome warrants. If so, identify where in the workflow structure the gaps arise.
5 Case Analysis
5.1 Case 1: Agentic Code Generation and Deployment
An agentic coding assistant receives the instruction “build a customer analytics data pipeline and deploy to production.” The system generates a multi-step plan, writes code across multiple files, tests iteratively, and deploys to a production environment—with human approval needed only at a final “merge” checkpoint. The deployed code contains a security vulnerability that causes a data breach.
Workflow Mapping: Initiation (task specification) → Planning (architectural design) → Execution (multiple coding, testing, revision phases) → Checkpoint (merge review) → Deployment.
Responsibility Analysis: The developer specified a functional goal without explicit security requirements. They bear some responsibility for not constraining the task appropriately, but the omission was reasonable given that security should be implicit in production code.
Checkpoint Responsibility: The reviewer who approved the merge had limited ability to identify the security vulnerability through code review of AI-generated code. The epistemic burden necessary to establish robust checkpoint responsibility was greater than what was feasible given the time constraints and AI’s opacity.
Delegation Responsibility: The organization’s decision to allow agentic code generation for production systems without security-specific checkpoints creates conditions for responsibility cascade.
Outcome Responsibility: The organization bears significant outcome responsibility through delegation responsibility, but individual actors have reduced responsibility due to epistemic and control limitations.
Cascade Pattern: The checkpoint reviewer is most visible in the causal chain and faces disproportionate scrutiny despite having lower control than the workflow structure indicates.
Governance Implications: Software liability frameworks traditionally assign responsibility to developers who write code and organizations that deploy it. When AI agents generate and deploy code with human oversight limited to merge approval, this traditional mapping breaks down. The human reviewer cannot practically verify the security properties of AI-generated code through standard review processes—the epistemic burden exceeds what is feasible. Emerging “secure by design” regulatory requirements, such as those proposed by CISA in the United States, may need to specify distinct review standards for AI-generated code, potentially including automated security scanning requirements that do not depend on human code review for assurance.
5.2 Case 2: Autonomous Vehicle Navigation and Collision
A Level 4 autonomous vehicle (robotaxi) operated by a technology company is transporting a passenger through an urban environment. The vehicle’s AI system continuously makes navigation decisions—route selection, speed adjustment, lane changes, and responses to other road users—without human driver intervention. The passenger has no control over driving operations. During the journey, the vehicle’s sensor system misclassified a partially obscured pedestrian behind a parked vehicle, and the AI’s trajectory planning algorithm failed to initiate timely braking, resulting in a collision that injured the pedestrian.
Workflow Mapping: Deployment approval (company decision) → Passenger activation (hail therobotaxi) → Continuous AI navigation (multiple perception-planning-action cycles) → No real-time human checkpoints → Result (pedestrian injury).
Responsibility Analysis: This case illustrates the responsibility cascade in its most condensed form—a continuous AI decision-making process with no human checkpoint opportunities between deployment and outcome.
Initiation Responsibility: The passenger who hailed the robotaxi exercised minimal initiation by specifying only a destination. Unlike the goal specification in Cases 1 and 2, the passenger exercised no control over how the AI would pursue the objective. The passenger’s responsibility for initiation is therefore minimal.
Checkpoint Responsibility: There are no real-time checkpoints in Level 4 autonomous operation. The passenger cannot intervene in driving decisions; the company’s remote monitoring (if any) occurs after the fact. This lack of checkpoint responsibility creates a clear gap—no human has the opportunity to exercise judgment when harm is about to happen.
Delegation Responsibility: This responsibility is the heaviest. The company that deploys the autonomous vehicle is responsible for: (a) the design of the AI system, including sensor capabilities and decision algorithms; (b) the decision to operate without human safety drivers; (c) the selection of operational domains (such as urban areas with pedestrians); and (d) the testing and validation procedures that certify the system as safe for deployment. Hevelke and Nida-Rümelin (2015) argue that when autonomous vehicles operate without meaningful human intervention capability, responsibility largely shifts to manufacturers and deployers.
Outcome Responsibility: The pedestrian’s injury requires attribution, but the cascade pattern makes attribution contested (Geisslinger et al., 2021; Himmelreich, 2018; McManus & Rutchick, 2019). The passengers cannot bear responsibility for the outcome proportional to the harm (they had no control). The AI system cannot bear moral responsibility (it lacks moral agency). The company bears outcome responsibility through delegation, but this creates accountability challenges: which individuals within the company are responsible?
