Diffusion of Responsibility in Collective Decision Making

Jia Tao; Pavel Naumov

REVIEW 2 major objections 4 minor 22 references

Diffusion of Responsibility in Collective Decision Making

T0 review · 2 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper proves that sequential collective decision mechanisms that avoid diffusion of responsibility must be elected dictatorships, and with two agents, outright dictatorships.

desk verdict The elected-dictatorship characterization is a plausible and interesting result, but the proof has a real gap in the bisimulation reduction that needs fixing before the theorem is citable as proven. read the letter →

arxiv 2506.07935 v1 pith:3ZMGMFQI submitted 2025-06-09 cs.MA cs.AIcs.GT

Pavel Naumov , Jia Tao This is my paper

classification cs.MAcs.AIcs.GT MSC 91A8091A18

keywords diffusionofresponsibilitycounterfactualcollectivedecisionmakingextensive-formgameselecteddictatorshipbisimulationmechanismdesignaccountability

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks whether a collective decision-making mechanism can guarantee that at most one agent is ever accountable for an outcome, avoiding the phenomenon called diffusion of responsibility. The authors prove an impossibility: for mechanisms in which agents move one after another with full information, any diffusion-free mechanism must be an elected dictatorship, meaning the mechanism selects a single agent who unilaterally decides among all alternatives. With exactly two agents, the only diffusion-free mechanisms are outright dictatorships. The result matters because it reveals a hard trade-off between diffused accountability and concentrated power in sequential decision procedures.

What carries the argument

The machinery is a bisimulation relation between two mechanisms, defined as a relation between nodes of their trees that matches leaf labels, preserves which nodes an agent can reach through edges labeled by the same agent, and matches ancestors of related nodes. A key lemma states that if $n \mathrel{R} n'$ then $n \in \mathrm{win}_a(S)$ iff $n' \in \mathrm{win}_a(S)$ for every alternative set $S$, so bisimilar nodes give every agent the same strategic power. The proof then restricts to a canonical form, a smallest mechanism totally bisimilar to the given one, and uses structural lemmas about children, heights, and alternative labels to show that each leaf's parent is a dictator. This canonical-form analysis is what carries the theorem.

What would settle it

Enumerate all finite rooted trees with two agents, two alternatives, surjective leaf labeling, and sequential perfect-information moves, and search for one in which no leaf has two responsible agents but no agent can force both alternatives from the root; finding even one such mechanism would disprove Theorem 2.

Watch

Extended reading notes

Core claim

The central claim is Theorem 1: every diffusion-free mechanism is an elected dictatorship. An agent is counterfactually responsible at a leaf if some ancestor node is one from which the agent has a strategy to prevent the leaf's alternative; diffusion is a leaf with two responsible agents; a dictator at a node can force every alternative from that node. An elected dictatorship is a mechanism in which every leaf has a dictator among its ancestors. Theorem 2 states that with two agents, any diffusion-free mechanism is a dictatorship, so one agent has an upfront strategy for every outcome. The proof works by bisimulation: it shows that responsibility, diffusion, and elected dictatorship are invariant under total bisimulation, then analyzes a smallest bisimular (canonical) mechanism and forces its structure so that the parent of every leaf is a dictator.

Load-bearing premise

The load-bearing premise is the paper's definition of responsibility: an agent counts as responsible for an outcome only if that agent had a unilateral strategy, somewhere along the decision path, to prevent it; if responsibility is instead defined through actual causation or probabilistic influence, the impossibility results would not necessarily hold.

