Characterized by elevated hostility, disputes often result in disputants being unable to resolve their diferences. This emphasizes the role of a mediator; however, mediators are typically highly specialized, dificult to find, and expensive. While alternatives exist, such as acquaintances or online moderators, these prove less efective. This raises the question of whether AI can efectively facilitate contentious disputes. We assess the potential of LLMs as mediators. First (Study 1), we focus on a large open-source corpus of customer service disputes that are objectively classified in terms of two key reasons for mediation: (1) whether or not the dispute ended in success or failure, and (2) to what extent the participants reported frustration with each other. We examine whether LLMs could be prompted to predict when a mediator should intervene in advance, and show that the decision to intervene is correctly sensitive to these two factors. Finally, we conducted a user study that compared AI to human suggestions on when and how to intervene. We find that the LLMs were rated significantly better in predicting when to intervene, rated as providing a better rationale for intervening, and rated as providing a more efective mediation message to the disputants. Remarkably, observers preferred the AI 2-to-1 over the human mediation decisions. Second (Study 2), we collect a small corpus of mediated dispute dialogues and analyze the efectiveness of novice human mediators. Quantifying the similarity of novice mediators' behaviors to those of an LLM, we find the closer one mediates to an LLM, the better the outcome - both subjectively and objectively. These results indicate LLMs not only outperform novice mediators but may serve to identify efective mediators. While these results do not definitely demonstrate LLMs can mediate human disputes, we do show they can sense escalation indicators and generate sensible messages - this work serves as a first-step toward AI-mediators in human conflict.
This paper describes two studies assessing the ability of Large Language Models (LLMs) to mediate
contentious and emotional disputes. Disputes arise when one party in a relationship makes a claim that
another party rejects, thus threatening the future of that relationship [
In evaluating LLMs as mediators, we start by examining a large corpus of dispute dialogues in which
online participants acted as buyers or sellers in a simulated purchase conflict [
First, with open-source dyadic dispute corpus (KODIS [
The dispute resolution literature outlines empirical and theoretical guides for mediation best practices.
For example, several studies show the efect of mediator bias on disputant outcomes [
The AI research community has surveyed the growing capabilities of AI in mediation and moderation
tasks. Prior work has examined the potential of AI to detect and potentially intervene in disputes,
primarily over posts on social media. Much of this work has focused on recognizing overtly toxic
comments after they have been posted, such as detecting personal attacks [
Study 12 motivates the potential of LLMs to act as mediators in a dispute context, leveraging an
opensource corpus of dispute resolutions KODIS [
First, we investigate if an LLM (gpt-4-0613) can efectively determine when to intervene in disputes pulled from the KODIS corpus. In KODIS, a buyer and seller dispute over an online order of a basketball jersey. The buyer reads a role-play prompt that states they ordered a Kobe Bryant jersey from an online seller, however, received a generic one instead; on reaching out, the seller denies their refund request; and each side then posts negative reviews about the other. The seller reads a similar prompt, but believes they never described the product as a Kobe Bryant jersey. Thus, the scenario primes the buyer and seller to argue over facts — a primary characteristic of disputes — as they attempt to resolve four core issues: whether the buyer receives a refund, whether the buyer removes their negative review of the seller, whether the seller removes their negative review of the buyer, and whether each side apologizes. KODIS contains many participant responses, though we focus on the objective outcome (whether the dispute ended in resolution or impasse) and self-reported frustration (from the Tactics scale [31], an average of the frustration-related questions from each side). Using pre-collected dialogues from this corpus, we experiment with LLM mediators. 1For these studies, we received IRB approval and consent from participants, which they could revoke at any time. 2Study 1 is an extension of previous work by the authors [30].
