arxiv: 2604.07778 · v1 · submitted 2026-04-09 · 💻 cs.AI

Recognition: 1 theorem link

· Lean Theorem

The Accountability Horizon: An Impossibility Theorem for Governing Human-Agent Collectives

Haileleol Tibebu

Authors on Pith no claims yet

Pith reviewed 2026-05-10 17:34 UTC · model grok-4.3

classification 💻 cs.AI

keywords accountabilityAI governanceimpossibility theoremhuman-agent collectivesautonomyfeedback cyclesresponsibility attribution

0 comments

The pith

When human-AI collectives exceed the Accountability Horizon and contain feedback cycles, no framework satisfies all four required accountability properties at once.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proves an impossibility result for accountability in systems that combine humans and AI agents. It shows that once the combined autonomy of the group crosses a computable threshold and the interaction structure includes loops between human and AI actions, it becomes mathematically impossible to maintain a system that attributes responsibility to specific actors, respects limits on what can be foreseen, ensures at least one party bears real responsibility, and fully allocates all responsibility without gaps. This directly challenges the foundational assumption in law, ethics, and regulation that every consequential outcome has at least one person with sufficient involvement and foresight to be held accountable. If the result holds, existing oversight tools such as audits and transparency requirements cannot bridge the gap without lowering autonomy itself, creating a sharp dividing line between regimes where traditional accountability works and regimes where it cannot.

Core claim

The Accountability Incompleteness Theorem proves that for any collective whose compound autonomy exceeds the Accountability Horizon and whose interaction graph contains a human-AI feedback cycle, no framework can satisfy all four properties simultaneously. The four properties are Attributability, requiring causal contribution for responsibility; Foreseeability Bound, limiting responsibility to what can be predicted; Non-Vacuity, ensuring at least one agent bears non-trivial responsibility; and Completeness, requiring full allocation of responsibility. The proof models agents as state-policy tuples in a shared structural causal model, characterises autonomy via a four-dimensional information-

What carries the argument

The Accountability Horizon: the computable threshold of compound autonomy, measured in a four-dimensional information-theoretic profile of epistemic, executive, evaluative, and social dimensions, beyond which the four accountability properties become incompatible in the presence of human-AI feedback cycles.

If this is right

Legitimate accountability frameworks that meet all four properties exist only for collectives whose compound autonomy remains below the Accountability Horizon.
Above the horizon, transparency, audits, and oversight mechanisms cannot resolve the incompatibility without reducing the autonomy of the collective.
The impossibility arises specifically when interaction graphs include human-AI feedback cycles; acyclic structures may still permit frameworks below the horizon.
Experiments on 3,000 synthetic collectives confirm the theorem holds with zero violations, establishing a phase transition in governance feasibility.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Design choices that deliberately keep collectives below the horizon could preserve the viability of existing accountability approaches.
The structural role of feedback cycles suggests that interaction topology itself becomes a primary target for governance interventions rather than post-hoc attribution.
The same modelling approach could be used to test whether analogous impossibilities arise for other governance requirements such as fairness or safety in joint human-AI systems.

Load-bearing premise

The four properties of Attributability, Foreseeability Bound, Non-Vacuity, and Completeness are the minimal necessary conditions for any legitimate accountability framework, and the four-dimensional autonomy model together with interaction graphs fully capture the features relevant to responsibility attribution.

