PaperlayerView exact revision on arXiv

Source-linked AI summary

AI Mediators Regulate Emotion and Create Value in Disputes

James Hale, Jonathan Gratch

arXiv:2609.17933v1cs.HC

TL;DR

Disputes can make compromise difficult, and little prior work tested whether LLM mediators regulate emotion and facilitate joint gains in heated disputes against novice humans. This study compares AI, human, and no mediation in a dispute task, finding AI mediators reduced negative emotion and may help realize joint gains when integrative potential is high.

  • Problem

    Prior LLM-mediation work had rarely tested emotion regulation and joint gains in heated disputes against novice humans, and had no mediation baselines.

    Source: C. AI Mediation · p. 2

  • Method

    The study compares AI, human, and no mediation in a dispute task, with human and LLM mediators given the same instructions.

    Source: 2) Mediator Perspective · p. 3, 2) Pre-Task · p. 3

  • Results

    AI mediators significantly reduced negative emotions, and findings suggest they can expand the pie relative to human mediators when integrative potential is high.

    Source: 1) Dampening of Negative Emotions · p. 6, C. AI Mediators Expand the Pie · p. 6

  • Takeaways & Limitations

    AI mediation shows potential to regulate emotion and support joint gains in heated disputes.

    Source: §VII · p. 7

    The buyer was played by GPT4o and mediators were novice crowd-sourced workers, so generalization to human-human disputes and comparison with trained professionals remain open.

    Source: §VII · p. 7

Abstract

from arXiv · show

In conflict and disputes, especially, emotion acts as a salient force in influencing outcomes. Prior work shows negative affect can obstruct collaborative behaviors, which typically lead to ``win-win'' outcomes. Thus, some suggest mediators may help regulate emotion and achieve joint gains. With the proliferation of AI, we posit LLMs may perform well at this task, with the added benefit of better accessibility compared with a human mediator. To examine the effectiveness of AI versus novice human mediators, we conduct a between-subjects experiment, where participants engage in a dispute mediated by a human, AI, or no mediator. We first analyze how well the mediators regulate emotions within a dispute -- finding AI mediators perform significantly better than humans at reducing negative emotion. We next examine whether AI mediators facilitate disputants better realizing joint gains in disputes with high integrative potential (IP) -- we find a marginally significant interaction between IP and condition (AI versus human), indicating LLMs may outperform humans at aiding disputants realize joint gains. Lastly, we perform an analysis of the messages the mediators sent, finding the AI sent significantly more messages suggesting trade-offs compared to the humans.

I. INTRODUCTION

Disputes combine competing goals with shared interests, but their emotional intensity can make compromise and resolution difficult. The paper examines whether mediators can manage emotion and, in turn, support collaboration and joint gains.

I. INTRODUCTION

  • Disputes mix competing goals and shared interests, so each side can maximize self-interest only through compromise.Perceived wrongs and anger can entrench positions and make resolution harder,.Source: §I · p. 1
  • Negative affect can obstruct joint gains, while collaborative behaviors such as information exchange can help disputants create them –.Positive affect has been linked to fewer contentious tactics and better joint outcomes, as well as more creative problem-solving.Source: §I · p. 1
  • Mediators are expected to support voluntary resolution by managing emotion, preventing escalation, clarifying miscommunication, and moving disputants away from unreasonable demands,,,,.The paper prompts both human and AI mediators to consider these goals.Source: §I · p. 1
  • The paper tests whether LLM mediators can influence emotional dynamics and whether they better regulate emotion and support joint gains.The authors focus on increasing positive affect and decreasing negative emotions.Source: §I · p. 1

A. Affect in Dispute Resolution

  • Dispute emotions can communicate internal states and contribute to escalatory spirals that entrench positions,.Prior online-dispute research links negative emotional attacks with lower resolution likelihood, while emotional effects may depend on disputants’ vulnerability.Source: A. Affect in Dispute Resolution · pp. 1–2
  • Professional mediators often promote free expression of emotion, which may reveal threatened values relevant to change,.The literature also examines how mediators manage emotion in practice –.Source: A. Affect in Dispute Resolution · p. 2

B. Integrative Potential in Disputes

Prior work links positive affect with collaborative behavior and joint gains, while research on LLM mediation has rarely tested emotion regulation and joint gains in heated disputes against novice humans and mediation baselines. The paper frames these as questions about emotion regulation and trade-offs across levels of joint-gain potential.

