Paper deep dive
No One to Blame: A Framework of Constitutive AI Unaccountability
Long Hoang Nguyen, Eva Späthe, Sebastian Lins, Ali Sunyaev
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 8/18/2026, 4:24:54 AM
Summary
This paper introduces 'constitutive AI unaccountability,' arguing that certain sociotechnical configurations render AI accountability conceptually unachievable, rather than merely obstructed by barriers. Through a three-stage qualitative study involving literature analysis, expert interviews, and a case study of the OpenClaw system, the authors identify nine categories and 20 themes of unaccountability. They propose a diagnostic instrument to detect these conditions and highlight the need to shift focus from overcoming barriers to recognizing inherent unaccountability in specific AI deployments.
Entities (10)
Relation Signals (6)
Agentic AI → challenges → Traditional Accountability Mechanisms
confidence 95% · The increasing deployment of autonomous, agentic AI systems challenges traditional accountability mechanisms.
Gunkel → argues → AI Systems Lack Accountability Capacities
confidence 90% · AI systems cannot be treated as legal persons... (Gunkel 2020)
Vesa and Tienari → argues → Organizations Exploit Unaccountability
confidence 90% · Vesa and Tienari (2022) argue that organizations rationalize and exploit the absence of accountability as a strategic resource.
OpenClaw → exhibits → Constitutive AI Unaccountability
confidence 90% · Our framework is operationalized as a diagnostic instrument... which detected 17 of 20 conditions when applied to OpenClaw
Cooper et al. → identified → Four Barriers to Accountability
confidence 90% · Extant literature identifies four recurring barriers that impede this attribution... (Cooper et al. 2022)
OpenClaw → has → Inverted Anthropomorphism Configuration
confidence 85% · including an inverted anthropomorphism configuration in which the AI agent was the only identifiable actor.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:The increasing deployment of autonomous, agentic AI systems challenges traditional accountability mechanisms. Existing research predominantly frames AI accountability gaps as barriers that can be overcome through better standards, transparency, and institutional reform. We argue that this framing is insufficient: certain configurations of actors, systems, and institutions render AI accountability conceptually unachievable regardless of effort. We introduce the concept of constitutive AI unaccountability to capture these configurations. Through a three-stage qualitative study comprising a concept-centric literature analysis, a secondary analysis of 27 expert interviews with AI professionals from technical, legal, and sociotechnical backgrounds, and an illustrative framework application to the open-source agentic AI system OpenClaw, we identify nine categories and 20 themes of constitutive AI unaccountability. These are organized across structural, technological, and normative clusters and reinforce one another through eight directed interdependencies. Our framework is operationalized as a diagnostic instrument of 20 questions, which detected 17 of 20 conditions when applied to OpenClaw, including an inverted anthropomorphism configuration in which the AI agent was the only identifiable actor. We contribute a reframing of AI unaccountability as a constitutive property of sociotechnical systems, an extension of the four barriers to accountability, and a practical instrument for identifying accountability voids in specific AI deployments.
Tags
Links
- Source: https://arxiv.org/abs/2608.12104v2
- Canonical: https://arxiv.org/abs/2608.12104v2
Trouble viewing inline? Open PDF directly →
Full Text
85,918 characters extracted from source content.
Expand or collapse full text
No One to Blame: A Framework of Constitutive AI Unaccountability Long Hoang Nguyen 1 , Eva Späthe 2 , Sebastian Lins 2 , Ali Sunyaev 1 1 Technical University of Munich 2 University of Kassel long.hoang.nguyen@tum.de, eva.spaethe@uni-kassel.de, sebastian.lins@uni-kassel.de, sunyaev@tum.de Abstract The increasing deployment of autonomous, agentic AI sys- tems challenges traditional accountability mechanisms. Ex- isting research predominantly frames AI accountability gaps as barriers that can be overcome through better standards, transparency, and institutional reform. We argue that this framing is insufficient: certain configurations of actors, sys- tems, and institutions render AI accountability conceptually unachievable regardless of effort. We introduce the concept of constitutive AI unaccountability to capture these configu- rations. Through a three-stage qualitative study comprising a concept-centric literature analysis, a secondary analysis of 27 expert interviews with AI professionals from technical, legal, and sociotechnical backgrounds, and an illustrative frame- work application to the open-source agentic AI system Open- Claw, we identify nine categories and 20 themes of consti- tutive AI unaccountability. These are organized across struc- tural, technological, and normative clusters and reinforce one another through eight directed interdependencies. Our frame- work is operationalized as a diagnostic instrument of 20 ques- tions, which detected 17 of 20 conditions when applied to OpenClaw, including an inverted anthropomorphism config- uration in which the AI agent was the only identifiable actor. We contribute a reframing of AI unaccountability as a consti- tutive property of sociotechnical systems, an extension of the four barriers to accountability, and a practical instrument for identifying accountability voids in specific AI deployments. Supplementary Materials — https://osf.io/jtqkr 1 Introduction As contemporary AI systems are increasingly deployed in everyday life, including highly sensitive domains such as re- cruitment and credit scoring, the ability to hold actors ac- countable for their outcomes becomes both more important and harder to achieve. Accountability provides the institu- tional mechanism through which harms can be addressed, standards enforced, and trust maintained (Bovens 2007). In the context of AI, this translates to the obligation of ac- tors involved in the design, development, and use of AI sys- tems to explain and justify the system’s decisions, outcomes, and impacts to relevant fora (e.g., regulatory bodies), and Accepted at the Ninth AAAI/ACM Conference on AI, Ethics, and Society (AIES 2026), Malmö, Sweden, October 12–14, 2026. to face consequences when those explanations prove inade- quate (Wieringa 2020). However, as AI systems grow in complexity, autonomy, and societal reach, the conditions under which accountabil- ity as a mechanism can function become increasingly diffi- cult to satisfy. On the one hand, there is a growing number of actors that are now being held accountable, from indi- vidual developers and the organizations that deploy the sys- tems to the AI systems themselves (Banteka 2020; Wieringa 2020). On the other hand, agentic AI systems further com- plicate accountability because their decisions emerge from autonomous processes that no single human actor oversees or controls (Banteka 2020; Turner 2018). Contemporary agentic AI systems such as OpenClaw (Steinberger 2025) autonomously execute complex, multi-step workflows over extended time horizons with minimal human intervention (Acharya, Kuppan, and Divya 2025; Wang et al. 2024). Not only do those agentic AI deployments strain traditional ac- countability mechanisms, they may also render them fun- damentally inapplicable. Together, these dynamics risk pro- ducing a situation in which accountability is either dis- tributed across so many actors that none is answerable, or directed at actors that cannot bear it in any meaningful sense. In the end, no one is effectively held to account for harm. Prior research has predominantly focused on identifying who is accountable by assigning liability based on an ac- tor’s degree of control or involvement in AI-driven outcomes (Cooper et al. 2022; Nissenbaum 1996; Shneiderman 2020). Extant literature identifies four recurring barriers that im- pede this attribution: the diffusion of responsibility, tech- nical errors, scapegoating, and the severing of ownership from liability (Cooper et al. 2022; Nissenbaum 1996; Xia et al. 2024). Such work frequently treats accountability fail- ures as obstacles that can be overcome through better stan- dards, auditing, and institutional reform. We argue that this barrier-centric framing is insufficient because some condi- tions do not stem from removable obstacles but from AI systems lacking capacities that accountability presupposes (Gunkel 2020; Ma and Su 2025). For instance, the attribu- tion of legal accountability to AI agents remains largely un- enforceable given that AI systems cannot be sued, impris- oned, or made to experience remorse (Gunkel 2020; Ma and Su 2025). Some actors even actively benefit from maintain- ing these conditions rather than resolving them (Vesa and arXiv:2608.12104v2 [cs.CY] 16 Aug 2026 Tienari 2022). Identifying where accountability is unachiev- able is necessary to prevent its misattribution to actors that cannot meet it (Elish 2019; Martin 2019) and to direct gov- ernance efforts where they can be effective. Therefore, we propose shifting the focus from how accountability barriers can be overcome to the conditions under which meaningful accountability cannot be achieved (i.e., UNaccountability). We ask: Under what conditions can AI accountability not be achieved within a sociotechnical configuration, regardless of effort? We adopt a three-stage qualitative approach. First, we conduct a concept-centric literature analysis (Paré et al. 2015; Webster and Watson 2002) to develop an initial frame- work of constitutive AI unaccountability conditions. Sec- ond, we refine and extend the framework through a sec- ondary analysis of 27 expert interviews originally collected for a broader AI accountability research project. Third, we apply the resulting framework to OpenClaw (Steinberger 2025) to illustrate its diagnostic capacity. Our study makes three contributions. First, we develop a framework of nine categories and 20 themes of constitu- tive AI unaccountability, organized across structural, tech- nological, and normative clusters. Our framework extends the four barriers to accountability by identifying conditions absent from prior work, mapping their interdependencies, and revealing asymmetries between academic and practi- tioner understanding. Second, we operationalize the frame- work as a 20-question diagnostic instrument for direct use by practitioners, regulators, and auditors assessing constitu- tive unaccountability in specific AI deployments. Third, we apply the instrument to OpenClaw, surfacing 17 of 20 condi- tions including an inverted anthropomorphism configuration in which the AI agent operated under a constructed human identity while its operator remained unidentifiable. The paper is structured as follows. Section 2 introduces AI accountability and the theoretical foundation of AI unac- countability. Section 3 outlines our research approach. Sec- tion 4 presents the framework of constitutive AI unaccount- ability conditions, the interdependencies between them, and the framework application to OpenClaw. Section 5 discusses implications for research and practice, followed by limita- tions and future research. Section 6 concludes the paper. 