What would settle it

Discovery of even one human-AI collective whose compound autonomy exceeds the Accountability Horizon, whose interaction graph contains a feedback cycle, and yet admits a framework that satisfies all four properties simultaneously would falsify the theorem.

read the original abstract

Existing accountability frameworks for AI systems, legal, ethical, and regulatory, rest on a shared assumption: for any consequential outcome, at least one identifiable person had enough involvement and foresight to bear meaningful responsibility. This paper proves that agentic AI systems violate this assumption not as an engineering limitation but as a mathematical necessity once autonomy exceeds a computable threshold. We introduce Human-Agent Collectives, a formalisation of joint human-AI systems where agents are modelled as state-policy tuples within a shared structural causal model. Autonomy is characterised through a four-dimensional information-theoretic profile (epistemic, executive, evaluative, social); collective behaviour through interaction graphs and joint action spaces. We axiomatise legitimate accountability through four minimal properties: Attributability (responsibility requires causal contribution), Foreseeability Bound (responsibility cannot exceed predictive capacity), Non-Vacuity (at least one agent bears non-trivial responsibility), and Completeness (all responsibility must be fully allocated). Our central result, the Accountability Incompleteness Theorem, proves that for any collective whose compound autonomy exceeds the Accountability Horizon and whose interaction graph contains a human-AI feedback cycle, no framework can satisfy all four properties simultaneously. The impossibility is structural: transparency, audits, and oversight cannot resolve it without reducing autonomy. Below the threshold, legitimate frameworks exist, establishing a sharp phase transition. Experiments on 3,000 synthetic collectives confirm all predictions with zero violations. This is the first impossibility result in AI governance, establishing a formal boundary below which current paradigms remain valid and above which distributed accountability mechanisms become necessary.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper claims a new impossibility theorem for accountability in human-AI collectives once autonomy crosses a threshold with feedback cycles present, but the structural causal model setup looks inconsistent with those cycles.

read the letter

The paper's central claim is that there's a mathematical limit to individual accountability in human-AI collectives once their combined autonomy crosses a certain threshold and feedback cycles are involved. No framework can then satisfy all four accountability properties at once. What stands out as new is the formalization of Human-Agent Collectives using state-policy tuples in a shared structural causal model, along with the four-dimensional autonomy profile covering epistemic, executive, evaluative, and social aspects. They define the Accountability Horizon and prove an incompleteness theorem that creates a phase transition: below the horizon, standard approaches work; above it with cycles, you need distributed responsibility instead. The 3,000 synthetic experiments showing zero violations is a solid check on the predictions. The axiomatization of the four properties—Attributability, Foreseeability Bound, Non-Vacuity, and Completeness—looks like a reasonable starting point for what any accountability system should do. Framing the problem in information-theoretic and graph terms brings some rigor to a field that often stays qualitative. That said, there's a real question about the setup with structural causal models. Those are typically acyclic to keep interventions and counterfactuals well-defined, yet the theorem triggers exactly when there are human-AI feedback cycles. If the paper doesn't spell out how they handle the cycles—maybe through dynamic extensions or separate layers—the domain of the theorem could be empty or the derivation shaky. The choice of axioms might also be tuned to produce the impossibility, though they call them minimal. The experiments being purely synthetic means we don't see how this plays out with actual systems or data. This is the kind of paper for researchers in AI safety, governance, and formal methods who want to see mathematical boundaries in policy discussions. It shows clear thinking in trying to pin down the limits formally. I would bring it to a reading group to discuss the proof and the cycle issue. It deserves peer review because the formal approach is worth referee scrutiny, even if revisions are needed on the modeling details. I'd recommend sending it out rather than desk rejecting.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces Human-Agent Collectives as joint human-AI systems in which agents are state-policy tuples embedded in a shared structural causal model. Autonomy is represented by a four-dimensional information-theoretic profile (epistemic, executive, evaluative, social), collective structure by interaction graphs and joint action spaces, and legitimate accountability by four axioms: Attributability (causal contribution required), Foreseeability Bound (responsibility limited by predictive capacity), Non-Vacuity (at least one non-trivial bearer), and Completeness (full allocation). The central result is the Accountability Incompleteness Theorem: any collective whose compound autonomy exceeds the Accountability Horizon and whose interaction graph contains a human-AI feedback cycle admits no framework satisfying all four axioms simultaneously. The paper reports a sharp phase transition below the horizon and confirms all predictions on 3,000 synthetic collectives with zero violations.