B. Integrative Potential in Disputes

  • Prior work associates positive affect with collaborative problem-solving and joint gains,, while negative feelings between negotiators correlate with fewer joint gains.This motivates a potential pathway from mediator emotion management through collaboration to mutual benefits.Source: B. Integrative Potential in Disputes · p. 2

C. AI Mediation

  • Earlier LLM-mediation studies examined intervention choices and messages, but little work tested emotion regulation and joint gains in heated disputes against novice humans, and none included mediation baselines.Prior studies used precollected dialogues, human-LLM interaction, or curated datasets –.Source: C. AI Mediation · p. 2

III. RESEARCH QUESTIONS

  • The paper asks whether LLM mediators regulate emotion more effectively than novice humans and whether they facilitate trade-offs across varying potential for joint gains.These are the study’s two research questions.Source: §III · p. 2

IV. METHODOLOGY

The study adapts the KODIS buyer-seller dispute into a task with four negotiable issues. Participants state their issue preferences beforehand, allowing disputes to vary in integrative potential.

A. Dispute Task

  • The study uses the KODIS simulated buyer-seller dispute, designed to include dispute markers such as impasses and emotional dynamics [25],.The scenario concerns a buyer seeking a refund for a jersey the seller says was correctly listed, with negative reviews worsening the relationship.Source: A. Dispute Task · p. 2, 1) Disputant Perspective · p. 3

1) Disputant Perspective:

  • Disputants negotiate over a refund, the buyer’s review, the seller’s review, and apologies.The issues have differing numbers of possible outcomes: three for refunds, two for each review, and four for apologies.Source: 1) Disputant Perspective · p. 3
  • Participants allocate 100 points among issues before the dispute, creating cases with varying integrative potential and scoring outcomes against their stated preferences.A bonus contingent on achieving self-defined priorities incentivizes fidelity to those preferences.Source: 1) Disputant Perspective · p. 3

2) Mediator Perspective:

The study recruited U.S.-based crowd-sourced participants and used human or LLM mediators who could intervene during disputes. Participants stated issue preferences before entering the platform, while the buyer role was consistently played by an AI confederate.

2) Mediator Perspective:

  • Human and LLM mediators received the same instructions, could access disputants’ preferences, and chose after each utterance whether to intervene and what to say.They were tasked with helping the parties reach agreement and facilitating agreements beneficial to both sides.Source: 2) Mediator Perspective · p. 3

1) Participants:

  • The study recruited 164 seller disputants and 50 mediators through Prolific, with participants residing in the United States.The reported sample averaged 37.9 years of age (SD = 12.6).Source: 1) Participants · p. 3

2) Pre-Task:

  • Before the dispute, participants received role instructions and sellers entered preferences over issues; GPT4o always played the buyer and was prompted to speak conversationally.The AI buyer was used to induce situations requiring mediation because many buyers in KODIS do not show anger,,.Source: 2) Pre-Task · pp. 3–4

3) Intervention:

The experiment compares disputes with no mediator, an AI mediator, or a human mediator. Mediators could intervene after disputant messages, using access to both parties’ preferences to support value creation.