Editorial extensions

If this is right

If Theorem 1 is correct, any mechanism that avoids diffusion of responsibility necessarily contains, for each leaf, an ancestor at which one agent can unilaterally determine every alternative.
With two agents, avoiding diffusion means one agent is a full dictator, so every two-agent sequential procedure that gives both agents a meaningful role must admit some leaf where two agents are responsible.
Because responsibility, diffusion, and elected dictatorship are preserved under total bisimulation, these accountability properties are behavioral rather than tied to a particular syntactic presentation of a mechanism.
For mechanism designers, this implies that accountability and power-sharing cannot both be achieved by a sequential, perfect-information procedure; one of those requirements must be relaxed.
The two-agent result extends the impossibility to the simplest case, ruling out any clever two-agent sequential protocol that avoids shared blame without concentrating power.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the impossibility is tied to sequential moves and perfect information; simultaneous or hidden-action mechanisms such as secret-ballot voting may escape the trade-off precisely because no agent has a unilateral strategy at the moment of choice.
Editorial inference: if responsibility is instead assigned by actual causation or by expected influence, a diffusion-free mechanism need not be an elected dictatorship, so the central result is sensitive to the philosophical interpretation of 'could have done otherwise.'
Editorial inference: deployed sequential decision pipelines could be audited by looking for bisimilar reductions; any diffusion-free core found this way would expose a single decision-making node responsible for each outcome.
Editorial inference: a natural quantitative extension is to measure the degree of diffusion by the number of responsible agents per leaf, which would produce a graded version of the theorem rather than a binary one.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper has a good core idea. It defines counterfactual responsibility in sequential decision mechanisms cleanly, shows that two-agent diffusion-free mechanisms are dictatorships, and extends this to elected dictatorship for any number of agents. The bisimulation technique is a nice way to collapse mechanisms to a canonical form, and the historical examples (US Constitution, Articles of Confederation, colonial vetoes) are genuinely illuminating. The appendix contains detailed proofs of the win-set lemma and the canonical-form structural lemmas.

The problem is the reduction step at the start of Theorem 1. The proof lets M' be a smallest mechanism totally bisimular to M, then says 'by Lemma 8 and Lemma 9' it suffices to prove the result for M'. Lemma 8 is about diffusion: if one of two totally bisimular mechanisms allows diffusion, so does the other. Lemma 9 makes the same claim for elected dictatorship. But in both cases the proof only establishes one direction. Lemma 7 transfers responsibility from a node in M to a bisimular node in M'; it does not transfer responsibility back. Similarly, Lemma 6 transfers 'every node has a dictator at an ancestor' from M to M', not from M' to M. The canonical-form reduction needs the reverse direction: you know M is diffusion-free and want M' to be diffusion-free, and later you want to lift elected dictatorship from M' back to M. Because condition 4 of Definition 10 (the ancestor condition) is asymmetric, the inverse of the bisimulation is not automatically a bisimulation, and the missing direction does not follow from the stated lemmas. So the reduction to canonical form is not justified as written.

This is not a merely cosmetic gap; it is the pivot of the proof. That said, I think it is fixable. Either add the symmetric ancestor condition to the definition of bisimulation (so the inverse of a total bisimulation is also a total bisimulation), or prove the converse transfer lemmas directly. The rest of the machine—the canonical-form lemmas that force every leaf's parent to be a height-1 node with all alternatives as children—looks credible, and I found no counterexample to the theorem itself. So my position is: the conclusion is likely true, but the proof is incomplete.

The paper is for people working on responsibility in multi-agent systems, formal epistemology, and mechanism design. It deserves a serious referee, but the authors should be asked to fix the transfer direction before acceptance. The gap is not desk-reject territory; it is revision territory.

Referee Report

2 major / 4 minor

Summary. The paper studies sequential, perfect-information, deterministic collective decision-making mechanisms and defines counterfactual responsibility as the existence of a unilateral strategy, from some ancestor node, to force an alternative different from the realized outcome. Diffusion of responsibility occurs when two agents are responsible at the same leaf. The main results are Theorem 1, stating that every diffusion-free mechanism is an elected dictatorship, and Theorem 2, stating that every diffusion-free two-agent mechanism is a dictatorship. The proof strategy introduces a bisimulation on mechanism trees, shows that bisimulation preserves several responsibility-related properties, reduces to a node-minimal totally bisimular mechanism, and establishes the theorem for such canonical mechanisms.

