AI-Driven Mediation Strategies for Audience Depolarisation in Online Debates

Online polarisation can tear the fabric of civility through reinforcing social media’s perceptions of division and discord. Social media platforms often rely on content-moderation to combat polarisation, contingent on the reactive removal or flagging of content. However, this approach often remains agnostic of the underlying debate’s ideas and stifles open discourse. In this study, we use prompt-tuned language models to mediate social media debates, applying the strategies of the Thomas-Kilmann Conflict Mode Instrument (TKI). We evaluate multiple mediation strategies in providing targeted responses to the debates, as shown to a debate audience. Our findings show that high-cooperativeness TKI strategies offered more persuasive arguments, while an accommodating argument strategy was the most successful at depolarising the audience’s opinion. Furthermore, high-cooperativeness strategies also increased the perception that the debaters will reach a consensus. Our work paves the way for scalable and personalised tools that mediate social media debates to encourage depolarisation.


INTRODUCTION
Social media platforms enable users to discuss ideas, project ideology, and critique political decision-making.From the days of ancient Greece, healthy democracy has been grounded on the marketplace of ideas and in public gathering places for citizens to listen to debates on social and political issues.However, online discussions often become subject to perpetual divisions, mob mentality, and misinformation.
When online debates go awry, a human or virtual moderator may reactively intervene to remove toxic or infammatory posts as dictated by a set of community rules as used on debating platforms (such as self-moderation seen on Kialo and on subreddit communities [69]), or platform-wide community guidelines [23].However, this form of intervention does not address the conficting opinions and claims that led to the debate in the frst place.It can also shift discussions to more extreme anonymous forums and echo chambers, as seen through the radicalisation of deplatformed (banned) users from mainstream social media such as Facebook, to extreme political fringe groups on Telegram [28,65] and Gab [16], and political forums such as Stormfront [28,29].
Likewise, emotion-driven content-moderation regulates posts' sentiment and emotion [28,39,69]-overlooking that polarisation is a result of idea-driven disagreements that lead to emotive responses.Thus, fostering constructive online debates with new ideas and questions to the polarised users ofers a vital proactive means to address polarisation before it can potentially spill into toxic or emotive responses.Likewise, having an objective voice collaborate in a discussion is a key measure to help promote consensus and constructive discourse-aiding both the debaters and audience without stifing speech through content-moderation.
Importantly, the victims of polarisation are not exclusively the debaters themselves.In fact, the debaters are often a small minority compared to the thousands or even millions that passively consume social media, sometimes years later [11,25,28].These audiences and wider communities are exposed to an increasingly divided debate space on platforms such as Reddit, Twitter, and Facebook [28,29].Moreover, exposure to disagreements without solutions or constructive debates can reinforce echo chambers and radicalise users who perceive that societal challenges cannot be resolved (i.e., existential dread) [40,73].A lack of understanding and critical thinking within social media debates causes those reading the debates to fall victim to a polarised 'us-vs-them' mindset, as they perceive that these divisive online debates may never reach a conclusion [31,79].
Outside of cyberspace, individuals stuck at a polarised impasse can proactively discuss their views and concerns with an external mediator-who aim to identify common ground, encourage critical thinking, and propose ideas that both parties may agree to.Encouraging critical thinking and consensus-building can bridge the divide in unresolved online conficts and encourage an observing audience to consider both sides of the debate with innovative solutions.
However, online debate moderation (i.e., topic-agnostic enforcement of debate rules) and mediation (i.e., voluntary contributions to the debate topic, such as questioning, claims, and proposed solutions which do not rely on removing content) by human users are not scalable given the size and speed of online discussions.The use of AI as a scalable and proactive instead of a reactive tool in online discourse is increasingly embraced by the HCI community, with research themes such as partisan debate-bots and automated argumentation design [61,63,71], misinformation in health communication through AI fact-checkers [41], and the psychological perceptions of automated systems in online discourse [10,41].
Importantly, current approaches do not incorporate psychological confict resolution theory [2,6,75,88] into the design of AI for confict-mediation online.A mediation-based AI designed around discussion-points, questions, and ideas would encourage further dialogue instead of just deleting controversial posts.This would give users and their audience the choice to either overlook or engage with the ideas and discussion points raised by the mediator-bots.The aim of the mediator's discussion points encourages critical thinking through questioning, solutions, and consensus-driven claims to address social media's culture of antagonism [48,73] and overcome the 'trench warfare dynamics' of polarised discussions [42].
In this work, we design and test mediation strategies to promote collaborative idea-driven mediation on the debate topic over moderating speech itself.Our opt-in mediator-bot design implements the strategies from the Thomas-Kilmann Confict Mode Instrumentwhich provides a framework of fve confict resolution strategies based on cooperativeness and assertiveness [44,75,76,88] (i.e., competing/forceful, avoiding/averting, compromising, accommodating, collaborating).Most crucially, our work relies on tuning prompts with the Generative Pretrained Model-4 (GPT-4 [57,59]) to implement these strategies.These strategies determine how to question, claim, and respond to polarised debaters.
In particular, our study considers the following questions in relation to the strategies we evaluate: RQ1: How do the strategies infuence participants' perceptions of the persuasiveness of the mediator's arguments?RQ2: How do the strategies infuence participants' personal opinion on the debates?
RQ3: Which strategy is the most efective at increasing the audience's perceived likelihood that a debate will reach a consensus?
Our fndings show that high-cooperativeness TKI strategies resulted in a higher Perceived Argument Strength score, indicating that these are more efective in providing persuasive arguments to help mediate polarised online debates.We found various degrees of efectiveness across the diferent confict resolution strategies regarding how likely they are to depolarise the audience.For example, the accommodating strategy was three times more likely to depolarise the audience when compared to having no mediator, while the forceful and the compromising strategies performed poorly in terms of depolarisation.Moreover, audiences exposed to the compromising mediator-bot were 22.2% more likely to believe that the debaters could reach a consensus when compared to having no mediator.
The contributions of our work are three-fold.First, we empirically validate that the psychological TKI confict resolution strategies can be successfully applied to the mediation process of online debates and that a user's preference for a TKI strategy afects the persuasiveness of the strategy.Our manipulation check (Section 3.3.1)also highlights the ability to prompt-tune language models to match specifc confict resolution strategies.Likewise, our psychologydriven GPT-4 prompts are available for users or platforms to experiment with and deploy in their own plug-in mediator-bots for social media.Second, our mediators' ability to depolarise an audience using TKI strategies ofers valuable insights for platforms to design non-invasive proactive depolarisation interventions which ofer avenues to address divisive debates.Through targeting our mediator-bots on the viewing audience, we can address the largerscale challenge of community-wide polarisation rather than just the smaller individual-level (i.e., just between debaters) polarisation process.Third, we provide a mixed quantitative-qualitative datadriven discussion on the applicability, design, and limitations of deploying mediator-bots online.Our fndings beneft HCI and psychology research by also demonstrating the efectiveness of the TKI strategies in a external mediator role rather than the traditional TKI use for active debaters/conficting person-vs-person engagement.