Cascade Pattern: Delegation vacuum combined with the lack of checkpoints. The ongoing nature of AI decision-making removes the need for checkpoints, placing all responsibility on the delegation level. However, it becomes difficult to assign organizational responsibility to specific individuals, creating conditions for a moral crumple zone.
Distinctive Features: This case differs from Cases 1 and 2 in ways that highlight the responsibility cascade concept. First, temporal compression: where agentic workflows in knowledge work take hours or days with distinct phases, autonomous vehicle decision-making happens in milliseconds with ongoing action. The responsibility cascade doesn’t operate across separate phases but within a single continuous process. Second, minimal initiation: the passenger only indicates a destination, passing the entire decision-making process to the AI. This removes the responsibility for initiation that partly grounds accountability in other agentic contexts. Nyholm and Smids (2016) argue that passengers in autonomous vehicles cannot be held responsible in the same way as drivers, since they lack the control condition necessary for traditional responsibility attribution. Third, physical harm: unlike financial or project-management harms in other cases, AV collisions cause bodily injury, raising the moral stakes of attribution failures. Zhai et al. (2024) find that people attribute more responsibility to autonomous vehicle users than to passengers in conventional taxis, even when the harm is identical, suggesting that public intuitions may not align with control-based philosophical frameworks.
Governance Implications: The autonomous vehicle case highlights why regulatory frameworks have struggled to address Level 4 and higher levels of autonomy. The lack of real-time human checkpoints focuses all responsibility at the delegation level—the organizational decisions to design, validate, and deploy the system. This focus supports regulatory approaches that hold deploying organizations strictly liable while requiring thorough safety validation before deployment. Essentially, when checkpoint responsibility is structurally unavailable due to system architecture, governance frameworks must ensure that delegation responsibility is sufficiently robust to carry the full burden of accountability. This might involve safety case methodologies, third-party audits, and ongoing monitoring requirements that serve as alternatives to real-time human judgment during operation.
5.3 Case 3: Agentic Project Management
An agentic system managing a product development workflow receives authorization to “optimize team resource allocation to meet the Q3 delivery deadline.” The system analyzes team capacity, reassigns personnel between tasks, adjusts milestone timelines, and communicates schedule changes—with human oversight limited to weekly summary reviews. The project fails when critical dependencies are broken by resource reassignments the system makes to meet the stated deadline.
Workflow Mapping:Initiation (optimization goal) → Planning (resource analysis) → Execution (multiple reassignment and communication phases) → Checkpoints (weekly reviews) → Outcome (project failure).
Responsibility Analysis:
Initiation Responsibility: The manager who set the optimization goal is responsible for defining success narrowly in terms of meeting deadlines, without restricting dependency preservation. However, preserving dependencies should be an implicit part of competent project management—the AI’s failure to recognize this reflects system limitations rather than a misdefinition of the goal.
Checkpoint Responsibility: Weekly reviewers received summary information that was insufficient to identify growing dependency risks. Their checkpoint responsibility is limited by inadequate information provision.
Delegation Responsibility: The organization’s decision to implement agentic project management with insufficiently granular checkpoints created predictable risks of coordination failures.
Outcome Responsibility: Team members affected by AI-directed reassignments bear no responsibility—they acted on what appeared to be legitimate instructions. Responsibility focuses on delegation and initiation modes, but gaps emerge from reasonable reliance on AI competence.
Cascade Pattern: Checkpoint buildup combined with a delegation vacuum. Multiple weekly reviewers share responsibility for outcomes, but none have sufficient information to provide meaningful oversight.
Governance Implications: As organizations adopt agentic AI for workflow management—such as project coordination, resource allocation, scheduling, and communication—internal governance frameworks must address how to maintain accountability for AI-mediated decisions that impact employees and organizational outcomes. The checkpoint accumulation pattern in this context shows that diffuse periodic reviews may create an illusion of oversight without establishing real responsibility: each weekly reviewer may cite limited information and time pressure, and no one bears clear accountability for the cumulative effect of AI decisions over the review period. Organizations might need to redesign checkpoint structures based on decision criticality and irreversibility rather than arbitrary time intervals, ensuring that important resource decisions get focused human attention regardless of when they happen in a workflow.
5.4 Cross-Case Analysis
Table 1 shows how the responsibility cascade operates differently depending on the workflow structure.