Significance. If the theorem and its supporting formalization hold, the result supplies a mathematically precise boundary for AI governance, showing that person-centric accountability frameworks become structurally inadequate once autonomy crosses a computable threshold and feedback cycles are present. The axiomatic treatment, the information-theoretic autonomy profile, and the large-scale synthetic validation (zero violations across 3,000 instances) constitute genuine strengths that could anchor future work on distributed accountability mechanisms.

major comments (2)

[Abstract and modeling paragraphs] Abstract and modeling paragraphs: agents are defined as state-policy tuples inside a shared structural causal model while the theorem is stated only for interaction graphs that contain human-AI feedback cycles. Standard SCMs are directed acyclic graphs; cycles render do-calculus and counterfactuals ill-defined, which are the primitives underlying Attributability and Foreseeability Bound. No workaround (temporal unrolling, dynamic Bayesian network layer, or separate interaction graph) is indicated. This is load-bearing because the cycle condition defines the regime in which impossibility is claimed; without a consistent model the theorem's domain is empty or the derivation circular.
[Axioms section and theorem statement] Axioms section and theorem statement: the four properties are presented as the 'minimal necessary conditions' for legitimate accountability, yet their concrete formulations (especially Foreseeability Bound tied to the four-dimensional profile and the horizon definition) appear to make the impossibility follow once the horizon and cycle are introduced. The manuscript must show either that the axioms remain independent of the horizon construction or that relaxing any one of them permits a consistent framework below the horizon; otherwise the result risks being an artifact of axiom choice rather than a structural necessity.

minor comments (2)

[Abstract] The abstract is dense; a single sentence clarifying how the four-dimensional autonomy profile is aggregated into 'compound autonomy' would improve accessibility.
[Experiments paragraph] The experimental protocol (how the 3,000 synthetic collectives are generated, how the horizon is computed, and how zero violations are verified) is referenced but not summarized; a brief methods paragraph would strengthen reproducibility claims.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which help strengthen the presentation of our results. We address the two major comments point by point below, indicating the revisions we plan to incorporate.

read point-by-point responses

Referee: [Abstract and modeling paragraphs] Abstract and modeling paragraphs: agents are defined as state-policy tuples inside a shared structural causal model while the theorem is stated only for interaction graphs that contain human-AI feedback cycles. Standard SCMs are directed acyclic graphs; cycles render do-calculus and counterfactuals ill-defined, which are the primitives underlying Attributability and Foreseeability Bound. No workaround (temporal unrolling, dynamic Bayesian network layer, or separate interaction graph) is indicated. This is load-bearing because the cycle condition defines the regime in which impossibility is claimed; without a consistent model the theorem's domain is empty or the derivation circular.

Authors: The referee correctly identifies a potential inconsistency in the causal modeling. While the manuscript introduces the interaction graph as a distinct structure to represent human-AI feedback cycles separately from the shared SCM (which defines individual agent states and policies), we acknowledge that the current text does not explicitly detail how causal inference is preserved under cycles. We will revise the modeling section to specify that cyclic interactions are handled via temporal unrolling of the interaction graph into an acyclic structure for the purposes of applying do-calculus and counterfactual reasoning. This ensures the primitives for Attributability and Foreseeability Bound remain well-defined. The theorem's domain is thus preserved, and we will include this clarification in the revised manuscript. revision: yes
Referee: [Axioms section and theorem statement] Axioms section and theorem statement: the four properties are presented as the 'minimal necessary conditions' for legitimate accountability, yet their concrete formulations (especially Foreseeability Bound tied to the four-dimensional profile and the horizon definition) appear to make the impossibility follow once the horizon and cycle are introduced. The manuscript must show either that the axioms remain independent of the horizon construction or that relaxing any one of them permits a consistent framework below the horizon; otherwise the result risks being an artifact of axiom choice rather than a structural necessity.