3) Intervention:

  • The experiment assigned human sellers to disputes with no mediator, an AI mediator, or a human mediator.The human mediators were crowd-sourced workers; the study focuses on outcomes for human disputants.Source: 3) Intervention · p. 4
  • The AI mediator used a smaller model to decide whether to intervene and a larger model to craft messages, with scores above seven triggering a message.The threshold came from an earlier offline KODIS study, where seven yielded more interventions in impasses; the two-model setup was intended to reduce latency.Source: 3) Intervention · p. 4, 4) Dispute Task · p. 4
  • Human and AI mediators received the same instructions, while the AI prompt emphasized neutrality and guiding parties toward voluntary agreement.The AI monitoring model rated whether to intervene on a one-to-ten scale.Source: 3) Intervention · p. 4

4) Dispute Task:

  • After each disputant utterance, a mediator could send a message or pass the turn, and could use both parties’ preferences to facilitate value creation.Disputants communicated in natural language with strict turn-taking; after eight exchanged messages, either could propose a final deal or walk away.Source: 4) Dispute Task · p. 4

5) Post-Task:

After the dispute, participants completed system-impression questions and were debriefed about whether other actors were AI. Objective outcomes included impasse rates and preference-based points for each disputant and jointly.

5) Post-Task:

  • After returning from the dispute platform, participants answered questions about their system impressions and received a debrief identifying AI counterparts when applicable.The debrief distinguished cases where both other actors were AI from cases where only the buyer was AI.Source: 5) Post-Task · p. 4

1) Objective Measures:

  • The study measured impasse rate and each disputant’s points based on the outcome and their stated preferences, coding walkaways as zero points.Joint points were defined as the sum of the buyer’s and seller’s points.Source: 1) Objective Measures · p. 4

2) Integrative Potential:

Integrative potential measures the extent to which disputants can create joint gains through trade-offs. The study also coded dialogue for emotion and assessed how well those labels aligned with human ratings.

2) Integrative Potential:

  • Integrative potential (IP) is reverse-scored cosine similarity between buyer and seller preferences, ranging from zero to one, with IP = 1 indicating the highest potential for value creation.The buyer’s preferences stayed constant while sellers’ preferences varied across rounds.Source: 2) Integrative Potential · p. 4
  • Table I provides examples of AI and human mediators suggesting trade-offs in disputes.The caption identifies the table’s scope but does not state which mediator’s examples are more effective.Source: Table I · p. 5

3) Affective Measures:

  • GPT-4.1 annotated each dialogue line for six dispute-relevant emotions, assigning intensity values that sum to one and allowing a neutral label.The categories were anger, joy, fear, sadness, surprise, and compassion.Source: 3) Affective Measures · p. 5
  • Human-rating agreement varied by emotion: anger correlated at r = .58 and compassion at r = .44, while fear showed no positive correlation.The study used 255 Prolific workers to rate 50 human-seller utterances; the authors describe anger and compassion as showing moderate agreement overall.Source: 3) Affective Measures · p. 5

4) Dialogue Coding:

The AI condition had the lowest observed impasse rate, but condition differences were not statistically significant. The study also coded mediator messages for suggestions of trade-offs.

4) Dialogue Coding:

  • The dialogue coding adapted NegotiAct and focused on whether mediators suggested trade-offs.The measure was the proportion of opportunities to intervene—one after each disputant utterance—in which the mediator sent a trade-off suggestion.Source: 4) Dialogue Coding · p. 5

A. Effect on Impasse

  • 47.3% of AI-mediated disputes ended in impasse, versus 54.3% with no mediation and 58.0% with human mediation; the difference was not significant, χ2(2, N = 164) = 1.262, p = .53.The AI condition averaged 7.1 interventions per dialogue, compared with 6.2 for human mediation.Source: Table II · p. 5, A. Effect on Impasse · p. 5

B. AI Mediators Better Control Emotion

AI mediation significantly reduced negative emotion and increased positive emotion; the human and no-mediation conditions did not show significant changes after the corresponding comparisons.