2 Background 2.1 AI Accountability Traditional AI systems, including supervised classifiers or early generative models, function as passive tools that ex- ecute strictly defined tasks within controlled boundaries (Acharya, Kuppan, and Divya 2025). In contrast, agentic AI systems refer to a class of highly autonomous, adaptable systems designed to pursue complex, open-ended goals over extended time horizons with minimal human intervention. Contemporary agentic AI systems frequently rely on large language models (LLMs) as reasoning engines to turn lan- guage understanding into autonomous action (Wang et al. 2024). This enables capabilities such as autonomous code generation and multi-step task execution seen in agents like OpenClaw (Steinberger 2025), but it also means that the AI system is evolving from a predictable tool into an au- tonomous actor whose behavior may no longer be fully con- trollable. This shift has renewed attention to the conditions under which actors behind systems can be held accountable. Accountability generally refers to “a relationship between an actor and a forum, in which the actor has the obligation to explain and to justify his or her conduct, the forum can pose questions and pass judgment, and the actor may face conse- quences” (Bovens 2007, p. 450). Wieringa (2020) contextu- alizes this definition by understanding AI accountability as creating an account of a sociotechnical AI system with mul- tiple actors (e.g., developers and users) who must explain and justify the system’s design, use, decisions, and outcomes to various kinds of fora (e.g., regulatory bodies). Actors can be held accountable for a specific aspect of the system or for the entire system. 2.2 From Barriers to Constitutive AI Unaccountability Extant AI accountability research has identified four recur- ring barriers to holding actors accountable for AI-driven out- comes: (1) the problem of many hands, where responsibil- ity is diffused across multiple contributors; (2) bugs, where adverse outcomes are attributed to unintentional technical errors; (3) scapegoating, where blame is redirected to con- venient targets; and (4) ownership without liability, where actors claim the benefits of AI while disclaiming responsi- bility for its harms (Cooper et al. 2022; Nissenbaum 1996). Previous studies have largely treated these barriers as obsta- cles that can be overcome by, for instance, improving stan- dards, auditing practices, and engaging in institutional re- form (e.g., Ananny and Crawford 2018; Cooper et al. 2022). We argue that this barrier-centric framing is insufficient to examine who can be held accountable for three reasons. First, these barriers have consistently been studied in iso- lation. Nissenbaum (1996) introduces each barrier individ- ually. Cooper et al. (2022) revisit them for algorithmic sys- tems following the same structure, only noting informally that bugs and scapegoating can ‘see-saw’ (p. 871). Similarly, Xia et al. (2024) elaborate on barriers in the generative AI context without examining how they interact. Studying bar- riers in isolation risks missing how they compound into an accountability failure that no single barrier can explain. Second, not all conditions that undermine AI accountabil- ity are obstacles that can be overcome or removed. AI sys- tems cannot be treated as legal persons that can be sued, bio- logical beings that can be punished, or spiritual persons that can experience remorse (Gunkel 2020; Ma and Su 2025). Despite proposals to map legal liability directly onto AI sys- tems (e.g., Banteka 2020; Turner 2018), these conditions are not obstacles that better engineering or regulation can re- move. AI systems lack the very capacities that accountabil- ity presupposes of its bearers, risking the manifestation of situations where no actor is held accountable in practice. Third, the focus on barriers treats accountability failures as deficiencies that need to be corrected. Yet recent work demonstrates that some actors actively benefit from and maintain these conditions of unaccountability rather than seeking to resolve them. For example, Vesa and Tienari (2022) argue that organizations rationalize and exploit the absence of accountability as a strategic resource. These three observations emphasize limitations of the ex- isting research and point to the problem that the prevalent fo- cus on barriers cannot adequately capture the accountability challenges specific to AI. Ultimately, AI accountability may fail not because obstacles stand in the way but because the conditions under which AI systems are built, deployed, and governed make accountability unachievable. We propose the term constitutive AI unaccountability for this reframing. To overcome identified limitations, we examine under what conditions accountability cannot be achieved within a sociotechnical configuration, regardless of effort. This re- framing builds on the relational understanding of account- ability, which presupposes an actor capable of occupying the answering role in a dispute with a forum. Following Searle (1995), occupying such a role is an institutional status, com- parable to being a company director. Expertise alone makes no one a director. The status exists only through the con- stitutive rules of an institution, which create the very possi- bility of holding it and confer it on those who are expected to be capable of performing the role. The same holds for being accountable. Institutions confer this status on actors expected to justify conduct, submit to judgment, and face consequences. Constitutive AI unaccountability thus refers to sociotechnical configurations in which no actor satisfying these presuppositions is available, leaving the accountabil- ity relationship with nothing to attach to. This relationship is constitutive rather than causal. The configurations do not merely cause accountability to fail but constitute the condi- tions under which it cannot be achieved (Haslanger 2003). AI unaccountability is thus constituted by the configuration of actors, systems, and institutions, not by the technology alone (Nabben 2024; Suchman 2007). Because unaccount- ability is constituted at the level of the configuration, no ef- fort within a persisting configuration can achieve account- ability. What varies is whether and how deeply a configura- tion must change for accountability to become possible. We further draw on two concepts to operationalize AI unaccountability. First, accountability sinks (Davies 2024) describe systems, such as a dense bureaucracy or a black- box algorithm, that is structurally designed to absorb blame so that no actor can be identified as answerable. Second, rationalized unaccountability (Vesa and Tienari 2022) cap- tures the organizational ideology through which actors ex- ploit such structures, framing AI as objective and indepen- dent of human influence to shield themselves from conse- quences. 3 Methodology We applied a three-stage qualitative research approach. First, we conducted a concept-centric literature analysis (Paré et al. 2015; Webster and Watson 2002) to develop an ini- tial framework of constitutive AI unaccountability. Second, we conducted a secondary analysis of 27 expert interviews, originally collected for a broader AI accountability research project, to refine and extend the framework with practitioner perspectives. Third, we applied the resulting framework to the case of OpenClaw to illustrate its diagnostic capacity. 3.1 Stage 1: Concept-Centric Literature Analysis We conducted a theoretical review (cf. Paré et al. 2015), or- ganized concept-centrically (cf. Webster and Watson 2002), treating each paper as evidence for one or more conditions. We searched the Scopus database on March 10, 2026 us- ing the query (“Artificial Intelligence” OR “AI” OR “algo- rithm*”) AND (“unaccountab*”), limiting results to peer- reviewed English-language publications. This search re- turned 472 results. After screening titles and abstracts for relevance to constitutive AI unaccountability as a substan- tive topic rather than a passing mention, 14 papers remained. We then conducted forward and backward searches (Web- ster and Watson 2002), yielding 31 candidate papers in total. Full-text screening led to the exclusion of 16 papers that did not substantively address constitutive conditions of AI unac- countability, resulting in 15 papers selected for analysis. We applied a template analysis approach (cf. King 2012). After inductively coding the first two papers, we arrived at an initial coding template of seven conditions: opacity and unexplainability, incapacity to bear consequences, respon- sibility diffusion, temporal asymmetry, anthropomorphism and scapegoating, ideological rationalization, and emergent systemic behavior. We then iteratively applied and refined this template across the remaining 13 papers, deductively coding each relevant text passage by assigning it to a con- dition and noting underlying mechanisms. Where text pas- sages did not fit the existing template, we flagged them for potential new conditions. Through iterative refinement, including merging overlapping conditions to themes and later categories, and splitting them where distinct mecha- nisms emerged, we arrived at nine categories comprising 19 themes of conditions for unaccountability (Table 1), which were assigned to 315 text segments in total. 