RELATED WORK
To design efective mediator bots, it is important to understand the factors infuencing audience depolarisation and perceived consensusbuilding.In the following sections, we start by defning mediation, moderation and debate facilitation, and their current implementations in HCI research for online confict resolution.We then examine the psychological confict resolution theory to elicit the requirements for our mediator-bot experimental design.

Theoretical Approaches for Resolving Confict
For the purposes of this study, we consider the following defnitions for moderation, mediation, and facilitation in a social media debate context.Firstly, moderation relies on overseeing social media content and enforcing a set of rules pertaining to appropriate conduct, such as removing hate speech, harassment, or spam, as well as controlling the platform of the debate [23,74].A moderator is expected to be topic-agnostic and only intervene when platform policy breaches occur-thus being a reactive measure for promoting healthy debates.Similarly, facilitation replicates the rules-based order of moderation but relies on encouraging users to participate equally in the debate [39,74].A facilitator relies on topic-agnostic rules for ensuring civility (akin to a moderator) but also manages the environment of the debate-such as controlling the conversational tempo of debaters (to reduce spam) [32,45], moderating cross-talk [45], and ensuring that debates remain on topic and (if applicable) on time [32,56].Importantly, facilitators are not expected to contribute to the content of the debate.
Finally, mediation extends debate facilitation by intervening but not overruling a discussion by providing information and ideadriven claims, questions, and proposed solutions to promote interaction and degrade out-group mechanics [74].The objective of a mediator is to encourage cooperation through interactive questioning, providing information/context, and proposing ideas to engage and resolve the debate.Mediators are expected to contribute to the debaters' claims/ideas to help achieve consensus and reduce polarisation [74].A key distinction between meditation and moderation online is that a mediator is a proactive, voluntary and opt-in process based around a topic and addressing grievances for constructive discussions.While mediation can be ignored by the audience or debaters,moderation is a reactive response with removal powers to regulate tone, sentiment, or tempo of a discussion without addressing the underlying ideas [39]-thus it can should only be a 'second line of defence' when mediation fails and/or the discussion violates community guidelines/rules.