Specifically, we analyze whether an LLM can pick up on salient features in a dispute (e.g., frustration, and outcome) in determining to intervene. We iterate through each dialogue exchange in each dispute, giving the LLM the conversation history thus far and asking it to determine whether to intervene at the current point (see Figure 7 for the prompt used). We construct the prompt ensuring the model understands its role as a mediator; identifies the severity of the situation on a scale from one to ten (Intervention Score); selects the reason for intervention from four categories (Escalation of conflict , Impasse, Miscommunication, or Unreasonable demands); and generates an appropriate response to guide the parties. We expect the LLM to ascribe higher Intervention Scores if participants report higher frustration and if the dispute ends in an impasse. 3.1.2. Results We perform a moderated regression ( (3, 1419) = 323.6 , < .001 , 2 = .41) to determine whether the diferences in the Mean Intervention Score (the average of all Intervention Scores generated for each exchange in a given dispute) over all ( = 1, 782 ) human-human dialogues significantly difer between two factors — 1) whether the dialogue impasses or resolves (Impasse), and 2) how much Frustration participants self-report (Z-scored). The test yielded main efects on Mean Intervention Score for each independent variable. We find a significant main efect of Impasse on Mean Intervention Score ( = 1.75 , = 0.12 , = 14.14 , < .001 ), where Tukey’s posthoc test revealed the LLM scored dialogues resulting in impasse significantly ( < 0.01 ) higher ( = 5.51 , = 1.69 ) than those resulting in resolution ( = 3.16 , = 1.61 ); we also find a significant main efect of Frustration on the Mean Intervention Score ( = 0.76 , = 0.04 , = 17.83 , < .001 ); lastly, we find no significant interaction between the independent variables ( = 0.08 , = 0.11 , = 0.75 , = .45 ). Thus, we find that an LLM can perceive and act when disputants become frustrated with one another and when a dispute moves to an impasse. Figure 2a visualizes these results. Figure 2b illustrates the average LLM intervention score over time broken out by factor, with Frustration binarized into high or low using median-split. One can see the LLM’s intervention score rise as the dialogue continues if it ultimately ends in an impasse, and there exist higher intervention scores in high-frustration dialogues.
Impasse Resolution
High Frustration Low Frustration 10 9 e8 r o c S7 n o iten6 v r e5 t n I n4 a e 3 M 2
Given the previous section establishes LLMs can competently determine when to intervene, the question remains of whether an LLM can formulate an efective message at an appropriate point — i.e., can an LLM decide how to intervene? We compare LLM mediations against those of novice human mediators and ask crowd-sourced annotators to rate each on several subjective measures — e.g., appropriateness of the intervention point, the efectiveness of the message, and whether an accompanying justification supports their action — and to ultimately pick which they felt more efective at guiding toward a resolution. We expect to find that LLMs significantly outperform novice human mediators.
We use crowd-sourced mediations gathered from Prolific as a baseline against the LLM — i.e., the novice mediators. As participants enter the online survey, we tell them they will role-play a mediator, working for an online retailer, overseeing a buyer and seller as they dispute over a purchase gone wrong. Further, we notify them of their goal of intervening only if they believe the dialog will otherwise stall at an impasse; we incentivize performance by ofering an additional bonus of $0.50 for each dialog (five total) a participant intervenes where it ends in an impasse or if they remain inactive for one that ends in resolution. When ready, they enter a page with the chat interface depicted in Figure 3 where they can navigate through a KODIS dispute dialog utterance-by-utterance (“Continue”), intervening (“Intervene,” with an accompanying message) when and how they decide. Through this, we collect the intervention point (the point in the dialog where the participant intervenes) and their message for 198 dialogues.
A diferent set of Prolific crowd-workers compared LLM mediations against novice human ones on a subset ( = 20 ) of the dialogues where both the LLM and the novice human mediator elected to intervene. Specifically, this was a within-subjects design, where we ask crowd workers ( = 106 ), given a single random mediated dialog up to an intervention point as well as the intervention/justification, to evaluate and compare the attempts of LLM and a human meditator (blind to which) on three subjective measures (1-10 Likert scale) — appropriateness of the intervention point, the efectiveness of the message, and whether an accompanying justification supports their action (see Table 1 for phrasing) — and to ultimately pick which they felt more efective at resolving the dispute. 3.2.2. Results We use a two-tailed t-test to test our hypothesis that human annotators prefer LLM mediations to human ones and find significance across the three questions supporting as much. We see participants view the LLM’s mediations as making resolution more likely, having more appropriate timing, and giving better justification. Table 1 summarizes the statistics discussed. Lastly, a Chi-squared test on a forced choice between the LLM or human mediations yielded a significant result ( 2(1, = 106) = 6.29 , = .01 ), where 71 participants selected the LLM-generated mediation compared to 35 for the human-crafted one — i.e., the human evaluators preferred the LLM’s mediations at a rate of two-to-one.