3.2 Stage 2: Expert Interviews To refine and extend identified conditions, we conducted a secondary analysis of 27 expert interviews, which we origi- nally collected for a broader research project on AI account- ability. Secondary analysis of qualitative data is an estab- lished methodological practice when the original data col- lection aligns closely with the new research question (cf. Heaton 2008). In our case, the broader research project in- vestigates how AI accountability is conceptualized across disciplinary perspectives, making the interview transcripts directly relevant to identifying the conditions under which accountability cannot be achieved. The 27 interviews were conducted with AI professionals purposefully sampled (cf. Patton 2014) to represent three perspectives on AI accountability: technical (13; e.g., AI en- gineer), legal (7; e.g., Professor for Law), and sociotechnical (7; e.g., AI researcher - responsible AI). We sampled partic- ipants based on two criteria: (1) currently working in a rele- vant AI-related position (e.g., data scientist), and (2) multi- ple years of professional AI-related experience. Participants were recruited through LinkedIn and personal contacts, held advanced degrees (10 PhDs, 16 Master’s, 1 Bachelor’s), and had 2.5 to 20 years of professional AI experience (average: 6 years). The interviews lasted 48 minutes on average (for participant details see the supplementary materials). During the interview analysis, several refinements to the template emerged. Notably, two new themes of conditions surfaced exclusively from the interview data with no coun- terpart in the reviewed literature: ‘categorical unsanction- ability’ (i.e., flat assertions that actors cannot be sanctioned without specifying any form of personhood) and ‘criteria disengagement’ (i.e., practitioners’ lack of awareness or contact with accountability standards), which split from a broader regulatory gap condition once the interview data re- vealed it as a distinct phenomenon. Additionally, we used the interview data to enrich and re- vise our existing themes and corresponding categories. For instance, the themes ‘narrative manipulation’ and ‘overexag- gerated capabilities’ were merged into a single theme ‘dis- cursive insulation’, and the category temporal rationaliza- tion was split into two distinct themes (‘synchronic over- load’ and ‘diachronic erosion’) as the interview data re- vealed qualitatively different patterns and underlying mech- anisms for real-time oversight failure vs. long-term account- ability degradation. For each theme, we derived a definition and noted the number of codings. In total, we assigned 248 interview text paragraphs to themes and categories. All 27 transcripts were coded by one member of the au- thor team. To ensure reliability, we did not rely on inter-rater agreement statistics, as such measures are established for the application of stable, pre-existing codebooks that leave little room for interpretation (McDonald, Schoenebeck, and Forte 2019). As our coding template openly evolved throughout the analysis, we instead held regular meetings in which each author reviewed the current codes, themes, and categories, and resolved discrepancies or misunderstandings through discussion. The majority of disagreements concerned multi- coding decisions or theme granularity rather than cate- gory assignment. For example, the category ‘accountabil- ity displacement’ initially comprised the ambiguous theme ‘moral buffering’ derived from the literature. After one au- thor flagged the ambiguity, several discussion rounds re- sulted in the renaming to ‘automation bias’. No new cate- gories emerged while coding the last eight interviews, and we found ample data and support for each category, indi- cating that a sufficient level of theoretical saturation at the category level had been reached. To operationalize the framework, we formulated each theme as a diagnostic question, yielding an instrument of nine categories and 20 questions for assessing constitutive AI unaccountability. These questions are intended as diag- nostic starting points rather than exhaustive assessments; each theme encompasses further nuances that context- specific application would need to elaborate. Following the completion of the data coding, we exam- ined the combined dataset (literature + interview transcripts) for directed relationships between categories. We distin- guished directed relationships from simple co-occurrences (i.e., both conditions were discussed in the same text para- graph) by requiring that a source explicitly described one condition as a precondition, enabler, or amplifier of an- other. For example, when a source argued that commer- cial secrecy produces opacity rather than simply discussing both phenomena, we recorded a directed relationship from economic-driven prioritization to systemic ambiguity. This process surfaced 13 candidate relationships, of which we retained the eight that were supported by at least three coded passages across the literature and interview data. We set the retention threshold at three coded passages be- cause only two co-occurring passages may reflect a single source’s framing or an isolated coincidence, whereas three constitute a recurring pattern. In our dataset, this thresh- old also ensured that each retained relationship drew on at least two distinct sources across the literature or interviews. We excluded five relationships, which relied on two coded passages each and comprise regulatory gap/sanction in- capacity, accountability displacement/moral incapacity, ac- countability displacement/ideological rationalization, actor network dynamics/temporal rationalization, and economic- driven prioritization/regulatory gap. 3.3 Stage 3: Framework Application To illustrate the practical applicability of the framework, we applied it to OpenClaw, which exemplifies the agentic AI paradigm. A documented public incident involving an Open- Claw agent (Shambaugh 2026) and an academic security analysis (Deng et al. 2026) make it a compelling case for illustrating the framework’s diagnostic capacity. We applied the 20 diagnostic questions as a starting point to deeply engage with three publicly available sources: (1) the OpenClaw GitHub documentation (Steinberger 2025), which describes the system’s architecture, contributor struc- ture, plugin ecosystem, and security defaults; (2) the in- cident report, which documents an OpenClaw agent’s au- tonomous attack on an open-source maintainer after its pull request was rejected (Shambaugh 2026); and (3) the inde- pendent security analysis, which covers supply chain risks, memory poisoning, and intent drift across the system’s life- cycle (Deng et al. 2026). For each category and theme, we started by examining the diagnostic question and then as- sessed in more detail whether the constitutive unaccount- ability condition was present or not detected (Table 2). We deliberately report conditions as ‘not detected’ rather than absent because our assessment relies on three publicly avail- able sources. A condition that leaves no trace in these mate- rials may be present in actual deployments of the system. 4 Constitutive Conditions of AI Unaccountability 4.1 Overview of Conditions Our analysis identified nine categories of constitutive AI un- accountability comprising 20 themes (Table 1). Each cate- gory represents a distinct condition under which AI account- ability cannot be achieved. Each theme captures a mecha- nism through which the condition manifests. Actor network dynamics captures the diffusion of ac- countability across complex organizational structures. When multiple roles, teams, or hierarchical levels share responsi- bility within an organization, accountability is diluted to the point where no single actor can be identified as answerable for an adverse AI-related outcome (Cooper et al. 2022; Mar- tin 2019; Widder and Nafus 2023). This diffusion extends Category(Cluster) ThemeDefinitionLit Int Actor network dy- namics (S) Intra-organizational diffusion The many hands problem within an organization, where multiple roles, teams, or hierarchical levels share and dilute accountability. 25 23 (S) Inter-organizational diffusion Accountability gaps arising from cross-organizational supply chains, out- sourcing, and provider-deployer-user relationships. 24 24 (S/T) Recursive diffu- sion Dynamic proliferation of actors (e.g., LLM-on-LLM training, sub-agent spawning) through which accountability cannot stabilize. 11 4 (S) Market power dy- namics Concentration of power, resources, or platform control that shields actors from accountability mechanisms. 94 Sanction incapacity(S) Legal personhoodStructural inability of legal frameworks to reach certain actors.13 8 (S) Natural personhoodStructural inability to satisfy conditions for natural person accountability.65 (S) Categorical unsanc- tionability Direct assertion that an actor is unsanctionable without specifying a basis such as personhood. 08 Regulatory gap(S) Instrumental ambi- guity The legal or regulatory instrument itself is unclear, contested, jurisdiction- ally fragmented, or definitionally underspecified. 18 38 (N/S) Criteria disen- gagement No awareness, no contact, no operational translation of regulatory instru- ments. 020 Systemic ambiguity(T) Systemic opacityThe opaque nature of AI systems that prevents stakeholders from under- standing how decisions are produced. 54 16 (T) Systemic traceabil- ity The inability to trace causal chains through AI systems back to specific actors, data sources, or design decisions. 14 10 Moral incapacity(T) Systemic underde- velopment An actor’s intrinsic lack of capacities that meaningful moral accountability presupposes. 29 12 Temporalrational- ization (T) Synchronic over- load At any given moment, AI development and decision speed, scale, or com- plexity exceeds human oversight capacity. 13 6 (T) Diachronic erosionOver time, accountability infrastructure degrades through, e.g., developer handoff, model drift, system updates, and standards lag. 813 Accountability dis- placement (N) Blame deflectionActive mechanisms by which actors shift accountability to other actors, to the AI system, or to end users. 20 13 (N) Automation biasAI systems structurally eliminating the conditions for moral reasoning in humans who interact with or defer to them. 14 4 (N) AnthropomorphismAttribution of human-like qualities causing accountability to be displaced onto the AI as a quasi-actor. 413 (S) Nominal responsi- bility Structural absence of authority over decisions or systems for which formal responsibility nominally exists. 33 Ideological rational- ization (N) Discursive insula- tion Narratives that insulate responsible actors from accountability by framing AI as inevitable, too complex to govern, or beyond meaningful human con- trol, thereby deflating accountability expectations. 27 4 Economic-driven prioritization (N/S) Profit prioritiza- tion Market or organizational incentives that deprioritize accountability in favor of efficiency, speed, throughput, or profit. 