6 Normative Implications
6.1 Design Principles for Responsibility-Preserving Agentic Systems
The SRAF analysis proposes several design principles for agentic systems aimed at maintaining meaningful responsibility attribution.
Transparency Principle. Each workflow step must provide enough explanation for humans at subsequent checkpoints to make informed judgments. This requires proactive communication of decision rationales in clear, understandable terms.
Reversibility Principle. Systems should be designed to allow human intervention, reversal, or redirection of workflows without significant cost. Irreversible actions need stricter checkpoint controls and cannot be delegated solely to AI execution phases.
Proportional Checkpointing Principle. Human checkpoints should be allocated based on the moral importance and irreversibility of subsequent actions. High-stakes decisions require more detailed human oversight, even if this reduces workflow efficiency.
Responsibility Accounting Principle. Systems should make the distribution of responsibility visible and traceable. Workflow logs should identify not only what decisions were made but also who was positioned to exercise control at each point and what information they had access to.
6.2 Toward Responsibility-Aware Agentic Architectures
These principles suggest concrete architectural features for responsibility-preserving agentic systems:
Explicit Responsibility. Workflow logs should include metadata identifying responsibility modes active at each step, the human actors involved, and the epistemic resources available to them.
Graduated Autonomy. System autonomy should be calibrated to responsibility stakes. Low-stakes, reversible actions may proceed with minimal human oversight; high-stakes or irreversible actions require robust checkpoint controls.
Epistemic Access. Checkpoints must be designed to provide genuine epistemic access, not merely the appearance of human approval. This may require pausing workflows to enable human investigation, providing comparative analysis rather than just recommendations, and flagging uncertainty rather than presenting false confidence.
Failure Mode Analysis. Before deployment, agentic systems should undergo responsibility failure mode analysis to identify how cascade patterns might emerge and what gaps might be created.
6.3 Implications for AI Governance
The SRAF framework proposes several principles for AI governance that go beyond the system-design recommendations outlined above.
Checkpoint Specification Requirements. Regulations requiring “human oversight” of AI systems should specify minimum standards for what constitutes a meaningful checkpoint. Token approval mechanisms that provide only the appearance of human control, confirmation buttons after a brief summary, and a weekly dashboard review of dozens of AI decisions should not satisfy regulatory oversight requirements. Governance frameworks might require that checkpoint approvers have access to specified categories of information (not merely AI-generated summaries but underlying data and reasoning), sufficient time for deliberation (proportional to decision stakes), and genuine authority to modify or terminate workflows without adverse consequences for exercising that authority. The responsibility cascade analysis suggests that checkpoints failing these conditions generate at most attenuated responsibility, which is insufficient to establish the accountability that oversight requirements aim to ensure.
Delegation Accountability Standards. Organizations deploying agentic AI should take greater responsibility to ensure that workflow designs maintain clear accountability. This implies a due diligence standard for delegation responsibilities: organizations should be required to demonstrate that they have prospectively analyzed potential responsibility cascades in their agentic deployments and put in place appropriate mitigation measures. Such analysis might be part of AI impact assessments that are increasingly mandated by governance frameworks. Organizations that deploy agentic systems without conducting such analysis or that design workflows with foreseeable responsibility gaps should bear increased liability for any resulting harms.
Outcome Documentation Requirements. When Agentic AI leads to negative outcomes, organizations should be required to produce responsibility attribution documentation that identifies the human actors involved at each workflow stage, the information available to them at relevant decision points, and the decisions they made. This documentation serves multiple purposes: it supports retrospective accountability by making responsibility distribution transparent, it enables organizational learning by identifying where cascade patterns emerged, and it provides evidence for regulatory enforcement. The SRAF diagnostic checklist (Appendix) offers a template for such documentation.
Proportionality Thresholds. Governance frameworks might establish tiered requirements based on the stakes of outcomes and reversibility. Low-stakes, easily reversible workflows—such as an AI assistant drafting an internal memo for human review—might proceed with minimal checkpointing requirements. High-stakes workflows that affect health, safety, financial security, or legal rights require robust human oversight regardless of efficiency costs. Irreversible actions—such as deploying code to production, executing financial transactions, or making binding commitments—should trigger enhanced checkpointing at the specific decision point, not just at workflow initiation or during periodic reviews. This proportionality principle aligns with risk-based approaches in the EU AI Act and NIST AI RMF, while offering more specific guidance on checkpoint design.