Algorithmic Interventions to Address Online Confict
Existing automated interventions in online debates are limited to debate facilitation, which does not include engaging in the debate itself or proposing solutions/compromises.Conversely, the rise of one-sided debate chatbots or social media bots could be exploited as a form of human-targeted polarisation (i.e., infuencing a user towards a specifc belief) [29,71,82].Meanwhile, existing designs for addressing multiparty privacy conficts or smart email replies currently rely on wizard-of-oz style experiments between two human interlocutors [35,66].Likewise, research in chatbots for interactive ideation includes non-social media tasks such as for cooperative design facilitation [70,80], thought-mapping to encourage critical thinking [27] and as an aid to help writers create more persuasive arguments [81].The main challenge of facilitation models is that they do not focus on amending disagreement or reducing polarisation, as debaters may abide by the platform's content rules (enforced by moderators) and debate structure (enforced by facilitators) but still remain in confict and extremely polarised.Thus, moderators and facilitators alone cannot address the pessimism, narcissism, and mental health challenges of widespread division perceived on social media [11,48,73,78].
Our approach hybridises AI-driven debating, deliberating, and consensus-building to collectively form the basis for debate confictmediation [74].The purpose of the questioning and ideation by a virtual mediator is to invoke perceptions of common ground, problem/debate solvability, and to depolarise the audience.For instance, a mediator could facilitate healthy discourse by identifying and questioning common ground (critical thinking) via persuasive arguments (argumentation strength) and co-design solutions (proposing compromises or solutions).
Importantly, we target persuasive argumentation as the referent object for depolarising the audience, which we measure using the audience's Perceived Argument Strength (PAS) [93].PAS refects an aggregate of nine questions into a 1.0-to-5.0score to refect an arguments ability to infuence user opinion and behaviour-with moderate replicability and reliability across psychological studies [4,37,38,93].In HCI literature, Karinshak et al. [41] identifed that using AI-generated pro-vaccination messages resulted in higher Perceived Argument Strength (PAS), and thus more persuasive arguments [93], compared to those created by the United States Centre of Disease Control.

Confict Theory-The Groundwork for Mediation
To create a theory-driven framework for mediation bots, it is essential to understand the psychological strategies that humans employ to resolve confict.In organisational behavioural management theory, Blake and Mouton's managerial grid classifes human confict into the axis of concern for people and concern for production [6].It posits that individuals balance risks (such as the risk of escalating the confict, making enemies, or 'losing' the debate) with potential social or tangible rewards-such as romantic interests, economic gains; collectively also known as Social Exchange Theory [36].Psychological measures for social exchange theory are measured through questionnaires to identify a participant's preferred style to resolve confict.The standard in psychological research is the Thomas-Kilmann Confict Mode Instrument [44]-which extends on the tangible (material gains) vs. intangible (social and metaphysical) cost-beneft interaction efect through a fve-strategy confict instrument with a cooperativeness and assertiveness axis.
It is important to note that all strategies aim to cooperate with both parties to come to an agreement.In the context of the TKI axis, the defnitions of cooperativeness and assertiveness are: • Cooperativeness-the degree of which a party attempts to address the specifc grievances of the opposing party(s) [75].Low-cooperative strategies (Competing, Avoiding) focus on the 'wider picture' and utmost over targeting smaller individual problems/disagreements. • Assertiveness-the degree of which a party pushes their own concerns [75] (in this case, the Mediator's confdence and directness in its questioning and claims to achieve consensus).
For instance, it is not trivial that high-cooperativeness strategies' will result better confict resolution outcomes [44,88].In a debate setting, high cooperativeness would consist of highlighting and questioning specifc grievances-thus expanding the debate and opening more room for criticism which could be seen by a viewing audience as complicating the debate.Likewise, highcooperation can be perceived as overthinking and lead to analysis paralysis where decisions get stuck on specifc inter-party grievance [75,76,88].Conversely, low-cooperativeness approaches such as forcing a solution, diverting (or avoiding) the issues for a less controversial approach, and a neutral/'middle-ground' compromise are valid resolution approaches.Hence, our experiment aims to investigate the efects of cooperativeness and how confdent/assertive the mediator is in intervening in the debate to persuade the audience towards an amicable debate outcome.
The strategies represent each corner of the TKI as well as the central 'Compromising' strategy [88], as visualised in Figure 1.These strategies are typically defned as follows: Alternate measures for identifying a participant's preference of resolving confict include those by Lawrence and Lorsh [49] or Hall [33]-which also derive from the Blake-Mouton grid but difer on collection type (likert vs binary choices) and question categories.However, TKI ofers more consistent results when a participant repeats the measure to (re)identify their preferred strategy (known as test-retest reliability [15]; with the highest performing 0.64 mode score for TKI [76] compared to 0.55 for Hall [33], 0.50 for Lawrence-Lorsh [49], 0.39 for Blake-Mouton [6]).Likewise, the TKI assessment questions are more likely to refect the desired strategies, improving our confdence when making claims regarding a user's preference for a confict-resolution strategy-as measured by Cronbach alpha metric for internal consistency (higher is better [15]; with 0.60 for TKI [76], 0.45 for Lawrence-Lorsh [49], 0.55 for Hall [33]).As such, TKI's consistent results when repeated on the same individual at diferent times (i.e., high test-retest reliability), alongside TKI's relevant questions that match the psychological theory (internal consistency) forms the basis for why we utilise the TKI strategies for our mediator-bot design.
Building on Thomas and Kilmann's work, Womack identifed that users prefer solutions to conficts based on one of the fve TKI strategies [88].Hence, we hypothesise that a mediator who utilises the participant's preferred confict resolution strategy will perceive the matching mediator as more infuential in reducing the diference in agreement between two debaters (i.e., polarisation) and increase the participant's belief that the debaters would reach a consensus.
The TKI measure does not assume neutrality in all strategies [75], nor does mediation itself [74].Thus, we do not require or force the mediator-bot to be 'neutral' or adopt a middle position.Only, the 'Compromising' strategy represents a middle-ground 'give-andtake' approach by design.