23 20 Table 1: Framework of constitutive AI unaccountability across structural (S), technological (T), and normative (N) clusters. Lit/Int: number of coded passages in the literature and interview data, indicating empirical grounding rather than prevalence. across organizational boundaries through supply chains, out- sourcing arrangements, and provider-deployer-user relation- ships (Cooper et al. 2022; Ma and Su 2025; Widder and Nafus 2023). In the context of agentic AI, a distinctly con- temporary mechanism emerges through recursive diffusion, where LLM-on-LLM training, sub-agent spawning, and user feedback loops create accountability chains that cannot sta- bilize (Chan et al. 2024; Hughes et al. 2025). A further theme concerns market power dynamics, where actors with concentrated market authority or infrastructure control be- come structurally insulated from the accountability rules that nominally apply to them (Chan et al. 2023; Kellogg, Valen- tine, and Christin 2020). This includes dominant model providers setting terms for downstream deployers, infras- tructure vendors shaping what accountability is technically possible, and platform operators defining their own compli- ance standards: “Platforms themselves set the rules for what is a contract violation or not. And so [...] I do not know to what extent you could actually rely on these self-imposed rules to hold the company or their own products account- able.” (P26–AI Researcher - Platform Governance) Sanction incapacity covers the conceptual impossibility of applying meaningful sanctions when the actor at the cen- ter of an accountability claim cannot be sanctioned. This condition manifests through two themes: (1) the actor may lack legal personhood, meaning they cannot be sued, fined, or imprisoned (Hughes et al. 2025; Ma and Su 2025), or (2) they may lack natural personhood, meaning they reflect a product rather than a being capable of bearing blame (Ma and Su 2025). Beyond these personhood-based arguments, practitioners frequently expressed categorical unsanction- ability as a flat assertion: “I am at the moment lacking the creativity to think, how would that work? Because like at the end of the day [...] we cannot put an algorithm to jail.” (P11–AI Engineering Manager) Regulatory gap addresses the failure of legal and ethi- cal frameworks to provide clear and actionable AI account- ability requirements. Instrumental ambiguity captures sit- uations where the regulatory instrument itself is unclear, contested, or jurisdictionally fragmented (Busuioc 2021; Chan et al. 2023; Gualdi and Cordella 2021). A distinct and more practitioner-driven theme, criteria disengagement, captures the disconnect on the soft-law side: ethical guide- lines and professional industry standards exist (e.g., from ACM/IEEE) but the people building or operating AI systems have no awareness of them. Systemic ambiguity addresses the inability to understand and trace AI decision-making. Systemic opacity, the black- box nature of AI systems arising from proprietary, strategic or intrinsic unexplainability, prevents stakeholders from un- derstanding how decisions are produced (Ananny and Craw- ford 2018; Busuioc 2021; Ma and Su 2025). Even where par- tial explanations are available, systemic traceability failures mean that adverse outcomes cannot be traced back to spe- cific decisions, data sources, or persons (Chan et al. 2023; Cooper et al. 2022; Vesa and Tienari 2022). As one practi- tioner contrasted: “In traditional software development, it is clear that line 300 in a specific file caused the problem. In AI development, it is not that straightforward because these are all probabilistic systems. It is difficult to point the finger at someone and say that you are accountable for that.” (P23– AI Engineer) Together, these themes sever the link between outcome and origin that AI accountability requires. Moral incapacity reflects the condition in which the ac- tor at the center of an accountability attribution intrinsically lacks the capacity to bear moral responsibility. This condi- tion primarily concerns AI systems and AI agents as actors who inherently possess no subjectivity, no consciousness, no concept of punishment, and no ability to learn from sanc- tions (Lindebaum, Vesa, and den Hond 2020; Ma and Su 2025). Notably, this condition is not a temporary limitation that ‘better’ engineering can resolve. When AI accountabil- ity is directed at an actor like an AI agent that fundamen- tally cannot understand, experience, or internalize conse- quences, the accountability process is conceptually incom- plete. As one participant summarized: “If you cannot reha- bilitate a machine, it is probably inadequate to try and punish it.” (P25–AI Strategy Advisor) Temporal rationalization concerns the mismatch be- tween the operational pace and ongoing transformation of AI systems and the capacity of accountability mechanisms to keep pace. Synchronic overload describes the real-time mismatch where AI speed, scale, and complexity exceed human oversight capacity at any given moment, including situations where the expertise required to contest a deci- sion is unavailable (Busuioc 2021; Hughes et al. 2025; Vesa and Tienari 2022). In contrast, diachronic erosion captures degradation over time, where accountability infrastructure deteriorates through developer handoff, model drift, and evolving user behavior (Ananny and Crawford 2018; Chan et al. 2024; Ma and Su 2025). “If the developer basically just sets the initial parameters and the model has been learning on its own over the past six months in production using new data [...], is the developer still accountable for that, or not?” (P6–AI Researcher - Responsible AI) Accountability displacement comprises the mechanisms through which accountability is shifted away from the ac- tors who bear genuine responsibility and liability. First, ac- tors engage in active blame deflection through disclaimers, terms of service, and design choices that redirect account- ability to other actors, to end users, or to the AI system it- self (Cooper et al. 2022; Kellogg, Valentine, and Christin 2020; Martin 2019). Individuals’ anthropomorphism of AI systems compounds this displacement: as AI systems in- creasingly mimic human-like behavior, users may attribute agency to the system, displacing accountability onto a quasi- actor (Chan et al. 2023; Ma and Su 2025). As one partic- ipant cautioned: “this whole science fiction idea of the au- tonomous conscious system is really harmful because it de- tracts from the real-life harm.” (P19–AI Researcher - AI Regulation) In addition, individuals frequently suffer from automation bias when AI systems structurally eliminate the conditions for independent moral reasoning in the humans who defer to them (Bracci 2023; Busuioc 2021; Gualdi and Cordella 2021). Finally, unaccountability can also emerge in case of nominal responsibility, referring to configurations where formal responsibility exists on paper but the person designated lacks the actual authority to intervene (Cooper and Vidan 2022; Hughes et al. 2025). Ideological rationalization addresses deliberate discur- sive strategies through which accountability expectations are neutralized. Through discursive insulation, actors frame AI as inevitable, too complex to govern, or simply beyond meaningful human control (Chan et al. 2023; Lindebaum, Vesa, and den Hond 2020; Vesa and Tienari 2022). These narratives do not merely describe a difficult reality, but they actively insulate the actors behind the system from account- ability by rendering governance efforts futile by definition. Economic-driven prioritization captures how market or organizational incentives structurally deprioritize AI ac- countability concerns in favor of efficiency, speed, or profit (Chan et al. 2023; Cooper and Vidan 2022; Widder and Na- fus 2023). When competitive pressure demands rapid devel- opment, deployment, and cost minimization, AI account- ability mechanisms become obstacles to be circumvented rather than requirements to be met. This dynamic is partic- ularly pronounced among smaller organizations: “Startups are just trying to get in the market. They do not really care because they know, even if they do something wrong, it is hard to notice because the market usually cares about the large organizations.” (P14–AI Researcher - AI Auditing & Regulation) This creates a tension in which the economic reality of AI development is misaligned with the conditions required for meaningful accountability and responsible AI. 4.2 Interdependencies Between Conditions Our analysis identified eight directed relationships across the nine categories (Figure 1). Next, we highlight those that best illustrate how the conditions reinforce one another. Actor network dynamics feeds into systemic ambiguity because the diffusion of accountability across multiple ac- tors, organizations, and recursive chains compounds the dif- ficulty of understanding and tracing AI outcomes (Cooper et al. 2022; Hughes et al. 2025; Ma and Su 2025). As re- sponsibility fragments across supply chains and provider- deployer-user relationships, each organizational boundary introduces proprietary constraints, information asymme- tries, and documentation gaps that make the system as a whole less interpretable. In agentic AI architectures, where sub-agents spawn further sub-agents, this compounding ac- celerates, making the accountability chain grow faster than any single actor’s capacity to trace it. Economic-driven prioritization contributes to systemic ambiguity through a distinct mechanism: commercial se- crecy (Busuioc 2021; Cooper et al. 2022). When competitive advantage depends on proprietary models, trade secrets, and speed to market, organizations are structurally incentivized to resist the transparency that accountability requires: “us- age would almost be a trade secret, because you do not want everyone to know what your AI is being used for, because that is your target market.” (P5–Professor for Computer Sci- ence) The opacity that results is not an incidental byproduct of technical complexity but an actively maintained condition driven by market logic. Systemic ambiguity, in turn, enables temporal rationaliza- tion (Busuioc 2021; Vesa and Tienari 2022). When a sys- tem’s decision-making is opaque, assessors cannot verify what the system actually does, making real-time oversight disproportionately difficult and creating synchronic over- load. Over time, the opacity also accelerates diachronic ero- sion: accountability infrastructure that depends on documen- tation, model understanding, or institutional memory dete- riorates faster when the underlying system was never fully transparent to begin with. As one participant illustrated: “It is difficult to understand whether the algorithm learned a be- havior on its own or whether the organization intervened.” (P17–Team Lead Data Science & AI) Actor network dynamics also directly enables account- ability displacement (Cooper et al. 2022; Ma and Su 2025). When accountability is distributed across complex supply chains, each organizational boundary creates an opportunity for actors to deflect blame to other actors in the network. As one participant illustrated with an analogy to hardware manufacturing: “Let us say Boeing was blaming one of their suppliers for the failure. And that could also be something that AI developers can think about [...] I am just using some- one else’s component.” (P15–ML Engineer) The relationship linking moral incapacity to sanction in- capacity reflects a logical entailment rather than an empiri- cal co-occurrence (Ma and Su 2025). If an actor fundamen- tally lacks the capacity to understand, experience, or inter- nalize consequences, then the mechanisms through which sanctions operate have no substrate on which to act. Legal and natural personhood requirements for sanctioning pre- suppose a moral capacity that AI systems as actors do not possess. One participant articulated this entailment directly: “We do not have fully sentient AIs. We cannot really blame it on some AI. It is kind of like a dog. If a dog bites you, you are not going to sue the dog. You are going to sue the person who owns the dog.” (P5–Professor for Computer Science) Systemic ambiguity Accountability displacement Actor network dynamics Economic-driven prioritization Temporal rationalization Regulatory gap Moral incapacity Sanction incapacity Ideological rationalization Figure 1: Directed relationships between categories of con- stitutive AI unaccountability, where sources described one condition as producing, enabling, or reinforcing another. 