Prohibition of Responsibility-Eliminating Designs. Finally, governance frameworks might prohibit workflow designs that foreseeably eliminate human accountability for consequential outcomes. If an agentic workflow is structured such that no human actor can satisfy the epistemic and control conditions for meaningful responsibility, yet the workflow produces outcomes with significant moral stakes, the workflow design itself is defective from an accountability perspective. Organizations should not be permitted to deploy accountability-eliminating workflows merely because they are efficient; the social value of maintaining attributable responsibility constrains permissible automation designs. This prohibition embodies the core normative principle behind meaningful human control: humans must stay truly responsible for AI-influenced outcomes in morally important areas.
6.4 Implications for Meaningful Human Control
The SRAF framework enhances our understanding of meaningful human control in agentic AI contexts. MHC cannot be achieved through token human involvement—checkpoints that merely give the appearance of human control without genuine epistemic access and decision-making authority do not establish real control. Furthermore, MHC should be viewed as distributed across different workflow phases rather than located at any single point. Achieving meaningful control of an agentic workflow requires effective control at the start (with well-specified goals), at checkpoints (with genuine decision authority based on adequate information), and at the delegation level (matching AI capabilities to task requirements). This distributed view of MHC has implications for regulatory and governance frameworks. Organizations deploying agentic AI should be required to demonstrate how meaningful human control is maintained throughout all workflow phases, not just that a human was nominally involved at some point.
7 Potential Objections
7.1 “No Techno-Responsibility Gap”
Some philosophers argue that a genuine responsibility gap is not caused by AI systems (Tigard, 2021; Simpson & Müller, 2016). They believe moral responsibility is flexible enough to include emerging technological entities; we can always find responsible human actors if we look carefully. This objection challenges strong claims that AI creates responsibility where none previously existed. However, the responsibility cascade concept does not suggest that AI creates responsibility out of nothing. Instead, it states that agentic workflows cause practical dilution, which current attribution practices cannot fully address. Even if responsibility can technically be assigned, the way agentic workflows are structured might make assignments difficult, spread across many actors, and so out of sync with common judgments of blameworthiness that real accountability problems remain.
7.2 “Collective Agency Solves This”
Drawing on List (2021), one could argue that collective agency frameworks can address responsibility gaps in agentic workflows. If the organization deploying agentic AI acts as a collective agent, it can bear responsibility even when no individual does. This objection points to a valuable resource but does not fully resolve the responsibility cascade. First, collective responsibility does not eliminate the need to understand how responsibility is distributed among individuals. Organizations act through their members, and individual accountability remains morally important—both for the individuals themselves and for organizational governance. Assigning responsibility to “the organization” without understanding how that responsibility connects to individual decisions leaves key moral questions unanswered. Second, hybrid human-AI workflows do not fit neatly into existing collective agency models. List’s framework and related accounts assume groups of individual agents, each contributing their own agency to collective action. The AI system influences outcomes without contributing to an agency in the morally relevant sense. This creates attribution challenges that collective frameworks, designed for human groups, do not directly address. The responsibility cascade concept specifically targets this hybrid structure—the switch between human decision points and AI-mediated actions—that distinguishes agentic workflows from traditional collective actions.
7.3 “Just Design Better Checkpoints”
An engineering-focused response might argue that the responsibility cascade is simply a design issue. Therefore, checkpoints that provide real epistemic access would solve the problem. This objection is partly accurate. Better checkpoint design can reduce some cascade effects (Santoni de Sio & van den Hoven, 2018). However, adding more checkpoints can worsen the distribution of responsibility by further fragmenting it. Meaningful checkpoints require time and mental resources. Practical limits exist on how much human attention can be dedicated to AI oversight. The challenge is not just to design more checkpoints but to understand which ones truly generate responsibility and which only give the illusion of human control.
7.4 “Strict Liability Solves This”
Legal scholars and policymakers might argue that strict liability regimes effectively handle the responsibility cascade. Under strict liability, organizations that deploy AI are legally responsible for harms caused by AI, regardless of fault. If organizations know they will be held liable for outcomes generated by agency AI, they have proper incentives to design responsible systems, implement sufficient checkpoints, and avoid deploying AI in high-risk scenarios. From this perspective, the philosophical issue of moral attribution becomes mostly irrelevant—what is important is ensuring that someone with resources and motivation bears the consequences for harm.