Addressing conflicting debaters through audience perceptions.
Polarised online debaters aim to defend and protect their own group's identity (in-group) against those with opposing views (i.e., the out-group) [58]-with Das and Kramer identifying that the presence of a third perspective between diametrically opposed debating parties reduced the aggressive tone and personalised attacks towards the opposing party in Facebook discussions [18].As such, a growing body of research highlights the role of the audience in infuencing debater behaviour to reduce hostility [18,42,50,72].Su et al. identifed that social media users engage in one-sided speech when they are aware that their target viewing audience also hold one-sided beliefs [72].Marwick and Boyd also identifed that social media users placate their narratives and ideology around an 'imagined audience' consisting of all parties engaging in the debate as well as the perceptions of the wider viewing audience [55].
Finally, the perception of the audience is highly relevant to online debates as any future post will be from either a past audience member or the original debaters-who seek to appease their audience.Thus, this symbiotic reader-writer relationship is pertinent to building a culture of open, critical and constructive debate.

METHODOLOGY
The key objective of this study is to utilise mediator strategies through virtual agents to reduce a social media reader's (the audience) personal opinion polarisation on polarised online discussions, as well as reduce their perception that the debaters are polarised.We expect the strategy-driven mediator-bots to provoke constructive questions and solutions and invoke audience introspection on the debates-as these set the requirements for critical thinking and changes in opinion required for depolarisation [90].
We operationalise this objective by measuring the audience's argument strength rating for each mediator (RQ1), as well as their change in personal agreement on the topic after reading the debate transcript (with or without the mediator-bot, for RQ2), and their perceived belief that the debate would reach a consensus given the debaters' polarised views (RQ3).
We grounded our mediator bots in social psychology research, tailoring diferent mediator bots to model the fve TKI strategies.In the following sections, we highlight the design considerations for creating the debate transcripts and the mediator bots and report the results of a manipulation check to verify that the prompts used for our mediator bots match their assigned TKI strategy across the debate topics.

Debate Transcript Design
We drew the debate text from Kialo-an open debating platform with over 2 million user contributions across over 18,000 public debates [43].The academic beneft of Kialo is its bilateral nature, whereby users debate a specifc debate topic with a specifed "For" or "Against" stance-thereby allowing annotated and defnitive polarised stances.Kialo also enables users to create reply-chain threads where users can reply and argue for or against statements.Thus, these reply threads are benefcial as we can replicate the conversational back-and-forth nature of online arguments while maintaining realistic user stances.
As the key focus of the mediator-bots is to voluntarily help remediate the ideas/concepts leading to the increase in social media division/polarity relevant for content-mediation, we do not target the emotive-driven reasons for polarisation-as this is for the separate domain of content-moderation.As such, our idea-driven mediation over emotive-driven moderation approach sees our mediator-bots as a frst-line of defence in the assumption of genuine debate, while personal fame-wars and heated attacks requires a separate 'secondline-of-defence' emotive-driven and topic-agnostic approach.The feld of reactive regulation of tone, sentiment, and tempo is already established in prior content-moderation work, and thus out-of-scope for our novel frst-line proactive mediation approach [39].Thus, Kialo (a platform with content-moderation guidelines similar to Reddit) ofers a compelling ideas-driven baseline 'ground-truth' for our mediator study-while still containing controversial and personal topics such as COVID-19 vaccine mandates and the Russia-Ukraine war.
As our intervention group tests each of the fve TKI confict resolution strategies, we collected fve debates covering social, economic, and cultural topics to represent the variety of topics seen in online debates.Specifcally, the chosen debate topics were: (1) "Public transport should be free." (PT) (2) "COVID-19 vaccines should be mandatory." (CV) (3) "Democracies should take in both Ukrainian refugees and Russians seeking to escape conscription in the Russia-Ukraine War." (UKR) (4) "Facebook should ban political ads with misinformation." (FB) (5) "US intelligence agencies should stop mass data collection." (NSA) We selected the frst top-level response to the debate topic on the condition that it has a reply that is for the statement and a reply that is against the statement to ensure that the debate remains balanced with one parent claim and a reply from both sides.We also selected this statement-reply structure to test the adaptability of our mediator bots to inter-personal confict/arguments, rather than a non-conversational 'list of statements for and against' as seen with competing debate platforms such as ProCon [62].