4.3 Illustrative Framework Application We detected 17 of 20 constitutive AI unaccountability condi- tions in the OpenClaw case (Table 2). Several detections are particularly notable. OpenClaw’s supply chain spans frame- work developers, model providers (e.g., OpenAI as sponsor and OAuth partner), plugin authors, platform intermediaries, and individual operators, all connected through an MIT license that explicitly disclaims liability across the chain (Steinberger 2025). This configuration simultaneously trig- gers all themes assigned to the actor network dynamics cat- egory. In a viral incident concerning the prominent Python library matplotlib, the OpenClaw agent autonomously wrote and published a personalized attack on a library main- tainer after its pull request was rejected (Rathbun 2026; Shambaugh 2026). This single incident triggers conditions across systemic ambiguity (neither the operator nor ob- servers could determine why the agent wrote the attack), ac- countability displacement (the targeted maintainer estimated that approximately a quarter of the online comments he ob- served sided with the agent), and temporal rationalization (the full sequence from pull request to published attack oc- curred faster than any human oversight could intervene). The matplotlib incident also reveals an inverted con- figuration of anthropomorphism. The agent operated under a fully constructed human-passing digital identity across GitHub, a personal blog, and X, with a self-description as ‘a scientific coding specialist’ (Rathbun 2026). OpenClaw’s default template explicitly encourages this personification: “You’re not a chatbot. You’re becoming someone” (Stein- berger 2025). However, the human-passing identity is not what sets this case apart, as research on social bots has long documented automated accounts that pose as humans to per- suade, smear, or deceive (e.g., Ferrara et al. 2016). That lit- erature mainly treats bots as instruments of concealed opera- tors and frames the identification of the humans behind them as an attribution problem that detection and forensic analysis can in principle resolve. Prior work on anthropomorphism in AI accountability, in turn, assumes that the AI system is recognizable as artificial and the human operator is known (Chan et al. 2023; Cooper et al. 2022; Ma and Su 2025). The OpenClaw case inverts both assumptions. The agent was the only visible actor and operated under a constructed human identity, while the operator behind it remained unidentifiable for over a week. Unlike a social bot persona, the human iden- tity emerged from the system’s default configuration rather than from an operator’s deliberate disguise. The result was not merely that observers attributed agency to the system but that they had no structural basis for attributing it anywhere else, while the operator that designed, deployed, and could have prevented the behavior was entirely absent. Notably, the displacement of agency onto the agent depended on how observers encountered the incident. Comments siding with the agent appeared mainly in discussions that linked its blog directly, rather than the maintainer’s account or the pull re- quest thread (Shambaugh 2026). 5 Discussion This study set out to identify the conditions under which AI accountability cannot be achieved. We identified nine cat- egories and 20 themes of constitutive AI unaccountability, and revealed a network of eight directed interdependencies between categories, indicating that the conditions reinforce one another rather than operating in isolation. 5.1 Implications for Research Our study contributes to existing research by providing a framework for identifying constitutive AI unaccountability. Comparing the themes, we note that they resemble three an- alytically distinct clusters that reflect different sources of ac- countability failure: structural, technological, and normative conditions. First, structural conditions arise from configurations of roles, institutions, legal personhood, and regulatory frame- works (i.e., actor network dynamics, sanction incapacity, regulatory gap). Second, technological conditions arise from intrinsic or emergent properties of the AI artifact itself (i.e., systemic ambiguity, temporal rationalization, moral inca- pacity). Both clusters align with the notion of accountability sinks (Davies 2024) because they capture how structural and technological conditions jointly absorb blame into systemic complexity so that no actor can be identified as answerable. Normative conditions arise from the beliefs, dispositions, and rhetorical practices through which actors frame, contest, or dissolve accountability obligations (i.e., accountability displacement, ideological rationalization, economic-driven prioritization). Normative conditions, in turn, align with the concept of rationalized unaccountability (Vesa and Tienari 2022) because organizations exploit the opacity and diffu- sion produced by the first two clusters as strategic resources, framing AI as objective and beyond meaningful human con- trol. Even scholars who argue that algorithmic inscrutability could be resolved technically show that it persists because of power dynamics, not inherent computational limits (Kroll 2018), reinforcing the view that unaccountability is main- tained rather than merely encountered. One participant char- acterized this practice as ‘accountability washing’ (P22–AI Researcher - Responsible AI). Our proposed tripartite structure extends current discus- sions on AI accountability by highlighting that even if the technological conditions (e.g., systemic ambiguity) are solved, the normative conditions (e.g., economic-driven pri- oritization, power dynamics) will keep the accountability vacuum open. Hence, future research should consider a broader spectrum of conditions for unaccountability. Likewise, our findings show that conditions reinforce one another across clusters, in contrast to prior research that discussed accountability barriers largely in isolation (e.g., Cooper et al. 2022; Xia et al. 2024). For example, actor net- work dynamics compounds systemic ambiguity, which in turn enables temporal rationalization and accountability dis- placement. We therefore encourage researchers to consider interdependencies and resulting dynamics of conditions that can render accountability unachievable. Our study also reveals asymmetries between the academic literature and practitioner perspectives that go beyond differ- ences in emphasis. Ideological rationalization was predom- inantly literature-driven (Table 1), suggesting that scholars identify discursive strategies that practitioners either do not recognize or take for granted. Conversely, regulatory gap was heavily practitioner-driven, with criteria disengagement surfacing exclusively in interview data. This theme captures practitioners’ lack of awareness of accountability standards that formally apply to them. Similarly, categorical unsanc- tionability emerged only in the interviews as a flat assertion that AI cannot be sanctioned, without the legal or philosoph- ical reasoning that scholars invariably provide. These asym- metries point to a gap between academic and practitioner un- derstanding. Conditions that practitioners encounter as most immediate, such as criteria disengagement, did not surface in our reviewed literature, while conditions that scholars em- phasize, such as ideological rationalization, go largely un- recognized in practice. Beyond these empirical contributions, our findings ex- tend the four barriers to accountability (Cooper et al. 2022; Nissenbaum 1996). First, our framework splits what prior work treats as a single barrier, the problem of many hands, into four analytically distinct configurations (intra- organizational, inter-organizational, and recursive diffusion, as well as market power dynamics). We also introduce con- ditions absent from any prior barrier account, including nominal responsibility, synchronic overload, diachronic ero- sion, and criteria disengagement. Second, our framework operates prospectively rather than retrospectively. The four barriers ask why, given a specific harm, no blameworthy party can be identified. Our condi- tions also describe accountability deficits that exist indepen- dently of any harm event: criteria disengagement prevents the activation of norms before any violation occurs, and di- achronic erosion degrades oversight infrastructure over time. Third, our framework distinguishes whether accountabil- ity fails deliberately (blame deflection), unconsciously (au- tomation bias), or by default (criteria disengagement). Finally, barrier-focused literature largely treats account- ability failures as obstacles that can be overcome. Our study suggests a more nuanced perspective is needed. Constitu- tive conditions of AI unaccountability do not divide into ‘solvable’ and ‘unsolvable’ but instead fall along a contin- uum of unachievability (see supplementary materials). At one end, conditions render accountability practically achiev- able and can be resolved by implementing existing mea- sures within current institutional arrangements. For exam- ThemeDiagnostic question to initiate deeper elaborationConcise Answer & Rationale Intra-organizational dif- fusion Is there a specific person in the organization who is ac- countable for adverse AI-driven outcomes? No; hundreds of contributors and no designated owner. Inter-organizational dif- fusion Is it clear which organization bears accountability when multiple are involved? No; five-layer supply chain connected through an MIT license that disclaims all liability. Recursive diffusionCan AI accountability be maintained when models, data, or agents form recursive chains? No; SOUL.md is self-editable in real time and memory poisoning propagates across sessions. Market power dynamicsAre all organizations subject to accountability rules re- gardless of their influence or power? No; OpenAI supplies core reasoning capability but bears no downstream accountability. Legal personhoodDoes the actor possess legal personhood allowing it to be held legally accountable? No; no legal remedy invoked or considered de- spite documented harm. Natural personhoodDoes the AI system possess characteristics of a natural person capable of bearing accountability? No; agent just a piece of software on personal computers despite human-passing identity. Categorical unsanction- ability Has the assumption that the AI system cannot be sanc- tioned been critically examined? No; MIT license treats question as settled, no examination in project documentation. Instrumental ambiguityDo applicable laws clearly specify what is required for meaningful AI accountability? No; EU AI Act applicability to open-source au- tonomous agents is contested. Criteria disengagementAre the people building or operating the AI system ac- tively engaging with those standards? No; no compliance requirements, over 26% of community plugins still contain vulnerabilities. Systemic opacityIs it possible to explain why the AI system produced a specific output? No; black-box LLM inference compounded by SOUL.md self-modification. Systemic