B. AI Mediators Better Control Emotion

  • For the no-mediation comparison, GPT-5-mini indicated where it would have intervened; surprise was excluded from both positive and negative emotion groupings.The no-mediation condition had no mediator messages, so the model's indicated intervention points served as a counterfactual baseline.Source: B. AI Mediators Better Control Emotion · pp. 5–6

1) Dampening of Negative Emotions:

  • AI mediators significantly reduced negative emotion after messages (M = 0.18, SD = 0.20) versus before (M = 0.24, SD = 0.20), unlike human and no-mediation conditions.The time-by-condition interaction was significant (F(2, 154) = 4.57, p = .01); Figure 3 depicts this interaction.Source: 1) Dampening of Negative Emotions · p. 6, Figure 3 · p. 6
  • Anger and fear showed stronger interaction evidence (p = .07 each) than sadness (p = .56), though seller intensity fell after mediator intervention for each emotion.The individual-emotion analysis reports declines after intervention across all three negative emotions.Source: 2) Proliferation of Positive Emotions · p. 6

2) Proliferation of Positive Emotions:

  • F(1, 154) = 9.05, p < .01 showed positive emotion rose overall; AI-mediated disputes increased from M = 0.09 to M = 0.14 (p = .01), unlike the other conditions.Positive emotion combined compassion and joy; the AI change was significant, while the human and no-mediation changes were not.Source: 2) Proliferation of Positive Emotions · p. 6

C. AI Mediators Expand the Pie

In disputes with high integrative potential, AI mediation was associated with greater joint value than human mediation, and AI mediators sent more trade-off suggestions.

1) IP Interacts with Joint Points:

  • β = 0.37, p = .056 for the joint-points condition-by-integrative-potential interaction; seller points showed a significant interaction (β = 0.39, p = .04), while buyer points did not (β = 0.27, p = .16).The regression compared AI- and human-mediated cases and used seller, buyer, and joint points as outcomes.Source: 1) IP Interacts with Joint Points · p. 6
  • Table III reports mean proportions and standard errors for opportunities each mediator used messages fitting NegotiAct codes [35].The table caption describes message-code proportions, not the outcome values in the supplied passage.Source: Table III · p. 7

C. AI Mediators Expand the Pie

2) AI Mediators Suggest More Trade-offs:

AI mediators suggested trade-offs more often than novice humans, and their messages served multiple functions. The authors discuss possible links to joint gains while noting that the mechanism is unconfirmed and that the study has ethical and generalizability boundaries.

2) AI Mediators Suggest More Trade-offs:

  • t = 8.00, p < .001: AI mediators suggested trade-offs at more intervention opportunities (M = 0.38, SD = 0.27) than humans (M = 0.07, SD = 0.10).The authors say trade-off language may help explain increased joint gains, but confirming that mechanism requires future studies.Source: 2) AI Mediators Suggest More Trade-offs · p. 7
  • AI messages averaged 3.12 codes each, compared with 1.07 for humans, and more often suggested trade-offs, actions, offer modifications, procedures, and time considerations.AI mediators also more often asked for substantiation, positional information, and clarification, and engaged in more active listening.Source: 2) AI Mediators Suggest More Trade-offs · p. 7

VI. DISCUSSION

  • The AI–human joint-gains interaction was marginally significant when integrative potential was high, but the study did not establish that emotion regulation caused those gains.The authors also report that AI mediators reduced negative emotion while human mediators did not, and that AI increased positive emotion while human and counterfactual conditions did not.Source: §VI · p. 7

VII. CONCLUSION & FUTURE WORK

  • The study used GPT4o as the buyer and novice crowd-sourced mediators, leaving generalization to human-human disputes and comparisons with trained mediators unresolved.It also examines text interactions only and did not find a significant effect on impasse rates.Source: §VII · p. 7

VIII. ETHICAL IMPACT STATEMENT

  • AI mediation must preserve voluntary participation and disputants’ self-determination; bias in language models may compromise their neutrality.The authors caution against imposing solutions rather than facilitating a resolution.Source: §VIII · p. 8
  • LLM-generated annotations may reflect an overly Western perspective, potentially producing incorrect inferences or worse outcomes in intercultural disputes.The study used American-based participants, and the authors urge extra care when deploying mediators or interpreting annotations in multicultural scenarios.Source: §VIII · p. 8
Loading 2609.17933v1…