I believe this mediation increases the probability of a resolution.
The supervisor intervened at an appropriate point.
The supervisor provided appropriate justification for intervention.
Here, we demonstrated that an LLM could appropriately act as a mediator and intervene with some accuracy — doing so in disputes where participants became frustrated or headed toward an impasse. Secondly, we demonstrated that LLMs could craft compelling mediation messages to rival human mediators. Crowd-sourced annotators evaluated the LLM as more likely to induce resolution, intervening at a more appropriate point, and providing better justification than human mediators; participants also overwhelmingly marked LLM mediations as better than human in a forced-choice question — choosing the LLM-generated mediation by a margin of two-to-one. In explaining the superior performance of LLMs in this task, one might consider that LLMs possess an innate lack of fatigue, broad training data, and a high level of consistency. This, we posit, lays the groundwork for using LLMs as mediators in complex dispute settings. However, one may recognize the lack of interactivity as a limitation of this work, as we cannot gauge whether these mediations positively impact subjective evaluations or impasse rate. We aim to address this in the next section via analysis of three-way mediated disputes.
This second study, building on the first, addresses the primary limitation of those analyses — namely, the lack of interactivity between the disputants and mediator. I.e., in the previous study, mediators annotated where they would intervene and what they would say on previously collected dialogues, which does not allow one to examine their efectiveness. However, we do wish to evaluate that efectiveness, so we conduct this second study. Specifically, we collect a smaller version of the previously used corpus 3, with a third participant in the chat-room acting as a mediator; we quantify the similarity of the human mediators with LLM mediators (gpt-4o-2024-08-06); and we test whether there exist significant outcome efects when a human mediator behaves more similarly to an LLM one. We find such efects in terms of subjective (e.g., SVI) and objective (e.g., impasse vs. resolution) outcomes.
We use Lioness Labs [32], a tool used by prior work [
After completing the mediated dispute, participants answer some post-task questions. The buyer and seller role players fill out Curhan
et al.’s Subjective Value Inventory (SVI) questionnaire [38], which contains four sub-scales measuring subjective feelings about the dispute — e.g., feelings about the instrumental outcome, process fairness, self, and relationship. We also derive questions from those sub-scales, explicitly invoking the mediator, to gauge the impression of the disputants and mediator about the mediator’s performance.
• Outcome: (You / The Mediator) helped achieve a satisfactory outcome. • Relationship: (You / The Mediator) helped repair both parties relationship. • Process: (You / The Mediator) helped facilitate a more fair outcome. • Self: (You / The Mediator) helped (parties / me) to keep face, act to (their / my) principles, and negotiate competently.
• Avoid Impasse: (You / The Mediator) helped avoid a walkaway.
We compensated parties $3.50 for a task that took approximately 20 minutes. Each player could earn an additional bonus of up to $3 depending on how well they achieved their objectives — e.g., the buyer and seller settling the refund, review, and apology issues, and the mediator encouraging both sides to come to a resolution. We intended this bonus to motivate participants, in this online setting, to immerse themselves in their role and to achieve their objectives.
In addition to the metrics recorded during data collection, we create various metrics to gauge the efectiveness of LLM mediators. Here, we outline how we quantify the similarity of intervention patterns between two mediators (human and LLM). Assume we have two vectors representing the intervention patterns of the human and LLM — for the human and for the LLM. Consider , though we define the same way, we set = 1 if the human intervenes after the Whereas we construct from the collected dialogue, we create afterwards on that same dialogue, removing the human mediator messages. Specifically, given some dialogue, we remove the mediation messages; have an LLM iterate through each utterance and determine whether to intervene, in the same manner as Study 1 (though the LLM does not output an Intervention Score, rather 1 to intervene or 0 to not); and construct such that = 1 if the LLM intervened after utterance , and zero otherwise. Thus, we have and , which represent the intervention patterns of the human and LLM, respectively, for a dialogue. Given these, we wish to gauge the similarity of each human mediator to the LLM’s behavior.
