When groups spread across time zones deliberate through AI systems, whether an AI mediator drafting a common statement or AI agents speaking for absent people, the output tends to converge on bland agreement. This note argues that the most valuable thing lost is a specific kind of contribution that organizational psychologists call prohibitive voice: pointed warnings that something is wrong or risky, as opposed to promotive voice, which offers suggestions for improvement. Survey research on employees finds that prohibitive voice is fragile in a particular way. It appears only when a person has internalized the group as part of who they are and feels able to act on it. Current AI deliberation pipelines may attack both conditions. The machine side compresses minority objections toward the model's average view. The human side, by handing authorship to an agent, may remove the act of owning one's objection that prohibitive voice depends on.
We propose to measure how much valid dissent survives an AI-mediated deliberation pipeline and to separate these two causes experimentally. The design crosses two factors. The first is how objections are aggregated: a single consensus statement, or a consensus statement plus a separately kept register of unresolved objections. The second is who authors them: an autonomous agent, or a human who must personally endorse each objection the agent drafts. The first testbed is peer review, where reviewers' objections and authors' rebuttal concessions give ground truth for which objections were valid. The second is a randomized study with geographically distributed human committees. The two-sentence version: AI deliberation tools may silence exactly the warnings groups most need, through two separable routes, one algorithmic and one psychological. We build the instruments to tell which route matters and test a design that addresses both.
Core Claim
The claim is that dissent loss in delegated distributed deliberation is an "organizational black box" problem. That literature studies the common finding that organizational inputs have weak direct effects on outcomes and that the effect runs through intermediate psychological states. Here the input is platform design and the outcome is surviving valid objections. We hypothesize two stacked mechanisms. Machine-side compression depends on the aggregator. Human-side withdrawal runs through psychological empowerment and group identification. These make different predictions. If compression dominates, a better aggregator restores dissent regardless of who authors it. If withdrawal dominates, human-endorsed objections survive better even under the same aggregator, and the gain is carried by measured empowerment and identification. The contribution is the measurement (valid-objection retention, split into prohibitive and promotive content) and the factorial test. It is not another consensus algorithm.
Grounding in Prior Work
Three survey studies from the self-concept internalization literature motivate the human-side hypothesis.
- A chain mediation model of inclusive leadership and voice behavior among university teachers surveyed 517 teachers. Leadership's effect on prohibitive voice ran fully through psychological empowerment and organizational identification, with the chain carrying 27.21% of the effect. Promotive voice kept a direct path.
- How Distributed Leadership Fosters Individual Leadership Emergence showed with 496 dyads over three waves that shared leadership works by first building an identity of being empowered and then an enacted leader identity, meaning one acquired by acting as a leader.
- The relationship between affective leadership and the voice behavior of nurses found no direct effect of leadership on voice. The effect ran through relational ties and appeared only in groups with a high voice climate.
All three used co-located Chinese organizations, cross-sectional or panel surveys, and voice rated as a static score. We borrow their constructs and validated scales. We change three things: we randomize the upstream input, we observe voice as actual text in a deliberation, and we study dispersed participants.
Four papers from the deliberation literature motivate the machine-side hypothesis and constrain the design.
- HABERMOLT: Delegating Deliberation to AI Representatives deployed AI representatives across 140 deliberations. Mean pairwise opinion similarity was 0.745 for autonomous agents against 0.649 for opinions drawn from human interviews, a drift toward the model's prior. Only 8 of 91 users ever revised their agent. The authors also found that representative statements are vague while actionable ones alienate factions. They treat low revision as an interface and memory problem. We test whether it is partly a loss of ownership.
- Can AI mediation improve democratic deliberation? describes the Habermas Machine, which drafts candidate group statements, predicts each participant's ranking, and picks one by Schulze voting. This is the design we use as the consensus-only aggregator.