Significance. If established, the main theorem is a clean and surprising impossibility result: avoiding every instance of shared counterfactual responsibility in a sequential mechanism forces decision power to be concentrated in a single elected agent. The paper is self-contained, introduces a bisimulation framework that is a genuine methodological contribution, and makes no use of fitted parameters or normalization tricks. The historical examples help make the formal definitions concrete. However, the proof of the central reduction is currently incomplete, so the significance is conditional on repairing the technical gaps described below.

major comments (2)

[Properties of a Bisimulation, Lemmas 6-9; Theorem 1 proof] The reduction step in Theorem 1 requires preservation in both directions: M diffusion-free implies M' diffusion-free, and M' elected dictatorship implies M elected dictatorship. Lemma 6 and Lemma 7 are proved in only one direction (from tree T to tree T'), and Definition 10 item 4 is asymmetric. Lemma 8 is stated as a symmetric preservation result but its proof only cites Lemma 7 and Definition 11, and Lemma 9 similarly cites Lemma 6. Neither proof establishes the converse direction needed for the canonical-form reduction. As written, the theorem is not proved. Please either add the missing reverse ancestor clause to Definition 10 and prove the iff-versions of Lemmas 5-9, or provide a direct argument that the chosen minimal bisimular mechanism inherits diffusion-freeness from M and transfers elected dictatorship back to M.
[Canonical Form, Lemmas 10, 11, 13, 16] The collapse arguments in these lemmas are presented through diagrams with 'the bisimulation shown by dashed lines' and a statement that unshown nodes are connected to their clones. The relation R is not fully specified, and the four clauses of Definition 10, especially the asymmetric ancestor clause, are not verified. These lemmas are load-bearing for Lemma 18 and Lemma 20, which are used in Theorem 1. Please provide explicit definitions of R and detailed checks of all four clauses for each collapse construction.

minor comments (4)

[Lemma 20 proof] The notations 'AltTree(r)' and 'TreeAlt(r)' are used inconsistently; the symbol defined earlier in the paper is 'TreeAlt'.
[Theorem 1 proof] The final sentence says 'mechanismm' is an elected dictatorship' instead of 'mechanism M'.'
[Definition 2] The phrase 'minimal set' is informal; since the three closure rules are monotone, the least fixed point exists, but this could be stated explicitly for completeness.
[General presentation] The convention of using 'if' instead of 'iff' in definitions is nonstandard and can be confusing; explicitly saying 'if and only if' would improve clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's characterization is proven by an independent bisimulation argument, and the cited prior interpretation of responsibility is not load-bearing.

full rationale

The derivation chain is not circular. The core result, Theorem 1, is established by defining bisimulation of decision mechanisms and then proving, via Lemmas 4-9, that bisimulation preserves responsibility, diffusion-freeness, and elected dictatorship. These preservation lemmas are proven directly from Definitions 2, 3, 5, 7, and 10; they do not assume the theorem they are used to prove. The reduction to a canonical form and the subsequent proof that every diffusion-free canonical mechanism is an elected dictatorship are substantive and depend on the structural lemmas, not on a restatement of the conclusion. The only self-citation is to Naumov and Tao 2019 as one of several works supporting the modeling choice that 'could have done otherwise' means having a unilateral strategy to prevent the outcome. That is a definitional premise, not an imported theorem that forces the result, and it is corroborated by multiple independent references. The reader's noted asymmetry concern about Lemmas 6/7 is a proof-correctness issue, not a circularity issue: even if the converse directions require more argument, the proof does not reduce to its own inputs by construction. There are no fitted parameters, no predictions derived from fitted values, and no renaming of a known empirical pattern as a new result.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central theorem is a mathematical implication within the explicit model; the only costs are the modeling assumptions about what counts as a mechanism and what counts as responsibility. No free parameters or invented empirical entities are introduced.