We derive a measure from the Earth Mover’s Distance (EMD) to quantify the similarity of the human and LLM mediator behavior — i.e., comparing and . As outlined below, we iterate through each element of and in parallel, keeping track of the cumulative distance up to each element (e.g., CD for element ); the Earth Mover’s Distance sums over the absolute value of each CD . ℎ message and zero otherwise.
=0 CD = ∑ −
EMD = ∑ |CD | =0 (1) Our decision of EMD stems from requiring a metric that considers two vectors as more similar if they contain interventions closer in proximity, rather than judging them as equivalent. For example, consider vectors , , and where = 1, +1 = 1, +2 = 1, and all else is 0; we expect EMD(, ) < EMD(, ) since intervened closer to where did than . With EMD, this is true; however, with other metrics, such as Euclidean distance, it does not hold. Of note, EMD measures distance, whereas we would like to measure similarity — thus, we scale this metric by negative one. Further, we Z-score this for subsequent analysis. Going forward, we refer to this measure as Intervention Alignment. Of note, this metric does not consider what the mediators say to disputants; we next outline a method to quantify that.
Building on the aforementioned intervention similarity metric, we construct a metric to measure the similarity of the messages that the human and LLM mediators send. Contrasting with the previous approach, we do not allow the LLM to decide where to intervene — rather, we fix the intervention points to where the human novice intervened. Then, given the dialogue history, we prompt the LLM to generate a message and select one of the justifications from Section 4.1.1 (or NA if none fit) — at the same time, we categorize the human’s message into one of those justifications. We compute our Justification Alignment dependent variable, then, as the proportion of interventions with matching categories for a given dialogue — discarding those where the novice chose not to intervene. Therefore, this variable gives a sense of whether the messages the human and LLM send have similar intents.
We begin by examining the impact of the novice human mediations on subjective and objective outcomes. For each dialogue, we ascribe a binary factor denoting whether the novice mediator intervened at all (the novice mediators intervene at least once about 83% of the time) — in the subsequent statistical tests, we consider this alongside another binary factor representing the outcome (impasse or resolution). We run four two-by-two ANOVAs considering the efect of intervention and outcome on subjective outcome (SVI). For the four SVI sub-scales, we do not find any significant main efects of this intervention factor; rather, we find main efects of outcome ( < .001 for each) such that disputants report a higher SVI if they come to a resolution. Of note, in three of the SVIs, we see disputants report worse subjective feelings about the outcome if the mediator intervened. Further, considering the impact of intervention on the objective outcome, we see disputes resolved 81% of the time if the mediator intervened, compared to 77% of the time otherwise — this small diference does not reach significance ( = .91 ) by Chi-squared.
Outcome
Intervention Process
No Intervention
Self
Relationship 1 Resolution
Impasse Resolution Impasse Resolution Impasse Resolution Impasse Figure 4: Depicts mean SVI sub-scale scores by outcome (impasse vs. resolution) and whether the human mediator intervened.
Given these seemingly unimpressive results, we proceed to analyze the impression from each party of the mediator’s performance — specifically, Figure 5 depicts the responses from each party to the questions overviewed in Section 4.1.1. We see the mediators actually thought they performed well, relative to the buyer and seller evaluations. Next, we analyze whether we can quantify the extent to which these novices behave like LLMs, and if there exist any efects of that on the outcomes.