This argument recognizes an important governance tool, but strict liability alone does not eliminate the need for responsibility analysis.
First, strict liability covers compensation for harm but not moral accountability in full. Knowing that an organization will pay damages or face penalties does not fulfill the moral obligation to determine who was responsible and why. Moral accountability involves more than just consequences; it requires understanding, acknowledgment, and the capacity for moral address, things that organizational liability alone cannot provide. The SRAF framework offers the conceptual tools for assigning responsibility to individuals, which remains morally important even when organizational liability ensures compensation.
Second, strict liability may produce harmful incentives around responsibility allocation. Organizations might set up human oversight to create liability shields—checkpoints designed to shift blame to individuals rather than foster genuine accountability. A checkbox approval process that is epistemically insufficient might still serve to deflect liability onto the approver if something goes wrong. SRAF helps identify when these structures create moral vulnerabilities instead of meaningful oversight.
Third, strict Liability does not provide guidance on internal organizational governance. Even under a strict liability regime, organizations must determine how to structure workflows, establish checkpoints, and assign responsibilities among employees. SRAF offers principles for these internal governance decisions that strict liability alone cannot provide. An organization that bears liability for AI outcomes still needs to decide how to effectively organize Responsibility and oversight. Designing responsibility-preserving workflows—the cascade patterns and mitigation strategies analyzed above—address this need. Fourth, strict liability regimes face their own attribution challenges in corporate contexts. When an organization bears liability, questions arise about how consequences should flow to individuals within the organization—executives, managers, designers, deployers, approvers. SRAF’s analysis of delegation responsibility and its relationship to design and deployment decisions informs how individual accountability might be structured within organizations facing strict liability for AI outcomes. Strict liability and responsibility analysis are thus complementary rather than substitutes. Strict liability ensures that harm does not go uncompensated; responsibility analysis ensures that accountability remains meaningful, that governance structures serve genuine oversight rather than blame-shifting, and that the moral dimensions of AI-mediated harm receive appropriate attention.
This raises a deeper question, however: if human-agentic AI teaming is organized exclusively according to the strict liability principle—which, as Hevelke and Nida-Rümelin (2015) suggest, may be the most socially appropriate framework for cases where passengers exercise no meaningful control—does the attribution of moral responsibility in the sense of blame or guilt become effectively empty? This concern parallels how we conceptualize systemic accidents: in complex sociotechnical systems, catastrophic failures often arise from the interaction of multiple components rather than from any identifiable individual’s negligence, and strict liability serves as the governance response precisely because individual blame attribution appears misplaced. If agentic AI workflows are understood as generating systemic-accident-like conditions—where harm emerges from complex interactions across human and AI decision nodes rather than from discrete individual failures—then one might conclude that moral responsibility in its blame-ascribing sense has indeed come to nothing. We resist this conclusion. While strict liability appropriately governs compensation and organizational incentives, the moral dimension of responsibility serves functions that strict liability cannot replace: it sustains the normative expectation that human agents remain attentive to the consequences of their decisions, it provides the basis for moral address between persons, and it underwrites the very rationale for designing systems that preserve meaningful human control. The responsibility cascade framework is intended, in part, to ensure that even within strict liability regimes, we retain the conceptual resources to identify where moral responsibility—including blame—is genuinely warranted and where its apparent absence signals a design deficiency rather than an inevitable feature of human-AI collaboration.
8 Conclusion
Agentic AI systems mark a significant shift in human-AI interaction, introducing new challenges for assigning moral responsibility. The responsibility cascade—the spread of responsibility over different stages of a workflow—is a unique phenomenon that current frameworks cannot fully capture. By distinguishing initiation, checkpoint, delegation, and outcome responsibilities, SRAF enables a more precise analysis of where responsibility lies within agentic workflows and where misattributions might occur. Identifying proper cascade patterns helps guide the design of agentic systems that maintain meaningful human accountability.
Several directions for future research arise from this analysis. Empirical studies on how humans perceive and assign responsibility in agentic workflows could test and improve the framework. Extending it to multi-agent systems, where AI agents delegate to other AI agents, would address emerging technological developments. Integrating with legal liability frameworks would relate philosophical insights to practical governance.