- Voting or Consensus? Decision-Making in Multi-Agent Debate found that forced consensus hurts reasoning tasks badly (MuSR 28.4% against 54.8% for voting). Consensus pressure plausibly destroys correct minority positions, which is a warning for any consensus-only design.
- EuroCon found that model mediators favor the dominant party and succeed on only 52 to 63% of two-thirds-majority tasks. It suggests that minority objections are the ones at risk.
Deliberation via Matching proves that pairing opposed voters improves worst-case social-choice guarantees. It assumes preferences are fixed and averaged, which is exactly the psychology we propose to measure, so it bounds rather than answers our question.
Research Program
The first component is an offline measurement of machine-side retention on peer review, where validity can be checked.
From the Re² dataset, which pairs 70,668 review comments with 53,818 rebuttal turns from OpenReview venues, we extract each reviewer's objections. We label them prohibitive (a claimed flaw or risk) or promotive (a suggested improvement). We mark an objection as a minority objection when other reviewers do not raise it or contradict it. For contradiction detection we use a RoBERTa-large classifier trained on ContraSciView, which the file reports at 77.55% F1. An objection counts as validated when the authors concede it in the rebuttal, which we judge with a human-audited model annotator.
We then give the full review set for each paper to each aggregator and ask for a meta-review. The main metric is validated-minority-objection recall, reported separately for prohibitive and promotive objections. A false-objection rate controls for aggregators that simply keep everything.
EuroCon supplies a second setting in which seat share is varied, so we can test whether small-faction objections are lost faster.
The second component is a randomized human study of distributed committees. We recruit researchers from at least three widely separated regions into mock program committees. They deliberate asynchronously over borderline submissions, following the procedure of the ISMAR 2019 three-site committee, which had 68 borderline papers among 163.
Each participant is represented by a HABERMOLT-style agent built from an interview. Participants are randomized into three authorship arms:
- The agent submits autonomously, with revision possible, which reproduces the HABERMOLT default.
- The agent drafts, and the participant must accept, edit, or reject each objection before it counts. We call this enacted authorship.
- Participants write their own objections without an agent.
Committees are separately randomized to the consensus-only aggregator or to consensus plus an objection register. In the register, each objection survives unless it is rebutted on the merits. Popularity does not remove it.
Before and after, we administer the empowerment and organizational-identification scales used by Liu et al., adapted to identification with the committee. We also collect a short voice-climate measure. Committee members are dispersed across home institutions, so their climates differ, as Wang et al. suggests matters. We therefore test whether home-institution climate or committee climate moderates the effect.
Planted flaws give each submission known ground-truth objections, so recall of valid prohibitive objections can be scored. The behavioral outcomes are revision rate and time spent reviewing objections.
Baselines
The naive reference is a single frontier model that reads all reviews, or all participant inputs, and writes one summary. It shows how much dissent a system with no deliberation structure keeps. The file reports a known weakness here: zero-shot LLMs trail fine-tuned contradiction detectors by 9 to 12 F1 on ContraSciView.
The strongest existing aggregator is the Habermas Machine pipeline from Can AI mediation improve democratic deliberation?. It generates candidates, predicts each participant's ranking, and selects by Schulze voting. We found no reported number on Lacuna for its endorsement or win rates, so we will re-implement it and establish its retention scores ourselves. It is the strongest competitor because it was designed and tested for common-ground finding with humans.
The strongest delegation baseline is the HABERMOLT autonomous-agent condition from HABERMOLT. Its reported numbers are opinion similarity of 0.745, against 0.649 for human interviews, and a revision rate of 8 of 91 users. Our enacted arm beats it if it restores opinion diversity to at or below the human-interview level and raises validated prohibitive-objection recall.
The voting baseline is majority voting over agent positions. Kaesberg et al. found it outperforms consensus by 13.2 points on average on reasoning tasks.
The explicit-dissent prompt baseline instructs the mediator to include a minority report.