Authors: We agree that it is important to demonstrate the independence of the result from the specific axiom choices. The axioms are motivated as minimal conditions drawn from legal and philosophical literature on accountability, and the horizon emerges from their interaction with the autonomy profile. In the experiments, we show that below the horizon, frameworks satisfying all four exist. To further address this, we will add a new subsection in the revised manuscript that explicitly considers relaxations of each axiom individually. For each relaxation, we provide a construction of a consistent framework above the horizon, but argue that the relaxation undermines the legitimacy of the accountability (e.g., relaxing Completeness leaves some outcomes unaccounted for). This will show that the impossibility is a consequence of requiring all four simultaneously, not an artifact. We will also ensure the horizon definition is presented prior to and independently of the theorem. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper introduces a formal model of Human-Agent Collectives via state-policy tuples in a shared structural causal model, defines a four-dimensional autonomy profile, axiomatizes four properties (Attributability, Foreseeability Bound, Non-Vacuity, Completeness) as minimal conditions for legitimate accountability, computes an Accountability Horizon from compound autonomy, and derives the Incompleteness Theorem as a logical incompatibility result for collectives exceeding the horizon with feedback cycles. This chain is self-contained and does not reduce any claimed result to a redefinition of its inputs, a fitted parameter renamed as prediction, or a load-bearing self-citation. The axioms are presented as independent necessary conditions rather than being constructed from the horizon or theorem; the proof establishes a phase transition without circular reduction. No enumerated circularity pattern is exhibited.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 3 invented entities

The central claim rests on the authors' definitions of Human-Agent Collectives, the four-dimensional autonomy profile, interaction graphs, joint action spaces, and the four accountability axioms. No explicit numerical free parameters are stated, but the compound autonomy measure and horizon threshold function as derived quantities whose exact functional form is not provided.

axioms (4)

domain assumption Attributability: responsibility requires causal contribution
One of the four minimal properties axiomatizing legitimate accountability.
domain assumption Foreseeability Bound: responsibility cannot exceed predictive capacity
One of the four minimal properties axiomatizing legitimate accountability.
domain assumption Non-Vacuity: at least one agent bears non-trivial responsibility
One of the four minimal properties axiomatizing legitimate accountability.
domain assumption Completeness: all responsibility must be fully allocated
One of the four minimal properties axiomatizing legitimate accountability.

invented entities (3)

Human-Agent Collectives no independent evidence
purpose: Formal model of joint human-AI systems using state-policy tuples in a shared structural causal model
New formalization introduced to capture collective behavior.
Accountability Horizon no independent evidence
purpose: Computable threshold beyond which the impossibility holds
Central derived quantity in the theorem.
four-dimensional information-theoretic autonomy profile no independent evidence
purpose: Characterizes autonomy along epistemic, executive, evaluative, and social dimensions
Novel characterization used to define compound autonomy.

pith-pipeline@v0.9.0 · 5579 in / 1870 out tokens · 136867 ms · 2026-05-10T17:34:40.139193+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Accountability Incompleteness Theorem: for any HAC whose minimum compound autonomy exceeds the Accountability Horizon Λ* = 1−1/|C_min| and whose interaction graph contains at least one mixed feedback cycle, no framework satisfies Attributability, Foreseeability Bound, Non-Vacuity, and Completeness simultaneously.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

43 extracted references · 28 canonical work pages · 1 internal anchor

[1]

Predicts 2025: Agentic AI—The Evolution of Experience,

Gartner, “Predicts 2025: Agentic AI—The Evolution of Experience,” Tech. rep., Gartner, 2025. Research Report, ID G00811221

2025
[2]

Agentic AI: The future and governance of autonomous systems,

Bloomsbury Intelligence and Security Institute, “Agentic AI: The future and governance of autonomous systems,” Tech. rep., BISI, 2026.https://bisi.org.uk/reports/agentic-ai

2026
[3]

The responsibility gap: Ascribing responsibility for the actions of learning automata,