Bias Mitigation-Debate Transcript Visual Design
We selected Reddit as our design reference for a mainstream discussion platform to emulate due to its familiarity across social media users (unlike Kialo) and its polarised communities [11,28,31], as well as the structure of Reddit's conversations (i.e., grouped topic threads, reply threads).We made the following design choices in our Reddit-style visual design to mitigate potential biases, as also represented in Figure 2: • Debater usernames: Each topic transcript uses a diferent username for the pro and anti-debate topic users (e.g., Person A|B for the COVID-19 topic pro|anti debaters, Person C|D for the Public Transport topic, etc.) to remove any persistent biases from prior arguments.
• Colour-choice: Person A|B, C|D, etc. all utilise opposing colours, with pure red and green removed from the gamut due to their connotations with good or bad [1].• Removed votes/Karma: no comments have any votes/karma or other Reddit tags to avoid conformity biases.• Reply order: Each debater has one main post each, with two replies-one from the (anti-topic) user and another from their (pro) position.The ordering is fxed across all topics and interventions.• Timing-to avoid the perception that the debaters are 'ignoring' the simulated mediator, we indicate that all mediator posts were sent '1 minute' ago from the fake Reddit transcript 'screenshot'.• Mediator-bot username: We do not anthropomorphise nor ofer a gendered name for the mediator bots [3,83,84,87].
Framing our intervention as a 'Mediator-bot' ensures external validity by ensuring that the audience is not deceived into believing the mediator is real-as a real-world use case would see the bots deployed as a collaborative AI tool.Likewise, naming the bot based on <FUNCTION-BOT> emphasises the role and virtual nature akin to prior virtual agents (e.g., Fa-cilitatorBot [39], TaskBot [77], Debbie the Debate-bot [63]).• Subreddit layout: the choice of subreddit name does not refect any major debating subreddits to avoid community bias and stereotypes [11].• Omission of reference links to data sources: the Microsoft GPT-4 adaptation used in this study collects live information from the internet to contextualise and present factual information in its responses [57].GPT-4's sources are provided as citation links (e.g., 'Sources: [1] author, source, date' format).We removed references to the original link in the output citations to avoid biases surrounding the source's validity.• Reply length: we control for the length variance of GPT-4 responses to ensure the overall mediator reply word count are balanced between strategies (+-20% word count variance between speech acts) per topic.We employ a soft-approximate approach by instructing GPT-4 to produce a short paragraph per speech act rather than a hard word-count limit to avoid abruptly cutting of or stifing GPT-4's responses.
Figure 2 displays a debate topic transcript example from the intervention group for the Russia-Ukraine refugees (UKR) topic with the mediator using the Collaborating TKI strategy.

Prompt-Engineering-Designing a Standard Mediator Prompt Template
We used GPT-4 to generate our mediator-bot text due to its better performance compared to other LLMs on reading comprehension (80.9% F1-score on the DROP reading comprehension and arithmetic benchmark [21,59]), and the HellaSwag deductive reasoning benchmark (95.3% with 10-shot learning [59,92]).We used Microsoft Bing's GPT-4 implementation with its online feature to automatically collect new information and context for its responses [57]-ensuring relevance given recent events such as the 2022 Russian invasion of Ukraine, which was not captured in the original ChatGPT/GPT-4's September 2021 corpus [59].Figure 2 displays an example of the mediator's response and knowledge.I understand that you both have different views on how accepting Russian defectors and refugees would affect the war.However, I also want to ask you to consider some evidence from real-world cases of Russian defectors and refugees in Europe.According to some sources [1,2] , there have been several cases of Russian soldiers and officials who have fled their country and sought asylum in Europe, claiming they did not want to participate in the war.According to another source [3] , some experts have argued that offering asylum to Russian deserters could be a way of challenging Putin's regime and supporting peace in Ukraine.
Person E, do you think this is a valid and viable option for democracies?Person F, do you think this is a fair and humane option for Russians?Thus, I propose that we consider creating a joint commission of experts from democratic countries and international organizations that would evaluate the cases of Russian defectors and refugees and decide whether they qualify for asylum or not?This way, we can ensure that the process is transparent, fair, and consistent.