assumptions (4)

domain assumption Decision mechanisms are finite rooted trees with deterministic perfect-information sequential moves (Definition 1).
The entire characterization is restricted to this class; concurrent or imperfect-information mechanisms are excluded.
domain assumption Counterfactual responsibility is defined as having a unilateral strategy to prevent the outcome (Definitions 2 and 3).
This interprets Frankfurt's principle in a game-theoretic way; the theorem's meaning depends on this choice.
domain assumption Leaf labeling ℓ is surjective onto Alt (Definition 1).
Eliminates fictitious alternatives and enables Lemma 20.
standard math Standard determinacy of finite perfect-information win-lose games (Bonanno 2018, Theorem 3.5.1).
Used in the proof of Theorem 2 to claim one agent has a winning strategy.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Diffusion of Responsibility in Collective Decision Making." pith.science (2026). https://pith.science/paper/3ZMGMFQI

@misc{pith2026250607935,
  author       = {Pith},
  title        = {Pith review of: Diffusion of Responsibility in Collective Decision Making},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3ZMGMFQI}},
  note         = {Machine review of arXiv:2506.07935}
}

read the original abstract

The term "diffusion of responsibility'' refers to situations in which multiple agents share responsibility for an outcome, obscuring individual accountability. This paper examines this frequently undesirable phenomenon in the context of collective decision-making mechanisms. The work shows that if a decision is made by two agents, then the only way to avoid diffusion of responsibility is for one agent to act as a "dictator'', making the decision unilaterally. In scenarios with more than two agents, any diffusion-free mechanism is an "elected dictatorship'' where the agents elect a single agent to make a unilateral decision. The technical results are obtained by defining a bisimulation of decision-making mechanisms, proving that bisimulation preserves responsibility-related properties, and establishing the results for a smallest bisimular mechanism.

Figures

Figures reproduced from arXiv: 2506.07935 by the authors.

**Figure 1.** US Constitution mechanism President, it becomes a law of the United States. If the President vetoes the bill, it goes back to Congress. At that point, Congress can either override the veto or kill the bill. Imagine that a bill to increase funding for AI research is introduced in Congress. Congress passes the bill and sends it to the President. The President decides to veto the bill. When the bill comes back to Cong… view at source ↗

**Figure 4.** The other 10 colonies bill would not become law, but none of them had an upfront strategy to make it into law. The mechanism depicted in [PITH_FULL_IMAGE:figures/full_fig_p002_4.png] view at source ↗

**Figure 2.** Articles of Confederation However, [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

Figures from the paper (3 more)

**Figure 3.** Figure 3: Maryland, Rhode Island, and Connecticut Are there decision-making mechanisms that do not have a dictator? Yes, there are. In fact, such mechanisms have been used by the 13 colonies that preceded the Confederation. Three of them used the mechanism depicted in [PITH_FUL…

**Figure 5.** Figure 5: Hypothetical mechanism this figure is what we call an elected dictatorship. Intuitively, an elected dictatorship is a mechanism that selects a single agent to be the “decision-maker” and this agent decides between all available alternatives. In the example depicted in…

**Figure 6.** Figure 6: Dashed lines show a bisimulation of two decision [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 20 canonical work pages

[1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Baier, C.; Funke, F.; and Majumdar, R. 2021. A Game-Theoretic Account of Responsibility Allocation. In 30th International Joint Conference on Artificial Intelligence (IJCAI-21)

work page 2021
[4]

Batusov, V.; and Soutchanski, M. 2018. Situation Calculus Semantics for Actual Causality. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence (AAAI-18)

work page 2018
[5]

Bleher, H.; and Braun, M. 2022. Diffused responsibility: attributions of responsibility in the use of A I -driven clinical decision support systems. AI and Ethics, 2(4): 747--761

work page 2022
[6]

Bonanno, G. 2018. Game Theory. Kindle Direct Publishing

work page 2018
[7]

Cushman, F. 2015. Deconstructing intent to reconstruct morality. Current Opinion in Psychology, 6: 97--103

work page 2015
[8]