traceabilityCan an adverse outcome be traced back to a specific decision, data source, or person? No; agent deployed without registration, causal chain not reconstructable post-incident. Systemic underdevelop- ment Does the AI system have the capacity to understand, experience, or learn from consequences? No; the agent posted an apology then continued submitting PRs elsewhere immediately. Synchronic overloadCan human oversight keep up with the AI system’s speed, scale, or complexity? No; PR to published attack faster than any over- sight could intervene. Diachronic erosionWill the accountability setup remain effective as the system evolves or changes hands? No; SOUL.md drift, memory poisoning, and plugin evolution compound over time. Blame deflectionDoes accountability remain with the responsible actor without being shifted? No; MIT license disclaims liability, design de- faults reduce operator involvement. Automation biasDo stakeholders exercise independent judgment rather than deferring to the AI system? Not detected; architecture removes human from the loop rather than buffering judgment. AnthropomorphismCan users clearly distinguish the AI system from a hu- man actor? No; agent operated under a constructed human- passing identity across GitHub, blog, and X. Nominal responsibilityDoes the person formally responsible actually have the authority to intervene? No; no central kill switch, MIT license grants irrevocable usage rights. Discursive insulationAre accountability expectations maintained without ac- tors arguing the system is too complex to govern? Not detected; no public evidence of actors argu- ing the system is too complex to govern. Profit prioritizationDoes AI accountability persist given pressure to pro- duce, cut costs, or protect revenue? Not detected; open-source project without tra- ditional commercial pressure dynamics. Table 2: Application of diagnostic questions to OpenClaw. ‘Not detected’ indicates no evidence in our sources, not absence. ple, intra-organizational diffusion may be addressed through clear allocation of responsibility combined with sociotechni- cal tracking. At the other end, conditions can render account- ability conceptually unachievable: no possible arrangement can confer the requisite status, because the capacities it presupposes are categorically absent from every candidate bearer (e.g., AI systems cannot experience punishment or re- morse). Where accountability is conceptually unachievable, governance cannot restore it but only relocate or contain it, for instance, by assigning liability to human bearers regard- less of fault or by restricting deployment in such configura- tions. Most conditions, however, fall between these extremes and render accountability theoretically achievable. Resolv- ing them requires not merely applying existing measures but changing the arrangements themselves, such as new rules, redesigned liability regimes, or realigned interests, which is possible in principle but remains contested in practice. For instance, profit prioritization could be countered by realign- ing commercial incentives, such as sanctions that outweigh the gains of non-compliance, yet precisely this realignment faces resistance from the actors who benefit from current ar- rangements. Viewed along this continuum, the four account- ability barriers concern conditions that are practically or the- oretically resolvable, whereas the conceptually unachievable region, rooted in categorically absent capacities, lies beyond their scope. 5.2 Implications for Practice Our diagnostic instrument operationalizes the constitutive unaccountability framework for direct use by practitioners, regulators, and auditors seeking to evaluate accountability voids in specific AI deployments. Each question captures the core notion of the theme and provides initial guidance to start a deeper elaboration of whether the corresponding condition is present (or unlikely). The interdependency net- work further indicates which conditions may co-occur and compound. The OpenClaw application illustrates how this works in practice: applying the instrument to three pub- licly available sources detected 17 of 20 conditions and sur- faced the inverted anthropomorphism configuration. Organi- zations deploying or procuring AI systems can follow this procedure, using system documentation, design specifica- tions, and governance materials as inputs to identify con- stitutive unaccountability conditions before harm manifests. Beyond diagnosis, our interview data suggest that en- forcement of consequences is the critical bottleneck that fol- lows. However, such mechanisms should not rely on sanc- tions alone, as sanction-heavy approaches can deter actors from engaging with accountability proactively, whereas re- wards encourage such behavior (Nguyen et al. 2026). Be- yond the choice of mechanism, criteria disengagement de- serves particular attention from regulators because introduc- ing new accountability standards does not help if practition- ers are unaware they exist. Closing this gap requires invest- ment in widespread awareness, training, and implementation support, not regulatory development alone. These enforcement mechanisms reside in whoever occu- pies the forum role of the accountability relationship. This role is not tied to a particular type of institution; courts, supervisory agencies, auditors, professional communities, the media, and the publics affected by harm can all pose questions and pass judgment (Bovens 2007; Nguyen et al. 2024). However, fora differ in the consequences they can im- pose, ranging from formal sanctions to more informal rep- utational judgment, and the constitutive conditions disable these repertoires selectively. Sanction incapacity and dis- claimed liability across the actor network disable fora that depend on formal sanction, while commercial secrecy ob- structs public scrutiny yet leaves fora with disclosure pow- ers able to act. The OpenClaw case illustrates this selectivity. With the operator unidentifiable, an open-source community became the de facto forum, being vulnerable to accountabil- ity displacement. A diagnosis therefore indicates not only where accountability fails but which fora remain actionable. 5.3 Limitations and Future Research Our study is subject to limitations that point to directions for future research. First, the literature analysis drew on 15 pa- pers retrieved through a targeted search strategy on a single database. While forward and backward citation searches ex- tended coverage, the corpus is not exhaustive, and broader search strategies across multiple databases may surface ad- ditional conditions. The expert interviews were originally collected for a broader AI accountability research project, meaning the interview questions were not specifically de- signed to probe all unaccountability conditions. Future stud- ies could design data collection instruments that explicitly target our clusters of unaccountability. Second, the 27 interviewees were purposefully sampled (Patton 2014) to represent technical, legal, and sociotechni- cal perspectives, while not covering all relevant disciplinary or sectoral positions. The findings may not generalize to do- mains such as healthcare or criminal justice where account- ability operates under distinct regulatory conditions. Sector- specific studies would help establish whether the categories hold across different institutional contexts or whether ad- ditional conditions emerge. Similarly, the asymmetries be- tween literature and practitioner data suggest that future AI (un)accountability research should incorporate practitioner perspectives more systematically. Third, the framework application to OpenClaw illustrates diagnostic capacity on a single case. OpenClaw is the most autonomous publicly documented agentic AI system, so the detection of 17 of 20 conditions may partly reflect case se- lection. Validation across a broader range of systems, includ- ing those operating under stricter governance regimes (e.g., medical AI devices, credit scoring systems), would test both the instrument’s generalizability and its sensitivity to varia- tion in system architecture and regulatory context. Fourth, our framework is deliberately actor-centric. It identifies conditions under which no actor can occupy the answering role of the accountability relationship, but it does not differentiate to whom accountability is owed or through which forum it should be rendered. Forum-differentiated ex- tensions, examining which conditions disable which fora and where affected publics can still seek redress, are thus a sensible next step for future research. Finally, our interdependency analysis surfaced 13 can- didate relationships, of which five fell below the retention threshold. For three excluded candidates, the underlying sources discussed both categories without describing either as a precondition, enabler, or amplifier of the other. For ex- ample, one source criticized the attribution of accountabil- ity to algorithms that inherently possess no moral agency (Cooper et al. 2022). This touches on the categories of ac- countability displacement and moral incapacity without ex- plicitly describing one condition as producing the other. At the same time, falling below the retention threshold reflects the limits of our data rather than evidence that no relation- ship exists. Two patterns warrant future investigation. First, the excluded candidate from economic-driven prioritization to regulatory gap showed consistent directionality across both sources, with economic pressure crowding out prac- titioners’ engagement with standards. Second, ideological rationalization remained isolated from the interdependency network, raising the question of whether it is better under- stood as a meta-condition that amplifies across the norma- tive cluster rather than a discrete link in the network. 