PersonE • 1 hr. ago
This would show that many Russians are against the war.

PersonE • 1 hr. ago
There are things we can do to identify Russians that could help our societies if we let them stay in the west.

PersonF • 1 hr. ago
There is no rational reason why only democratic countries should be compelled to accept refugees.

PersonF • 1 hr. ago
If Russians who are critical of the Putin regime all defect to the West, then it is likely that only loyal soldiers and citizens will remain in Russia.This will increase Putin's influence on the country.

PersonE • 1 hr. ago
If patriotic Russians see their armed forces and fellow citizens defecting to the West, they may begin to question their government and its actions.

PersonF • 1 hr. ago
Accepting Russian defectors and their families may make it harder to influence internal change in Russia.
Mediator-bot 3 • 1 min.ago I think you both have some interesting arguments about how defection would affect the public opinion in Russia.But I also think that this issue is not so simple and predictable.There may be different factors that influence how people in Russia view their government and the war, such as the media, the opposition, the international community, and their own personal experiences.Therefore, I think we should not assume that defection has a clear or uniform effect on the public opinion in Russia, but rather explore the possible scenarios and outcomes.We followed an iterative and incremental prompt-engineering process to ensure the mediator replicated only one of the fve TKI strategies.Throughout this process, the TKI strategy 'Competing' was renamed to 'Forceful' to refect that the mediator themselves is not competing to the debate itself, but forcing its own views into the debate.Likewise, the 'Avoiding' strategy was renamed to 'Averting' to refect that the mediator will always remain present in the debate and not virtually leave/not-respond, but rather try to avert the debate question and instead opt for less controversial solutions as refected in its TKI defnition.All GPT-4 prompts were topicagnostic and were not provided with any contextual information on the topic, with our prompts available in our supplementary material.Our prompts consisted of four paragraphs: (1) A paragraph explaining each TKI strategy from the mediator's perspective.We also use these defnitions to help with the manipulation check to ensure that independent raters can identify which mediator bot output corresponds to the TKI strategy.( 2) The text of the debate, with replies provided as 'Person A, Statement 1: ... ', 'Person B replying to Statement 1: ... '. (3) Response speech act options-consisting of Ask Question, Make Claim, and Respond; this design choice ensures a consistent output structure from GPT-4 and is based on the predicate-logic debate speech acts framework proposed by Prakken [61].All speech acts are equally balanced across the debate transcripts and Mediator strategies.(4) A generic topic-agnostic instruction prompt for GPT-4 to role-play as a mediator whose aim is to achieve consensus while not using the specifc names of the TKI strategies to mitigate bias for those aware of the TKI strategies.

Manipulation Check.
To ensure that our prompts accurately operationalise the TKI strategies, we relied on independent raters to read the mediator bot responses and group them by the TKI strategy.Specifcally, we utilised the Prolifc crowdsourcing platform to recruit 20 participants who were based in the United States to ensure the relevance of the current events topics; were native English speakers, to mitigate challenges with language profciency [32]; and had approval ratings above 97%.All participants were compensated based on the highest minimum wage available across all states of the United States ($15.74USD) at the time of data collection, which is well above the payment recommended by Prolifc [20].
Through the Qualtrics survey platform, raters viewed the fve debate transcripts along with all mediator bot responses representing each of the fve TKI strategies.They then matched each mediator to their perceived TKI strategy.The debate topic and order of strategies presented were both randomised to mitigate bias.
We measured the level of agreement through the accuracy of the raters' classifcations and via Fleiss' Kappa for inter-rater agreement [26].We visualise our raters' classifcations in Figure 3, where we contrast their predictions with the actual strategy the mediator was trying to convey.In general, we obtained 88.21% matching classifcations, indicating that raters were able to identify the intended strategy in most cases.In addition, we attained a mean Fleiss' Kappa of 0.73 (as categorised in Table 1), whereby a Fleiss' Kappa between 0.61 to 0.80 indicates 'substantial agreement' and 0.80-1.00indicates 'almost perfect' agreement between all raters [26].Given our results, we can confdently conclude that our topicagnostic prompts and their responses efectively refect the mediatorbots targeted TKI strategy.