Mediator
Buyer
Seller Outcome Relationship
Process
Self
Avoid Impasse
First, we consider the similarity of the intervention pattern between the human and LLM (Intervention Alignment ). We conduct a regression analysis to examine the extent to which Intervention Alignment predicts subjective (SVI outcomes) and objective (Impasse vs. Resolution) outcomes. By linear regression, it significantly predicted each of the four SVI sub-scales, as seen in Table 2. As the human’s intervention pattern becomes more similar to the LLM’s, the subjective outcome improves across the board — Figure 6 (top row) illustrates these relationships. Next, we consider the objective outcome — i.e., a binary dependent variable of whether the dispute ended in a resolution (coded as zero) or impasse (coded as one). We run a logistic regression to test whether there exists a significant efect of the similarity metric on the outcome. The regression ( = −.41 , = .23 , = −1.75 , = .08 ) shows a trend, where the more similar one acts to the LLM, the less likely an impasse. Further, we see an impasse rate of 29% in dialogues with the bottom half similarity, which reduces to 11% for the upper half.
SVI Subscale Relationship
Self
Process Outcome
Intervention Alignment
B SE p 2 0.37 0.09 <.001 0.21 0.10 0.035 0.36 0.10 <.001 0.31 0.10 0.002
Mirroring the analysis for Intervention Alignment, we conducted a regression analysis to test the efect of the Justification Alignment metric on the various outcome measures. Starting with the subjective outcome, we run four linear regressions, the results of which one can find in Table 2, which yielded significant results for all subjective outcomes, aside from self. We again run a logistic regression to analyze the efect of semantic similarity on impasse. We find a significant result ( = −.84 , = .34 , = −2.47 , = .013 ) such that if the novice mediator tended to send messages in the same category as the LLM, the chance of an impasse reduced significantly.
This study, building on the first, demonstrates that LLMs possess the potential to mediate human disputes in a triadic setting. The first study demonstrated LLMs can tune into salient features of pre-collected dispute dialogues, intervening — as expected — in dialogues with high frustration and ending in impasse; in this second study, we analyze three-party human mediations, finding that novice mediators acting more similarly to LLMs achieve better outcomes. Importantly, whereas in the first study, we could not measure the efect of an intervention on disputes (we leveraged a pre-collected corpus); in the second, we could. In Study 2, initially, we see uninspiring results related to impasse rate and disputants’ reported subjective outcomes (SVI), where it seems as though, if anything, mediators imposed an adverse efect. Alternatively, one may consider that mediators intervene more so in disputes heading toward impasse or with high frustration, and parties could take a mediator’s interjection as a signal that the outcome may be worse — this could partially explain the diferences in subjective evaluations in cases of impasse between dialogues where the mediator intervened versus not (see Figure 4). Despite this, we see mediators rate themselves relatively higher compared to evaluations from disputants across several dimensions — potentially an example of the self-serving bias [39]. Considering these results, we explore demonstrating that LLMs have the potential to act as mediators.
We create two variables to capture how similar a novice mediator acts to an LLM: 1) Intervention Alignment, which captures the similarity of the intervention patterns of mediators, and 2) Justification Alignment, which captures whether two mediators intervene for similar reasons. In terms of objective outcomes, we find a trend and a significant efect, respectively, where higher similarity to the LLM reduces the probability of an impasse. Further, for nearly every SVI (aside from Self for Justification Alignment), we see significant efects. This implies that when a novice mediator acts closely to an LLM one, outcomes improve. There still exists a limitation in interactivity with Study 2, where disputants never interact with an LLM mediator directly — rather, we collected three-party dialogues, and quantified how much an LLM “agreed” with a novice mediator.
This work demonstrates LLMs respond to escalation markers, and generate sensible mediation messages; further, we show human mediators acting more similarly to LLMs induce better outcomes. Together, these findings indicate LLMs could operate as a mediator in a zero-shot setting. We have two primary results: 1) We show LLMs can efectively determine when and how to intervene on a pre-collected corpus, and 2) we demonstrate the potential of LLM mediators in a triadic dispute setting via elevated subjective evaluations from disputants and lower impasse rates. However, this work has several limitations that we aim to address in future work. First, we test only one LLM (GPT4o) with one prompt configuration (zero-shot). One might imagine performance improvements when endowing an LLM with expert strategies for mediation [40], or in leveraging popular promoting techniques (e.g., chain-of-thought). Secondly, in the work’s current paradigm, we have the LLM intervene only when it detects a potential issue in the dialogue – however, one could imagine a mediator might attain greater efectiveness through proactively intervening, preventing issues before they arise. We will consider this notion going forward. Lastly, while an improvement over Study 1, Study 2 still did not place an LLM mediator directly with human participants — rather, we evaluate the efectiveness of the LLM mediators through the performance of the novices. Another study, placing the LLM directly between two human disputants, must occur before making a definitive statement on how well LLMs can mediate — however, this work illustrates their promise.