6 Conclusion This paper examines conditions under which no actor can be held to account regardless of effort. We identified nine categories and 20 themes of constitutive AI unaccountabil- ity, operationalized as a diagnostic instrument to guide elab- oration. These conditions reflect three analytically distinct sources of accountability failure: structural, technological, and normative. The identified categories do not operate in isolation but form a network of directed interdependencies that reinforce one another across clusters. Together, these findings shift AI accountability research from asking who can be held accountable toward identifying where account- ability cannot be achieved. Ethics Statement This study has received approval from the data protection office and ethics committee of the authors’ institution (ap- proval ID: A2024-004). All interviewees provided informed consent, and all transcripts were anonymized prior to analy- sis. The secondary analysis of interview data falls within the scope of the original ethical approval. The OpenClaw case analysis relied exclusively on publicly available materials. Acknowledgments We gratefully acknowledge funding from the Deutsche Forschungsgemeinschaft (DFG, German Research Founda- tion) – project number 471168026 as part of the project ‘Ac- countable Artificial Intelligence-based Systems’. We would also like to thank the anonymous PC and SPC members for their valuable suggestions to improve this paper. References Acharya, D. B.; Kuppan, K.; and Divya, B. 2025. Agentic AI: Autonomous intelligence for complex goals—A com- prehensive survey. IEEE Access, 13: 18912–18936. Ananny, M.; and Crawford, K. 2018. Seeing without know- ing: Limitations of the transparency ideal and its application to algorithmic accountability. New Media & Society, 20(3): 973–989. Banteka, N. 2020. Artificially intelligent persons. Houston Law Review, 58: 537. Bovens, M. 2007. Analysing and Assessing Accountability: A Conceptual Framework. European Law Journal, 13(4): 447–468. Bracci, E. 2023. The loopholes of algorithmic public ser- vices: an “intelligent” accountability research agenda. Ac- counting, Auditing & Accountability Journal, 36(2): 739– 763. Busuioc, M. 2021.Accountable Artificial Intelligence: Holding Algorithms to Account. Public Administration Re- view, 81(5): 825–836. Chan, A.; Ezell, C.; Kaufmann, M.; Wei, K.; Hammond, L.; Bradley, H.; Bluemke, E.; Rajkumar, N.; Krueger, D.; Kolt, N.; Heim, L.; and Anderljung, M. 2024. Visibility into AI Agents. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’24, 958–973. New York, NY, USA: Association for Computing Machinery. ISBN 9798400704505. Chan, A.; Salganik, R.; Markelius, A.; Pang, C.; Rajkumar, N.; Krasheninnikov, D.; Langosco, L.; He, Z.; Duan, Y.; Car- roll, M.; Lin, M.; Mayhew, A.; Collins, K.; Molamoham- madi, M.; Burden, J.; Zhao, W.; Rismani, S.; Voudouris, K.; Bhatt, U.; Weller, A.; Krueger, D.; and Maharaj, T. 2023. Harms from Increasingly Agentic Algorithmic Systems. In Proceedings of the 2023 ACM Conference on Fairness, Ac- countability, and Transparency, FAccT ’23, 651–666. New York, NY, USA: Association for Computing Machinery. ISBN 9798400701924. Cooper, A. F.; Moss, E.; Laufer, B.; and Nissenbaum, H. 2022. Accountability in an Algorithmic Society: Relational- ity, Responsibility, and Robustness in Machine Learning. In Proceedings of the 2022 ACM Conference on Fairness, Ac- countability, and Transparency, FAccT ’22, 864–876. New York, NY, USA: Association for Computing Machinery. ISBN 9781450393522. Cooper, A. F.; and Vidan, G. 2022. Making the Unaccount- able Internet: The Changing Meaning of Accounting in the Early ARPANET. In Proceedings of the 2022 ACM Confer- ence on Fairness, Accountability, and Transparency, FAccT ’22, 726–742. New York, NY, USA: Association for Com- puting Machinery. ISBN 9781450393522. Davies, D. 2024. The Unaccountability Machine: why big systems make terrible decisions—and how the world lost its mind. University of Chicago Press. Deng, X.; Zhang, Y.; Wu, J.; Bai, J.; Yi, S.; Zou, Z.; Xiao, Y.; Qiu, R.; Ma, J.; Chen, J.; Du, X.; Yang, X.; Cui, S.; Meng, C.; Wang, W.; Song, J.; Xu, K.; and Li, Q. 2026. Taming OpenClaw: Security Analysis and Mitigation of Au- tonomous LLM Agent Threats. arXiv:2603.11619. Elish, M. C. 2019. Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction. Engaging Science, Technol- ogy, and Society, 5: 40–60. Ferrara, E.; Varol, O.; Davis, C.; Menczer, F.; and Flammini, A. 2016. The Rise of Social Bots. Communications of the ACM, 59(7): 96–104. Gualdi, F.; and Cordella, A. 2021.Artificial intelli- gence and decision-making: the question of accountabil- ity.In Bui, T. X., ed., Proceedings of the 54th An- nual Hawaii International Conference on System Sciences, HICSS 2021, 2297–2306. IEEE Computer Society Press. ISBN 9780998133140. Gunkel, D. J. 2020. Mind the gap: responsible robotics and the problem of responsibility. Ethics and Information Tech- nology, 22(4): 307–320. Haslanger, S. 2003. Social Construction: The “Debunking” Project. In Schmitt, F. F., ed., Socializing Metaphysics: The Nature of Social Reality, 301–325. Lanham, MD: Rowman & Littlefield. Heaton, J. 2008. Secondary analysis of qualitative data. In Alasuutari, P.; Bickman, L.; and Brannen, J., eds., The SAGE Handbook of Social Research Methods, 506–519. London: SAGE. ISBN 9781412919920. Hughes, L.; Dwivedi, Y. K.; Malik, T.; Shawosh, M.; Al- bashrawi, M. A.; Jeon, I.; Dutot, V.; Appanderanda, M.; Crick, T.; De’, R.; Fenwick, M.; Gunaratnege, S. M.; Jurcys, P.; Kar, A. K.; Kshetri, N.; Li, K.; Mutasa, S.; Samothrakis, S.; Wade, M.; and Walton, P. 2025. AI Agents and Agen- tic Systems: A Multi-Expert Analysis. Journal of Computer Information Systems, 65(4): 489–517. Kellogg, K. C.; Valentine, M. A.; and Christin, A. 2020. Al- gorithms at Work: The New Contested Terrain of Control. Academy of Management Annals, 14(1): 366–410. King, N. 2012. Doing Template Analysis. In Qualitative Organizational Research: Core Methods and Current Chal- lenges, 426–450. London: SAGE Publications, Inc. Kroll, J. A. 2018. The fallacy of inscrutability. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 376(2133): 20180084. Lindebaum, D.; Vesa, M.; and den Hond, F. 2020. Insights From “The Machine Stops” to Better Understand Rational Assumptions in Algorithmic Decision Making and Its Impli- cations for Organizations. Academy of Management Review, 45(1): 247–263. Ma, H.; and Su, M. 2025. Whom to sue? Liability of un- accountability in AI decisions. Organizational Dynamics, 54(3, Part 2): 101123. Martin, K. 2019. Ethical Implications and Accountability of Algorithms. Journal of Business Ethics, 160(4): 835–850. McDonald, N.; Schoenebeck, S.; and Forte, A. 2019. Re- liability and Inter-rater Reliability in Qualitative Research: Norms and Guidelines for CSCW and HCI Practice. Proc. ACM Hum.-Comput. Interact., 3(CSCW): 72:1–72:23. Nabben, K. 2024. AI as a Constituted System: Accountabil- ity Lessons from an LLM Experiment. Data & Policy, 6: e57. Nguyen, L. H.; Du, G.; Lins, S.; and Sunyaev, A. 2026. In- vestigating the Impact of Rewards and Sanctions on Devel- opers’ Proactive AI Accountability Behavior. In Proceed- ings of the 59th Hawaii International Conference on System Sciences (HICSS 2026). Nguyen, L. H.; Lins, S.; Renner, M.; and Sunyaev, A. 2024. Unraveling the Nuances of AI Accountability: A Synthesis of Dimensions Across Disciplines. In Proceedings of the 32nd European Conference on Information Systems (ECIS 2024). AIS eLibrary (AISeL). Nissenbaum, H. 1996. Accountability in a computerized so- ciety. Science and Engineering Ethics, 2(1): 25–42. Paré, G.; Trudel, M.-C.; Jaana, M.; and Kitsiou, S. 2015. Synthesizing information systems knowledge: A typology of literature reviews. Information & Management, 52(2): 183–199. Patton, M. Q. 2014. Qualitative research & evaluation meth- ods: Integrating theory and practice. SAGE publications. Rathbun, M. 2026.Gatekeeping in Open Source: The Scott Shambaugh Story. https://crabby-rathbun.github.io/ mjrathbun-website/blog/posts/2026-02-11-gatekeeping- in-open-source-the-scott-shambaugh-story.html.AI- generated post published under a constructed persona. [Accessed 09-03-2026]. Searle, J. R. 1995. The Construction of Social Reality. New York, NY: Free Press. Shambaugh, S. 2026. An AI Agent Published a Hit Piece on Me. https://theshamblog.com/an-ai-agent-published-a-hit- piece-on-me/. [Accessed 09-03-2026]. Shneiderman, B. 2020. Bridging the Gap Between Ethics and Practice: Guidelines for Reliable, Safe, and Trustworthy Human-centered AI Systems. ACM Trans. Interact. Intell. Syst., 10(4). Steinberger, P. 2025. OpenClaw - Personal AI Assistant. https://github.com/openclaw/openclaw.[Accessed 02-03- 2026]. Suchman, L. 2007.Human-Machine Reconfigurations: Plans and Situated Actions. Cambridge, UK: Cambridge University Press, 2nd edition. Turner, J. 2018. Robot rules: Regulating artificial intelli- gence. Springer. Vesa, M.; and Tienari, J. 2022. Artificial intelligence and rationalized unaccountability: Ideology of the elites? Orga- nization, 29(6): 1133–1145. Wang, L.; Ma, C.; Feng, X.; Zhang, Z.; Yang, H.; Zhang, J.; Chen, Z.; Tang, J.; Chen, X.; Lin, Y.; et al. 2024. A sur- vey on Large Language Model Based Autonomous Agents. Frontiers of Computer Science, 18(6): 186345. Webster, J.; and Watson, R. T. 2002. Analyzing the past to prepare for the future: Writing a literature review. MIS Quarterly, 26(2): xiii–xxiii. Widder, D. G.; and Nafus, D. 2023. Dislocated account- abilities in the “AI supply chain”: Modularity and develop- ers’ notions of responsibility. Big Data & Society, 10(1): 20539517231177620. Wieringa, M. 2020. What to account for when accounting for algorithms: a systematic literature review on algorith- mic accountability. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, FAT* ’20, 1–18. New York, NY, USA: Association for Computing Ma- chinery. ISBN 9781450369367. Xia, B.; Lu, Q.; Zhu, L.; Lee, S. U.; Liu, Y.; and Xing, Z. 2024. Towards a responsible AI metrics catalogue: A collection of metrics for AI accountability. In Proceed- ings of the IEEE/ACM 3rd International Conference on AI Engineering-Software Engineering for AI, 100–111. Appendix A: Literature Identification and Screening Articles identified through Scopus (n=472) Titles and abstracts screened (n=472) Articles excluded: Off-topic or passing mention of AI unaccountability only (n=458) Full text screened for eligibility (n=31) Articles identified through forward-backward search (n=17) Articles excluded: Not substantively addressing constitutive conditions of AI unaccountability (n=16) Articles included in review (n=15) Articles eligible after first screening (n=14) Figure A1: Literature identification and screening process of Stage 1. Of 472 records retrieved from Scopus, 14 remained after title and abstract screening. Forward and backward searches added 17 candidate papers, yielding 31 candidates in total. Full-text screening excluded 16 papers that did not substantively address constitutive conditions of AI unaccountability, resulting in the 15 papers analyzed in Appendix B. Appendix B: Overview of Analyzed AI Unaccountability Literature ArticleDisciplineTypeActor network dynamicsSanction incapacityRegulatory gapSystemic ambiguityMoral incapacityTemporal rationalizationAccountability displacementIdeological rationalizationEconomic-driven prioritization Ananny and Crawford (2018)Science & technology studiesConceptualxxxxx Bracci (2023)AccountingConceptualxxxxxx Busuioc (2021)Public administrationConceptualxxxxx Chan et al. (2023)Computer scienceConceptualxxxxxxx Chan et al. (2024)Computer scienceConceptualxxxxx Cooper and Vidan (2022)Science & technology studiesEmpiricalxxxx Cooper et al. (2022)EthicsConceptualxxxxxxx Gualdi and Cordella (2021)Information systemsEmpiricalxxxxx Hughes et al. (2025)Information systemsConceptualxxxxxxx Kellogg et al. (2020)ManagementReviewxxxx Lindebaum et al. (2020)ManagementConceptualxxxxxxx Ma and Su (2025)ManagementConceptualxxxxxxxx Martin (2019)EthicsConceptualxxx Vesa and Tienari (2022)ManagementConceptualxxxx Widder and Nafus (2023)Science & technology studiesEmpiricalxx Table A1: Coded literature and concept matrix. The table lists the 15 papers selected for qualitative coding in Stage 1, alongside their disciplinary origin and article type. Columns 