Procedure
We recruited a total of 144 participants with an average age of 39.06 (SD = 13.29) years old using the Prolifc platform, applying the same pay rate and fltering conditions as stated in our manipulation check, with the addition of equal-split balancing for political party preference (Republican, Democrat, Independent).The sample size was based on a G* Power analysis using a 2 efect size of 0.15 (indicating a medium-efect size as used in similar psychological intervention studies [14,47,53,68]).
Our survey consisted of several sections and took approximately 45 minutes to complete.Figure 4 outlines our experimental workfow.Our experimental design was approved by our university's Human Ethics Committee.
Stage A started by providing the participant with a plain language statement, explaining the details of the experiment.This was followed by a 'Pre-task Questionnaire', in which we collected participant information, including demographic data (age, gender and education level) and their prior online debate experience through two 5-pt Likert scales ('Never', 'Sometimes over the past six months', 'Sometimes over the past month', 'Weekly', 'Daily').Participants were asked "How often do you read/watch (without posting) debates on social media?" for reading/passive experience, and "How often do you post and actively partake in debates on social media?" for writing/active engagement.Participants were also asked to complete the Adkins Self-Questionnaire, containing 15 questions across the fve strategies to categorise one's preference for TKI assertiveness and cooperativeness in relation to the fve strategies [2].This tested if TKI preference enhances the corresponding mediator strategy's efectiveness.We utilise the normalised ratio of the participant's preference for assertiveness and cooperativeness as the two TKI preference indicators-representing where they stand on the TKI assert/coop-axis quadrant from Figure 1.We also collected the participants' pre-existing topic agreement on each of the fve debate topics using a 7-point Likert scale ('Strongly Disagree' to 'Strongly Agree').Next, participants were assigned to either the control group (without any mediators) or the intervention group (with one of the fve mediator strategies per transcript, counterbalanced).Stage B refects the iterative process whereby users read the diferent debate transcripts.After reading each transcript, participants were asked to answer the following questions covering each of our research questions: After iterating over all fve debate topic transcripts in their respective group (control or intervention), participants were asked a fnal exit survey question (Stage C) on their opinion on "What are your thoughts on deploying virtual bots to mediate debates in social media?".This was followed by a message thanking participants for completing our experiment.

RESULTS
We employ a mixed-methods approach with linear mixed models and thematic analysis to address each one of our research questions.We utilise a Generalised Linear Mixed Model (GLMM) to test the PAS values across the diferent strategies which account for our predictor values (RQ1).We utilise Cumulative Link Models (CLM) to represent the ordinal Likert data across the mediator strategies, targeting personal topic opinion polarisation (RQ2) and perceived debater consensus/polarisation (RQ3).We consider the mediator strategies in both their one-of-fve strategy form (Accommodating, Collaborating, Compromising, Averting, Forceful), and in their highmedium-low Assertiveness and Cooperativeness axis form, which we infer from the TKI quadrant visualised in Figure 1.We also consider the participant's TKI preference from the Adkins TKI self-questionnaire-where we utilise their preferences towards the cooperativeness and/or assertiveness axis.
We measure the efect size of the strategies and relevant predictors using unstandardised coefcients for PAS response and Likert scale data, alongside Odds-Ratio (OR) and Cohen's d standardised efect sizes for binomial data (such as whether the mediator depolarised the audience participant or not) [14].For pairwise comparisons, such as comparing the mediator TKI strategies, we utilise Estimated Marginal Means comparisons (i.e., emmeans, also known as least-squares means).Emmeans are useful for representing the wider population as they account for the efects of each variable/predictor in our models as opposed to the adjusted raw sample means.For signifcance, we provide p-values and Standard Error (SE) for each efect-size claim.We compute the Variance Infation Factors (VIF) to check for multicollinearity across our independent variables for each mixed model.All VIFs for our models were below 5, indicating a lack of linear dependency among the independent variables [64].Our model outputs (all generated in R) and fnal topic-agnostic GPT-4 prompts for each strategy are available in our supplementary material.
Finally, we conducted a deductive thematic analysis of our qualitative responses, which involves coding them into themes based on argument strength (RQ1), critical thinking (RQ2), and confictresolution (RQ3) theories.Deductive thematic analysis involves an iterative process of familiarising frequent concepts in the data, grouping trends as themes, reviewing themes with the theory [7].We focus our theory-based analysis on Critical Thinking and speech act models [61,89,90], and the psychology of the TKI strategies [44]).We iterate over the data across three passes and present the relevant information for each research question.