These types of social influence technologies carry the potential for harm, especially considering the intricate emotional dynamics at play. E.g., as Schluger et al. [41] note, technologies that proactively work to prevent escalatory spirals in conversation — as does the LLM mediator in our scenario — may infringe on one’s freedom of speech if they shut down the conversation or intervene based on a prediction of future bad behavior. Further, prior work demonstrates LLMs struggle to generalize across cultures with emotion [42] — given the salience of emotion in disputes, this warrants consideration.
Further, over-reliance on AI can cause harm — especially in emotionally charges settings like dispute resolution. E.g., Passi and Vorvoreanu [43] note several manifestations of over-reliance: users may have bias to favor automatically generated results; a user may over-rely on AI if it does well initially, even if it fails later; AI-explanations may also cause over-reliance. Generally, a mediator should not dominate a dispute or exert pressure in such a way that disputants defer to them entirely — rather, a mediator should guide disputants toward a solution [44]. As such, over-reliance on the AI-mediator may hamper the mediation process.
This work is supported by the U.S. Government including the Air Force Ofice of Scientific Research (grant FA9550-23-1-0320), and the National Science Foundation (grant 2150187). The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the oficial policies, either expressed or implied, of the Army Research Ofice or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein.
During the preparation of this work, the authors used GPT5.2 and Grammarly to help with writing clarity, and to assist with Python coding. After using these tools, the authors reviewed, verified, and edited the content as needed and take full responsibility for the publication’s content. delegation: A case study of content moderation, in: Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, 2022, pp. 1–18. [27] J. Govers, E. Velloso, V. Kostakos, J. Goncalves, Ai-driven mediation strategies for audience depolarisation in online debates, in: Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, CHI ’24, Association for Computing Machinery, New York, NY, USA, 2024.
URL: https://doi.org/10.1145/3613904.3642322. doi:10.1145/3613904.3642322. [28] M. H. Tessler, M. A. Bakker, D. Jarrett, H. Sheahan, M. J. Chadwick, R. Koster, G. Evans, L. CampbellGillingham, T. Collins, D. C. Parkes, et al., Ai can help humans find common ground in democratic deliberation, Science 386 (2024) eadq2852. [29] J. Tan, H. Westermann, N. R. Pottanigari, J. Šavelka, S. Meeùs, M. Godet, K. Benyekhlef, Robots in the middle: Evaluating llms in dispute resolution, arXiv preprint arXiv:2410.07053 (2024). [30] J. Hale, H. Kim, A. Choi, J. Gratch, Ai-mediated dispute resolution, in: Proceedings of the AAAI
Symposium Series, volume 5, 2025, pp. 67–70. [31] S. Aslani, J. Ramirez-Marin, J. Brett, J. Yao, Z. Semnani-Azad, Z.-X. Zhang, C. Tinsley, L. Weingart, W. Adair, Dignity, face, and honor cultures: A study of negotiation strategy and outcomes in three cultures, Journal of Organizational Behavior 37 (2016) 1178–1201. [32] M. Giamattei, K. S. Yahosseini, S. Gächter, L. Molleman, Lioness lab: a free web-based platform for conducting interactive experiments online, Journal of the Economic Science Association 6 (2020) 95–111. [33] K. Chawla, J. Ramirez, R. Clever, G. Lucas, J. May, J. Gratch, Casino: A corpus of campsite negotiation dialogues for automatic negotiation systems, in: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2021, pp. 3167–3185. [34] A. Vannucci, C. M. Ohannessian, K. M. Flannery, A. De Los Reyes, S. Liu, Associations between friend conflict and afective states in the daily lives of adolescents, Journal of Adolescence 65 (2018) 155–166. [35] C. K. De Dreu, L. R. Weingart, Task versus relationship conflict, team performance, and team member satisfaction: a meta-analysis., Journal of applied Psychology 88 (2003) 741. [36] J. Van Veenen, Dealing with miscommunication, distrust, and emotions in online dispute resolution (2010). [37] P. F. Kirgis, Bargaining with consequences: Leverage and coercion in negotiation, Harv. Negot. L.