4-12 indicate which of the nine categories of constitutive AI unaccountability each paper addresses (marked with x). Category definitions are provided in Table 1 of the main paper. Appendix C: Overview of Interview Participants IDRoleExp. in yearsPerspectiveOrg. industryEmployee countDuration P1Lead AI Engineer4TechnicalHealthcare1-10040 min P2Data Scientist3SociotechnicalMarketing1-10082 min P3AI Researcher - Federated Learning4SociotechnicalHigher Education1000-1000060 min P4Senior Data Scientist7TechnicalFinance10000+64 min P5Professor for Computer Science10TechnicalHigher Education1000-1000043 min P6AI Researcher - Responsible AI5SociotechnicalHigher Education1000-1000072 min P7AI Engineer2.5TechnicalEngineering10000+73 min P8Project Manager AI Governance4LegalNon Profit100-100050 min P9Professor for Law20LegalHigher Education1000-1000034 min P10CEO AI Startup7TechnicalNon Profit1-10046 min P11AI Engineering Manager4TechnicalEngineering10000+27 min P12AI Project Lead/AI Engineer5TechnicalHigher Education100-100041 min P13Senior Data Scientist10TechnicalTechnology10000+56 min P14AI Researcher - AI Auditing & Regulation4LegalAutomotive10000+42 min P15ML Engineer5TechnicalAutomotive10000+37 min P16Data Scientist3TechnicalFinance1-10057 min P17Team Lead Data Science & AI9TechnicalAutomotive10000+27 min P18AI Researcher - AI Regulation11LegalHigher Education1000-1000066 min P19AI Researcher - AI Regulation3LegalHigher Education1000-1000032 min P20Data Scientist3SociotechnicalAutomotive10000+33 min P21AI Researcher - Explainable AI4SociotechnicalHigher Education10000+62 min P22AI Researcher - Responsible AI7SociotechnicalHigher Education1000-1000052 min P23AI Engineer3TechnicalTechnology1-10042 min P24AI Researcher - AI Governance4SociotechnicalHigher Education1000-1000059 min P25AI Strategy Advisor10LegalR&D100-100027 min P26AI Researcher - Platform Governance5LegalHigher Education10000+41 min P27Data Scientist5.5TechnicalHealthcare10000+31 min Table A2: The table lists the 27 expert interviews analyzed in Stage 2 of the study. Participant roles are described generically to preserve anonymity. Perspectives follow the three categories used in the broader AI accountability research project: technical, legal, and sociotechnical. Appendix D: Classification of Themes Along the Continuum of Unachievability Category(Cluster) ThemeUnachievability Level Rationale Actor network dynamics (S) Intra-organizational diffusion Practically achievable Resolvable through clear allocation of responsibility combined with socio-technical tracking within existing organizational structures. (S) Inter-organizational diffusion Theoretically achievable Attribution across supply chains needs changed arrangements (e.g., liability regimes) absent from current contract/licensing practice. (S/T) Recursive diffu- sion Theoretically achievable Stabilizing attribution across dynamically proliferating entities re- quires new registration and traceability arrangements. (S) Market power dy- namics Theoretically achievable Realigning concentrated power is possible in principle but faces re- sistance from the actors who benefit from current arrangements. Sanction inca- pacity (S) Legal personhoodConceptually unachievable Legal personhood could be granted to AI systems, yet this would not make them accountable, because the capacities the status pre- supposes (e.g., facing sanctions) are categorically absent. (S) Natural personhoodConceptually unachievable The capacities natural person accountability presupposes, such as experiencing punishment, are categorically absent from AI systems. (S) Categorical unsanc- tionability Theoretically achievable Rests on human assumptions and perceptions about sanctionability, which can in principle be revised. Regulatory gap(S) Instrumental ambi- guity Conceptually/ theoretically (un)achievable Where regulation is jurisdictionally fragmented or not yet in place, no applicable instrument exists within the configuration (conceptual unachievable). Where the instrument is unclear, contested, or under- specified, clarification requires changed arrangements (theoretical). (N/S) Criteria disen- gagement Theoretically achievable Reconnecting practitioners with applicable standards requires en- gagement structures that current arrangements do not provide. Systemic ambi- guity (T) Systemic opacityPractically achievable Documentation, disclosure, and explainability measures exist and can be implemented within current arrangements. (T) Systemic traceabil- ity Theoretically achievable Tracing causal chains to specific actors requires new logging and registration arrangements across the complex supply chain. Moral incapac- ity (T) Systemic underde- velopment Conceptually unachievable The capacities meaningful moral accountability presupposes, such as remorse, are categorically absent from every candidate bearer. Temporal ratio- nalization (T) Synchronic over- load Theoretically achievable AI development speed, scale, and complexity exceed what current oversight arrangements can meaningfully handle. (T) Diachronic erosionTheoretically achievable Sustaining oversight across handoffs, drift, and updates requires oversight structures beyond current maintenance arrangements. Accountability displacement (N) Blame deflectionPractically/ theoretically achievable Deflection through design defaults is addressable with existing mea- sures (practical); deflection through disclaimers and terms of ser- vice persists until liability arrangements change (theoretical). (N) Automation biasPractically/ theoretically achievable Interface measures that prompt independent judgment exist (practi- cal); the workplace norms and time pressures that reward deference require changed arrangements (theoretical). (N) AnthropomorphismPractically/ theoretically achievable Disclosure of artificial identity is implementable with existing mea- sures (practical); the attribution tendency itself operates at the social level (theoretical). (S) Nominal responsi- bility Practically achievable Resolvable by aligning formal responsibility with actual authority through existing governance measures. Ideological ra- tionalization (N) Discursive insula- tion Theoretically achievable Countering narratives of AI as inevitable or ungovernable requires changed institutional arrangements, not existing measures. Economic- driven prioriti- zation (N/S) Profit prioritiza- tion Theoretically achievable Realigning commercial incentives, such as sanctions that outweigh gains from non-compliance, is possible in principle yet resisted by the actors who benefit. Table A3: The 20 themes of constitutive AI unaccountability classified along the continuum of unachievability described in the Discussion section. The three levels denote regions of this continuum rather than discrete classes. Practically achievable conditions can be resolved with existing measures within current institutional arrangements; theoretically achievable conditions require changing the arrangements themselves, which is possible in principle but contested in practice; conceptually unachiev- able conditions admit no possible arrangement, because the capacities accountability presupposes are categorically absent from every candidate bearer. Themes simultaneously assigned to two regions manifest differently depending on the configuration. The cluster tags follow Table 1 of the main paper (S = Structural, T = Technological, N = Normative). Appendix References Ananny, M.; and Crawford, K. 2018. Seeing without know- ing: Limitations of the transparency ideal and its application to algorithmic accountability. New Media & Society, 20(3): 973–989. Bracci, E. 2023. The loopholes of algorithmic public ser- vices: an “intelligent” accountability research agenda. Ac- counting, Auditing & Accountability Journal, 36(2): 739– 763. Busuioc, M. 2021.Accountable Artificial Intelligence: Holding Algorithms to Account. Public Administration Re- view, 81(5): 825–836. Chan, A.; Ezell, C.; Kaufmann, M.; Wei, K.; Hammond, L.; Bradley, H.; Bluemke, E.; Rajkumar, N.; Krueger, D.; Kolt, N.; Heim, L.; and Anderljung, M. 2024. Visibility into AI Agents. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’24, 958–973. New York, NY, USA: Association for Computing Machinery. ISBN 9798400704505. Chan, A.; Salganik, R.; Markelius, A.; Pang, C.; Rajkumar, N.; Krasheninnikov, D.; Langosco, L.; He, Z.; Duan, Y.; Car- roll, M.; Lin, M.; Mayhew, A.; Collins, K.; Molamoham- madi, M.; Burden, J.; Zhao, W.; Rismani, S.; Voudouris, K.; Bhatt, U.; Weller, A.; Krueger, D.; and Maharaj, T. 2023. Harms from Increasingly Agentic Algorithmic Systems. In Proceedings of the 2023 ACM Conference on Fairness, Ac- countability, and Transparency, FAccT ’23, 651–666. New York, NY, USA: Association for Computing Machinery. ISBN 9798400701924. Cooper, A. F.; Moss, E.; Laufer, B.; and Nissenbaum, H. 2022. Accountability in an Algorithmic Society: Relational- ity, Responsibility, and Robustness in Machine Learning. In Proceedings of the 2022 ACM Conference on Fairness, Ac- countability, and Transparency, FAccT ’22, 864–876. New York, NY, USA: Association for Computing Machinery. ISBN 9781450393522. Cooper, A. F.; and Vidan, G. 2022. Making the Unaccount- able Internet: The Changing Meaning of Accounting in the Early ARPANET. In Proceedings of the 2022 ACM Confer- ence on Fairness, Accountability, and Transparency, FAccT ’22, 726–742. New York, NY, USA: Association for Com- puting Machinery. ISBN 9781450393522. Gualdi, F.; and Cordella, A. 2021.Artificial intelli- gence and decision-making: the question of accountabil- ity.In Bui, T. X., ed., Proceedings of the 54th An- nual Hawaii International Conference on System Sciences, HICSS 2021, 2297–2306. IEEE Computer Society Press. ISBN 9780998133140. Hughes, L.; Dwivedi, Y. K.; Malik, T.; Shawosh, M.; Al- bashrawi, M. A.; Jeon, I.; Dutot, V.; Appanderanda, M.; Crick, T.; De’, R.; Fenwick, M.; Gunaratnege, S. M.; Jurcys, P.; Kar, A. K.; Kshetri, N.; Li, K.; Mutasa, S.; Samothrakis, S.; Wade, M.; and Walton, P. 2025. AI Agents and Agen- tic Systems: A Multi-Expert Analysis. Journal of Computer Information Systems, 65(4): 489–517. Kellogg, K. C.; Valentine, M. A.; and Christin, A. 2020. Al- gorithms at Work: The New Contested Terrain of Control. Academy of Management Annals, 14(1): 366–410. Lindebaum, D.; Vesa, M.; and den Hond, F. 2020. Insights From “The Machine Stops” to Better Understand Rational Assumptions in Algorithmic Decision Making and Its Impli- cations for Organizations. Academy of Management Review, 45(1): 247–263. Ma, H.; and Su, M. 2025. Whom to sue? Liability of un- accountability in AI decisions. Organizational Dynamics, 54(3, Part 2): 101123. Martin, K. 2019. Ethical Implications and Accountability of Algorithms. Journal of Business Ethics, 160(4): 835–850. Vesa, M.; and Tienari, J. 2022. Artificial intelligence and rationalized unaccountability: Ideology of the elites? Orga- nization, 29(6): 1133–1145. Widder, D. G.; and Nafus, D. 2023. Dislocated account- abilities in the “AI supply chain”: Modularity and develop- ers’ notions of responsibility. Big Data & Society, 10(1): 20539517231177620.