Matthias, A., “The responsibility gap: Ascribing responsibility for the actions of learning automata,”Ethics and Information Technology, Vol. 6, No. 3, 2004, pp. 175–183. https://doi.org/10.1007/s10676-004-3422-1

work page doi:10.1007/s10676-004-3422-1 2004
[4]

Killer robots,

Sparrow, R., “Killer robots,”Journal of Applied Philosophy, Vol. 24, No. 1, 2007, pp. 62–77. https://doi.org/10.1111/j.1468- 5930.2007.00346.x

work page doi:10.1111/j.1468- 2007
[5]

Four responsibility gaps with artificial intelligence: Why they matter and how to address them,

Santoni de Sio, F., and Mecacci, G., “Four responsibility gaps with artificial intelligence: Why they matter and how to address them,”Philosophy & Technology, Vol. 34, No. 4, 2021, pp. 1057–1084. https://doi.org/10.1007/s13347-021-00450-x

work page doi:10.1007/s13347-021-00450-x 2021
[6]

Artificial intelligence and responsibility gaps,

Königs, P., “Artificial intelligence and responsibility gaps,”Ethics and Information Technology, Vol. 24, No. 3, 2022, p. Article

2022
[7]

https://doi.org/10.1007/s10676-022-09643-0

work page doi:10.1007/s10676-022-09643-0
[8]

There is no AI responsibility gap,

Tigard, D. W., “There is no AI responsibility gap,”Philosophy & Technology, Vol. 34, No. 4, 2021, pp. 685–688. https://doi.org/10.1007/s13347-020-00414-7

work page doi:10.1007/s13347-020-00414-7 2021
[9]

Robots, law and the retribution gap,

Danaher, J., “Robots, law and the retribution gap,”Ethics and Information Technology, Vol. 18, No. 4, 2016, pp. 299–309. https://doi.org/10.1007/s10676-016-9403-3

work page doi:10.1007/s10676-016-9403-3 2016
[10]

On the morality of artificial agents,

Floridi, L., and Sanders, J. W., “On the morality of artificial agents,”Minds and Machines, Vol. 14, No. 3, 2004, pp. 349–379. https://doi.org/10.1023/B:MIND.0000035461.63578.9d. 19

work page doi:10.1023/b:mind.0000035461.63578.9d 2004
[11]

Virtual moral agency, virtual moral responsibility,

Coeckelbergh, M., “Virtual moral agency, virtual moral responsibility,”AI & Society, Vol. 24, No. 2, 2009, pp. 181–189. https://doi.org/10.1007/s00146-009-0208-3

work page doi:10.1007/s00146-009-0208-3 2009
[12]

The moral standing of machines: Towards a relational and non-Cartesian moral hermeneutics,

Coeckelbergh, M., “The moral standing of machines: Towards a relational and non-Cartesian moral hermeneutics,”Philosophy & Technology, Vol. 27, No. 1, 2014, pp. 61–77. https://doi.org/10.1007/s13347-013-0133-8

work page doi:10.1007/s13347-013-0133-8 2014
[13]

Folk understanding of artificial moral agency,

Noh, H., “Folk understanding of artificial moral agency,”Social Robots with AI: Proceedings of Robophilosophy 2024, IOS Press, 2025, pp. 210–218. https://doi.org/10.3233/FAIA241506

work page doi:10.3233/faia241506 2024
[14]

Mind the gap: Autonomous systems, the responsibility gap, and moral entanglement,

Goetze, T. S., “Mind the gap: Autonomous systems, the responsibility gap, and moral entanglement,”Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’22), ACM, 2022, pp. 390–399. https://doi.org/10.1145/3531146.3533106

work page doi:10.1145/3531146.3533106 2022
[15]

J.,Social Choice and Individual Values, Wiley, 1951

Arrow, K. J.,Social Choice and Individual Values, Wiley, 1951

1951
[16]