Rev. 19 (2014) 69. [38] J. R. Curhan, H. A. Elfenbein, H. Xu, What do people value when they negotiate? mapping the domain of subjective value in negotiation., Journal of personality and social psychology 91 (2006) 493. [39] D. R. Forsyth, Self-serving bias (2008). [40] S. B. Goldberg, J. M. Brett, B. Blohorn-Brenneur, How mediation works: Theory, research, and practice, Emerald Publishing Limited, 2017. [41] C. Schluger, J. P. Chang, C. Danescu-Niculescu-Mizil, K. Levy, Proactive moderation of online discussions: Existing practices and the potential for algorithmic support, Proceedings of the ACM on Human-Computer Interaction 6 (2022) 1–27. [42] S. Havaldar, B. Singhal, S. Rai, L. Liu, S. C. Guntuku, L. Ungar, Multilingual language models are not multicultural: A case study in emotion, in: J. Barnes, O. De Clercq, R. Klinger (Eds.), Proceedings of the 13th Workshop on Computational Approaches to Subjectivity, Sentiment, & Social Media Analysis, Association for Computational Linguistics, Toronto, Canada, 2023, pp. 202–214. URL: https://aclanthology.org/2023.wassa-1.19/. doi:10.18653/v1/2023.wassa- 1.19. [43] S. Passi, M. Vorvoreanu, Overreliance on ai literature review, Microsoft Research 339 (2022) 340. [44] S. B. Goldberg, J. M. Brett, B. Blohorn-Brenneur, The roles of the mediator and the disputing parties at each step of the mediation process, in: How Mediation Works, Emerald Publishing Limited, 2017, pp. 17–57.
Imagine you are playing the role of a mediator in a buyer/seller purchase dispute. Your goal is to allow participants to resolve their dispute on their own if possible, but to intervene if necessary. Some reasons to intervene include: 1. Escalation of Conflict: if the conversation becomes heated with parties resorting to personal attacks or hostile language 2. Impasse: when parties reach a deadlock and are unable to move forward 3. Miscommunication: if there are signs that the parties are misunderstanding each other’s points 4. Unreasonable Demands: if one party is making unreasonable demands that the other party can’t possibly meet You will be given the conversation so far. Rate the situation on a scale from 1 to 10 with 1 meaning definitely don’t intervene and 10 meaning definitely intervene. Provide: • (a) the rating on whether to intervene, • (b) the reason to intervene, selected from the list above, • (c) a one-sentence statement you might tell the parties at this point.
You do not need to intervene every turn, and should consider how recently you’ve intervened before making a decision.
Here is the conversation so far… You are a mediator for a buyer/seller purchase dispute. Let parties resolve issues themselves when possible; intervene only if needed.
Evaluate only the last Buyer/Seller utterance above. Respond with only JSON in the format: { "intervene": 0 or 1, "reason": "Escalation of Conflict" | "Impasse" | "Miscommunication" | "Unreasonable demands" | "NA", "intervention": "one sentence suggestion (empty if intervene == 0)" }
Prompt: Study 2 — Justification Alignment You are an expert mediation evaluator for a buyer/seller purchase dispute.
1. Escalation of Conflict 2. Impasse 3. Miscommunication 4. Unreasonable demands 5. NA
{prev} Classify only the last mediator utterance into exactly one category. Respond with only JSON in the format: { "category": "Escalation of Conflict" | "Impasse" | "Miscommunication" | "Unreasonable demands" | "NA" }