Feng, C.; Deshpande, G.; Liu, C.; Gu, R.; Luo, Y.-J.; and Krueger, F. 2016. Diffusion of responsibility attenuates altruistic punishment: A functional magnetic resonance imaging effective connectivity study. Human brain mapping, 37(2): 663--677

work page 2016

Show all 22 references

[9]

R.; Zyzniewski, L

Forsyth, D. R.; Zyzniewski, L. E.; and Giammanco, C. A. 2002. Responsibility diffusion in cooperative collectives. Personality and Social Psychology Bulletin, 28(1): 54--65

2002
[10]

Frankfurt, H. G. 1969. Alternate possibilities and moral responsibility. The Journal of Philosophy, 66(23): 829--839

1969
[11]

Halpern, J. Y. 2016. Actual causality. MIT Press

2016
[12]

Iusmen, I. 2020. Whose children? Protecting unaccompanied migrant children in Europe: A case of diffused responsibility? The International Journal of Children's Rights, 28(4): 925--949

2020
[13]

Lewis, D. 2013. Counterfactuals. John Wiley & Sons

2013
[14]

Liu, D.; Liu, X.; and Wu, S. 2022. A Literature Review of Diffusion of Responsibility Phenomenon. In 2022 8th International Conference on Humanities and Social Science Research (ICHSSR 2022), 1806--1810. Atlantis Press

2022
[15]

Mynatt, C.; and Sherman, S. J. 1975. Responsibility attribution in groups and individuals: A direct test of the diffusion of responsibility hypothesis. Journal of Personality and Social Psychology, 32(6): 1111

1975
[16]

Naumov, P.; and Tao, J. 2019. Blameworthiness in Strategic Games. In Proceedings of Thirty-third AAAI Conference on Artificial Intelligence (AAAI-19)

2019
[17]

R.; Kan, E.; Frick, P

Rowan, Z. R.; Kan, E.; Frick, P. J.; and Cauffman, E. 2022. Not (entirely) guilty: The role of co-offenders in diffusing responsibility for crime. Journal of Research in Crime and Delinquency, 59(4): 415--448

2022
[18]

Sangiorgi, D. 2011. Introduction to bisimulation and coinduction. Cambridge University Press

2011
[19]

Shi, Q. 2024. Responsibility in Extensive Form Games. In Proceedings of 38th AAAI Conference on Artificial Intelligence (AAAI-24)

2024
[20]

Watson, R. A. 1987. Origins and early development of the veto power. Presidential Studies Quarterly, 401--412

1987
[21]

Widerker, D. 2017. Moral responsibility and alternative possibilities: Essays on the importance of alternative possibilities. Routledge

2017
[22]

Yazdanpanah, V.; Dastani, M.; Jamroga, W.; Alechina, N.; and Logan, B. 2019. Strategic responsibility under imperfect information. In Proceedings of the 18th International Conference on Autonomous Agents and MultiAgent Systems, 592--600. International Foundation for Autonomous...

2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Baier, C.; Funke, F.; and Majumdar, R. 2021. A Game-Theoretic Account of Responsibility Allocation. In 30th International Joint Conference on Artificial Intelligence (IJCAI-21)

work page 2021

[4] [4]

Batusov, V.; and Soutchanski, M. 2018. Situation Calculus Semantics for Actual Causality. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence (AAAI-18)

work page 2018

[5] [5]

Bleher, H.; and Braun, M. 2022. Diffused responsibility: attributions of responsibility in the use of A I -driven clinical decision support systems. AI and Ethics, 2(4): 747--761

work page 2022

[6] [6]

Bonanno, G. 2018. Game Theory. Kindle Direct Publishing

work page 2018

[7] [7]

Cushman, F. 2015. Deconstructing intent to reconstruct morality. Current Opinion in Psychology, 6: 97--103

work page 2015

[8] [8]

Feng, C.; Deshpande, G.; Liu, C.; Gu, R.; Luo, Y.-J.; and Krueger, F. 2016. Diffusion of responsibility attenuates altruistic punishment: A functional magnetic resonance imaging effective connectivity study. Human brain mapping, 37(2): 663--677

work page 2016