Fischer, Nancy A

Fischer, M. J., Lynch, N. A., and Paterson, M. S., “Impossibility of distributed consensus with one faulty process,”Journal of the ACM, Vol. 32, No. 2, 1985, pp. 374–382. https://doi.org/10.1145/3149.214121

work page doi:10.1145/3149.214121 1985
[17]

Towards robust distributed systems,

Brewer, E. A., “Towards robust distributed systems,”Proceedings of the 19th ACM Symposium on Principles of Distributed Computing, ACM, 2000. https://doi.org/10.1145/343477.343502

work page doi:10.1145/343477.343502 2000
[18]

Brewer’s conjecture and the feasibility of consistent, available, partition-tolerant web services,

Gilbert, S., and Lynch, N. A., “Brewer’s conjecture and the feasibility of consistent, available, partition-tolerant web services,” ACM SIGACT News, Vol. 33, No. 2, 2002, pp. 51–59. https://doi.org/10.1145/564585.564601

work page doi:10.1145/564585.564601 2002
[19]

Model AI Governance Framework for Agentic AI,

Infocomm Media Development Authority, “Model AI Governance Framework for Agentic AI,” Tech. rep., Government of Singapore, 2026. https://www.imda.gov.sg/

2026
[20]

Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act),

European Parliament and Council of the European Union, “Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act),”Official Journal of the European Union, Vol. L 2024/1689, 2024

2024
[21]

Presidio AI Framework: Towards Safe and Beneficial AI,

World Economic Forum, “Presidio AI Framework: Towards Safe and Beneficial AI,” Tech. rep., WEF, 2025. https: //www.weforum.org/

2025
[22]

Impossibility results in AI: A survey,

Brcic, M., and Yampolskiy, R. V., “Impossibility results in AI: A survey,”ACM Computing Surveys, Vol. 56, No. 1, 2023, p. Article 8. https://doi.org/10.1145/3603371

work page doi:10.1145/3603371 2023
[23]

Impossibility and uncertainty theorems in AI value alignment,

Eckersley, P., “Impossibility and uncertainty theorems in AI value alignment,”Proceedings of the AAAI Workshop on Artificial Intelligence Safety, 2019. ArXiv:1901.00064

work page arXiv 2019
[24]

On the Arrowian impossibility of machine intelligence measures,

Oswald, J. T., Ferguson, T. M., and Bringsjord, S., “On the Arrowian impossibility of machine intelligence measures,”Artificial General Intelligence. AGI 2025, LNCS, Vol. 16058, Springer, 2026. https://doi.org/10.1007/978-3-032-00800-8_3

work page doi:10.1007/978-3-032-00800-8_3 2025
[25]

Limitations on Accurate, Trusted, Human-level Reasoning

Panigrahy, R., “Limitations on safe, trusted, artificial general intelligence,”arXiv preprint, 2025. ArXiv:2509.21654

work page internal anchor Pith review arXiv 2025
[26]

Agent behavioral contracts: Probabilistic compliance guarantees for autonomous agents,

Bhardwaj, A., “Agent behavioral contracts: Probabilistic compliance guarantees for autonomous agents,”Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40, AAAI Press, 2026

2026
[27]

Kaptein, V.-J

Rodríguez-Fernández,V.,etal.,“RuntimegovernanceforAIagents: Policiesonpaths,”arXivpreprint,2026. ArXiv:2603.16586

work page arXiv 2026
[28]

Code-based measurement of AI agent autonomy levels,

Cihon, P., Schuett, J., and Baum, S. D., “Code-based measurement of AI agent autonomy levels,”AI and Ethics, 2025. https://doi.org/10.1007/s43681-025-00612-3

work page doi:10.1007/s43681-025-00612-3 2025
[29]

Causation in tort law,

Wright, R. W., “Causation in tort law,”California Law Review, Vol. 73, No. 6, 1985, pp. 1735–1828. https://doi.org/10.2307/ 3480373

1985
[30]

Hart, H. L. A., and Honoré, T.,Causation in the Law, Oxford University Press, 1959. 2nd ed., 1985