As agentic AI systems become more common in critical areas—such as healthcare, finance, project management, software development, and transportation—the importance of responsibility attribution has never been greater. We cannot and should not avoid deploying these systems; their benefits are too great, and the pressures to adopt them are too strong. Nonetheless, we can implement them in ways that uphold meaningful human accountability, ensure harm is not detached from responsibility, and protect individuals from becoming moral crumple zones for system failures beyond their control.
The responsibility cascade concept, as explored by philosophers, opens new pathways for analyzing moral responsibility in hybrid human-AI environments. It extends traditional frameworks to include the temporal and sequential aspects of agentic workflows. For policymakers, this suggests that governance structures requiring “human oversight” must focus on the quality, structure, and epistemic validity of that oversight—not merely its formal existence. Regulations should define checkpoint standards based on the stakes of decisions and ensure that oversight promotes genuine accountability rather than superficial compliance. For organizational leaders, SRAF provides a diagnostic tool to identify potential responsibility gaps in agentic deployments and how workflow design can help maintain attribution. For system designers, the framework encourages the development of architectural features—such as graduated autonomy calibrated to the stakes, checkpoint interfaces that support genuine epistemic access, and explicit responsibility metadata in workflow logs—that bolster human accountability rather than undermine it.
The responsibility cascade is not an unavoidable outcome of agentic AI but a design challenge that can be addressed through thoughtful governance at multiple levels: system architecture, organizational policy, and regulatory frameworks. Ultimately, resolving this challenge requires sustained interdisciplinary alignment. Future progress depends on bridging the gap between philosophical conceptualization, social-scientific empirical analysis, legal liability frameworks, technical systemic design, and policy-driven governance structures.
Data Availability
Not applicable.
References
Abou Ali, M., Dornaika, F., & Charafeddine, J. (2025). Agentic AI: A comprehensive survey of architectures, applications, and future directions. Artificial Intelligence Review, 59(1), 11. https://doi.org/10.1007/s10462-025-11422-4
Ayad, R., & Plaks, J. E. (2025). Attributions of intent and moral responsibility to AI agents. Computers in Human Behavior: Artificial Humans, 3, 100107. https://doi.org/10.1016/j.chbah.2024.100107
Bonnefon, J. F., Shariff, A., & Rahwan, I. (2016). The social dilemma of autonomous vehicles. Science, 352(6293), 1573–1576. https://doi.org/10.1126/science.aaf2654
Elish, M. C. (2019). Moral crumple zones: Cautionary tales in human-robot interaction. Engaging Science, Technology, and Society, 5, 40–60. https://doi.org/10.17351/ests2019.260
Fischer, J. M., & Ravizza, M. (1998). Responsibility and control: A theory of moral responsibility. Cambridge University Press.
Geisslinger, M., Poszler, F., Betz, J., Lütge, C., & Lienkamp, M. (2021). Autonomous driving ethics: From trolley problem to ethics of risk. Philosophy & Technology, 34(4), 1033–1055. https://doi.org/10.1007/s13347-021-00449-4
Hevelke, A., & Nida-Rümelin, J. (2015). Responsibility for crashes of autonomous vehicles: An ethical analysis. Science and Engineering Ethics, 21(3), 619–630. https://doi.org/10.1007/s11948-014-9565-5
Himmelreich, J. (2018). Never mind the trolley: The ethics of autonomous vehicles in mundane situations. Ethical Theory and Moral Practice, 21(3), 669–684. https://doi.org/10.1007/s10677-018-9896-4
Hughes, L., Dwivedi, Y. K., Li, K., Appanderanda, M., Al-Bashrawi, M. A., & Chae, I. (2025). AI agents and agentic systems: redefining global it management. Journal of Global Information Technology Management, 28(3), 175–185. https://doi.org/10.1080/1097198X.2025.2524286
Lange, B., Keeling, G., Manzini, A., & McCroskery, A. (2025). We need accountability in human–AI agent relationships. Npj Artificial Intelligence, 1(1), 38. https://doi.org/10.1038/s44387-025-00041-7
List, C. (2021). Group agency and artificial intelligence. Philosophy & Technology, 34(4), 1213–1242. https://doi.org/10.1007/s13347-021-00454-7