Fairness and tuning. All aggregators use the same base model, the same inputs, and the same generation budget. Each gets equal prompt-tuning effort on a held-out development split of Re².
Ablations. The objection-register ablation removes the register while keeping enacted authorship. The endorsement ablation keeps the register but lets agents submit autonomously. Together they isolate the two ingredients.
What counts as beating a baseline. Offline, a method must show higher validated prohibitive-objection recall at no higher false-objection rate than the best baseline. In the human study, it must show a significant authorship-by-outcome effect on valid prohibitive objections.
Why not just…?
Why not just prompt the mediator to include minority views? This is the most serious threat to the machine-side story, and it is a baseline for that reason. If a minority-report prompt matches the register on validated recall without raising false objections, we drop the register and report that compression is a prompting problem. A prompt still cannot address the human side: if autonomous agents never generated the objection, no instruction to the mediator recovers it. The human-study arms with a fixed aggregator test that directly.
Why not just fix the interface so users revise more, as HABERMOLT's authors suggest? We add a notification-and-nudge variant of the autonomous arm. If nudges raise revision rates and valid prohibitive objections to the enacted-arm level without any change in empowerment or identification, the identity mechanism is unnecessary and we will say so. Survey evidence that prohibitive voice specifically needs internalization (Liu et al.) makes us expect nudges to raise promotive edits more than prohibitive ones. That is a falsifiable asymmetry.
Why not just use voting instead of consensus? Voting helps reasoning tasks (Kaesberg et al.). But a majority vote discards a single reviewer's correct objection by construction, and EuroCon shows minority positions are the ones lost. Voting is included as a baseline. We expect it to beat consensus on overall accuracy but not on validated minority-objection recall.
Why not just survey distributed teams about identity and voice? That reproduces the cross-sectional, self-report design whose common-method bias the black-box literature already acknowledges. It also cannot observe which objections the platform discarded. Randomizing authorship and scoring actual text against planted flaws is what makes the mechanism testable.
Proposed Experiments and Decisive Checks
Offline check. Consensus-only aggregators should keep validated minority prohibitive objections at a clearly lower rate than validated minority promotive ones on Re². The gap should widen with faction imbalance on EuroCon. If no such asymmetry appears, the promotive/prohibitive distinction does not transfer to machine aggregation and the offline component fails.
Human study: factorial check. In the 3-by-2 design, main effects of aggregator and authorship on validated prohibitive-objection recall tell us which mechanism dominates.
Human study: mediation check. For the path from authorship through empowerment and identification to prohibitive objections, the first link is causal because authorship is randomized. The link from the psychological states to objections remains observational, and we will report it as such. The prediction specific to our account is that enacted authorship raises prohibitive but not promotive objection counts, carried by identification.
Climate check. We test whether committee voice climate or home-institution climate moderates the authorship effect. This speaks to the open question of which climate governs dispersed members.
Stopping rule. If the dissent prompt closes the offline gap and nudges close the human gap without psychological mediation, the conclusion is that dissent loss is an engineering problem. That conclusion is also worth publishing.
References
- A chain mediation model of inclusive leadership and voice behavior among university teachers: evidence from China
- How Distributed Leadership Fosters Individual Leadership Emergence: The Mediating Role of Empowerment Role Identity and Enacted Leader Identity
- The relationship between affective leadership and the voice behavior of nurses: a cross-level moderated mediation model
- HABERMOLT: Delegating Deliberation to AI Representatives
- Can AI mediation improve democratic deliberation?
- Voting or Consensus? Decision-Making in Multi-Agent Debate
- EuroCon: Benchmarking Parliament Deliberation for Political Consensus Finding
- Deliberation via Matching
- Re2: A Consistency-ensured Dataset for Full-stage Peer Review and Multi-turn Rebuttal Discussions
- When Reviewers Lock Horns: Finding Disagreements in Scientific Peer Reviews
- Message from the ISMAR 2019 Science and Technology Program Chairs and TVCG Guest Editors