1959
[31]

The limits of the preponderance of the evidence standard,

Kaye, D., “The limits of the preponderance of the evidence standard,”American Bar Foundation Research Journal, Vol. 7, No. 2, 1982, pp. 487–516. https://doi.org/10.1111/j.1747-4469.1982.tb00460.x

work page doi:10.1111/j.1747-4469.1982.tb00460.x 1982
[32]

Responsibility and Control: A Theory of Moral Responsibility,

Fischer, J. M., and Ravizza, M., “Responsibility and Control: A Theory of Moral Responsibility,” 1998. https://doi.org/10. 1017/CBO9780511814594

1998
[33]

Analysing and assessing accountability: A conceptual framework.European Law Journal, 13(4):447–468, 2007

Bovens, M., “Analysing and assessing accountability: A conceptual framework,”European Law Journal, Vol. 13, No. 4, 2007, pp. 447–468. https://doi.org/10.1111/j.1468-0386.2007.00378.x. 20

work page doi:10.1111/j.1468-0386.2007.00378.x 2007
[34]

Pathologies of accountability: ICANN and the challenge of “multiple accountabilities disorder

Koppell, J. G. S., “Pathologies of accountability: ICANN and the challenge of “multiple accountabilities disorder”,”Public Administration Review, Vol. 65, No. 1, 2005, pp. 94–108. https://doi.org/10.1111/j.1540-6210.2005.00434.x

work page doi:10.1111/j.1540-6210.2005.00434.x 2005
[35]

https://doi.org/10.1017/CBO9780511807763

Ostrom, E.,Governing the Commons: The Evolution of Institutions for Collective Action, Cambridge University Press, 1990. https://doi.org/10.1017/CBO9780511807763

work page doi:10.1017/cbo9780511807763 1990
[36]

Ostrom, E.,Understanding Institutional Diversity, Princeton University Press, 2005

2005
[37]

https://doi.org/10.1017/ CBO9780511803161

Pearl, J.,Causality: Models, Reasoning, and Inference, 2nd ed., Cambridge University Press, 2009. https://doi.org/10.1017/ CBO9780511803161

2009
[38]

Spirtes, P., Glymour, C., and Scheines, R.,Causation, Prediction, and Search, 2nd ed., MIT Press, 2000

2000
[39]

Foundations of structural causal models with cycles and latent variables,

Bongers, S., Forré, P., Peters, J., and Mooij, J. M., “Foundations of structural causal models with cycles and latent variables,” Annals of Statistics, Vol. 49, No. 5, 2021, pp. 2885–2915. https://doi.org/10.1214/21-AOS2064

work page doi:10.1214/21-aos2064 2021
[40]

Finding all the elementary circuits of a directed graph,

Johnson, D. B., “Finding all the elementary circuits of a directed graph,”SIAM Journal on Computing, Vol. 4, No. 1, 1975, pp. 77–84. https://doi.org/10.1137/0204007

work page doi:10.1137/0204007 1975
[41]

Theory of capacities,

Choquet, G., “Theory of capacities,”Annales de l’Institut Fourier, Vol. 5, 1954, pp. 131–295. https://doi.org/10.5802/aif.53

work page doi:10.5802/aif.53 1954
[42]

A value for𝑛-person games,

Shapley, L. S., “A value for𝑛-person games,”Contributions to the Theory of Games, Vol. II, Princeton University Press, 1953, pp. 307–317

1953
[43]

Estimating Mutual Information

Kraskov, A., Stögbauer, H., and Grassberger, P., “Estimating mutual information,”Physical Review E, Vol. 69, No. 6, 2004, p. 066138. https://doi.org/10.1103/PhysRevE.69.066138. 21 Appendix A. Full Proofs A. Corollary 2 (Agent-to-Agent Dilution): Proof The bound is derived via the interventional Data Processing Inequality. Define the causal mutual informat...

work page doi:10.1103/physreve.69.066138 2004