Matthias, A. (2004). The responsibility gap: Ascribing responsibility for the actions of learning automata. Ethics and Information Technology, 6(3), 175–183. https://doi.org/10.1007/s10676-004-3422-1
McManus, R. M., & Rutchick, A. M. (2019). Autonomous vehicles and the attribution of moral responsibility. Social Psychological and Personality Science, 10(3), 345–352. https://doi.org/10.1177/1948550618755875
Nyholm, S., & Smids, J. (2016). The ethics of accident-algorithms for self-driving cars: An applied trolley problem? Ethical Theory and Moral Practice, 19(5), 1275–1289. https://doi.org/10.1007/s10677-016-9745-2
Salehi, S., Singh, Y., Habibi, P., & Erickson, B. J. (2025). Beyond single systems: How multi-agent AI is reshaping ethics in radiology. Bioengineering, 12(10), 1100. https://doi.org/10.3390/bioengineering12101100
Santoni de Sio, F., & Mecacci, G. (2021). Four responsibility gaps with artificial intelligence: Why they matter and how to address them. Philosophy & Technology, 34(4), 1057–1084. https://doi.org/10.1007/s13347-021-00450-x
Santoni de Sio, F., & van den Hoven, J. (2018). Meaningful human control over autonomous systems: A philosophical account. Frontiers in Robotics and AI, 5, 15. https://doi.org/10.3389/frobt.2018.00015
Simpson, T. W., & Müller, V. C. (2016). Just war and robots’ killings. The Philosophical Quarterly, 66(263), 302–322. https://doi.org/10.1093/pq/pqv075
Talbert, M. Moral responsibility. In E. N. Zalta, & U. Nodelman (Eds.), The Stanford encyclopedia of philosophy (Summer 2026 edition). https://plato.stanford.edu/archives/sum2026/entries/moral-responsibility
Thompson, D. F. (1980). Moral responsibility of public officials: The problem of many hands. The American Political Science Review, 74(4), 905–916. https://doi.org/10.2307/1954312
Tigard, D. W. (2021). There is no techno-responsibility gap. Philosophy & Technology, 34(3), 589–607. https://doi.org/10.1007/s13347-020-00414-7
Van de Poel, I., Nihlén Fahlquist, J., Doorn, N., Zwart, S., & Royakkers, L. (2012). The problem of many hands: Climate change as an example. Science and Engineering Ethics, 18(1), 49–67. https://doi.org/10.1007/s11948-011-9276-0
Van den Bosch, K., Van Diggelen, J., Verdult, S., Haije, T., & Van der Waa, J. (2025). Measuring meaningful human control in human–AI teaming: Effects of team design in AI-assisted pandemic triage. AI and Ethics, 5(3), 3329–3353. https://doi.org/10.1007/s43681-024-00647-8
Van Der Waa, J., Verdult, S., Van Den Bosch, K., Van Diggelen, J., Haije, T., Van Der Stigchel, B., & Cocu, I. (2021). Moral decision making in human-agent teams: Human control and the role of explanations. Frontiers in Robotics and AI, 8, 640647. https://doi.org/10.3389/frobt.2021.640647
Van Diggelen, J., Neerincx, M., Peeters, M., & Schraagen, J. M. (2018). Developing effective and resilient human-agent teamwork using team design patterns. IEEE Intelligent Systems, 34(2), 15–24. https://doi.org/10.1109/MIS.2018.2886671
Venkatesh, A. N., & Jayavardhan, G. V. (2025). Future of work: Managing ethical challenges of Agentic AI and super intelligence in organizations. Advances in Consumer Research, 2, 5277–5284.
Verbeek, P. P. (2011). Moralizing technology: Understanding and designing the morality of things. University of Chicago Press.
Zhai, S., Wang, L., & Liu, P. (2024). Not in control, but liable? Attributing human responsibility for fully automated vehicle accidents. Engineering, 33, 121–132. https://doi.org/10.1016/j.eng.2023.10.008
Acknowledgements
None.
Funding
Open access funding provided by Northeastern University Library
Author information
Authors and Affiliations
Contributions
Single author.
Corresponding author
Ethics declarations
Ethics Approval
Not applicable.
Consent to Participate
Not applicable.
Consent for Publication
Not applicable.
Conflicts of Interest
The author declares no conflicts of interest.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Electronic supplementary material
Below is the link to the electronic supplementary material.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/
About this article
Cite this article
Kalluri, R. The Responsibility Cascade: Moral Attribution and Governance Challenges in Agentic AI Systems. Digit. Soc. 5, 47 (2026). https://doi.org/10.1007/s44206-026-00288-w
Received:
Accepted:
Published:
Version of record:
DOI: https://doi.org/10.1007/s44206-026-00288-w