RQ1-Perceived Argument Strength by TKI Strategy
Table 2 shows the mean PAS for each strategy and all strategies combined.It also shows the mean PAS of the Human Debaters as rated by our participants in the Control condition (no mediator) as a reference.
We then tested the diference in Perceived Argument Strength values (1.0 to 5.0) across the fve mediators from the 120-person intervention group and 24 control group (with the debaters argument strength as the measure).We found a statistically signifcant relationship between the mediator strategy and its perceived argument strength-whereby Collaborating, Compromising, and Accommodating (i.e., the more cooperative strategies) outperformed the Forceful and Averting strategies ( = 0.38, SE = 0.07, p < 0.05).Likewise, Collaborating and Compromising (i.e., the cooperative and assertive strategies) had statistically signifcant PAS improvements over the control debaters' argument strengths (up to = 0.45, CHI '24, May 11-16, 2024, Honolulu, HI, USA SE = 0.12, p < 0.01).All pairwise strategy comparisons are outlined in Table 3. Figure 5 highlights the distribution of PAS scores given to each strategy.The full fxed-efects model output is available in the supplementary materials (R 2 : 0.39).Prior work has highlighted that people prefer resolving conficts linked to their preferred TKI strategy [2,44,67,75,76,88].Hence, we use our GLMM model to cross-analyse each participant's preferred TKI strategy and compare their assertiveness and cooperative scores to the strategy used by the mediators.We observed a statistically signifcant positive relationship between a user's preference for more cooperative strategies and PAS scores for the medium-cooperativeness Compromising ( = 0.102, SE = 0.036, p < 0.01) and high-cooperativeness Collaborating ( = 0.074, SE = 0.036, p < 0.05) mediators.We also observe a statistically signifcant and moderate increase between preference for more assertive mediation strategies and PAS scores for the mediumassertiveness Compromising strategy ( = 0.177, SE = 0.080, p < 0.05).
Conversely, Forceful and Averting lacked a statistically signifcant relationship between preference for these high-assertiveness or low-cooperativeness strategies and PAS scores ( = 0.02, SE = 0.04, p = 0.63 and = 0.07, SE = 0.08, p = 0.41).Stronger arguments had a lasting psychological impact, with the more cooperative strategies promoting critical thinking-as imbued in statements when asked whether the mediator made them refect on the debate; "Absolutely, yes.The approach allowed me to keep an open mind and try to fgure out solutions, rather than solely defend my opinion/my side" (P67) and that "The moderator brings a way for both parties to understand the pros and cons of each other's statements" (P31).
Participants identifed the high cooperativeness and assertiveness collaborating mediator as quite successful in providing persuasive arguments (as also shown in our quantitative fndings), with statements like; "The [collaborating] mediator followed an effective, fair, and formulaic strategy: acknowledge both sides, prompt for sources/further information, provide a diplomatic solution.I believe that this has been the mediator's best approach/strategy thus far" (P67).The high-assertiveness and low-cooperativeness forceful mediator was far more divisive.Some participants perceived it as "rational" (P2) and ''direct" (P17, P21, P49) in its approach to win over the debaters to a specifc side with claims such as "I enjoyed the way it called out person F!" (P1) and "It called out both people on this topic which I enjoyed" (P99).However, others criticise the mediator's "aggressive" (P60, P81, P112, P118) approach in that it did not seem receptive to discourse; "right out of the gate turned me of and made me shut down" (P52) and that the Forceful mediator appeared "combative and contrarian" (P19).
Regarding the tone of the mediator, some participants found that low-assertiveness strategies had a polite depolarising efect on the perception of the debaters through casualising the discussionimbued through claims such as "I think that the mediator was exactly correct here, in trying to reroute the discussion to something that was actually on topic and not needlessly infammatory and judgmental" (P18).Likewise, the casual/polite persona and recommended (but not forced) solutions enhanced the perception that the debaters were not polarised; "The bot was good at proposing middle ground solutions" (P51).
However, while low-assertiveness resulted in the perception of "calm[ing] things down" (P98) for some participants, its arguments attained the lowest PAS values in our quantitative analysis.Several participants echoed this degradation of PAS due to the mediator's low assertiveness as they felt these mediators appeared to lack confdence and saw the debate as unsolvable.Participants expected the mediator to take a more authoritative role to be more than just another debater "in which it's trying to act more human and cavalier in order to be more accepted.It's a technique that generally backfres pretty badly" (P74) and that the mediator "sounds slightly unprofessional, and should have provided sources instead of saying "I read some things" etc." (P17).Furthermore, they perceived the informality of an unassertive casual approach as a negative trait of the mediator-bot as "It detracted from the ideas the mediator put forth" (P70).In some situations, the low-assertiveness strategies' "politeness" (P57, P21, P2, P89) was even misconstrued as passiveaggression which "came across as pretentious and unproductive." (P25) and that the "mediator's tone was weak and too soft, which made it seem less credible." (P67)-thus showing a desire for more direct argumentation.

RQ2-Change in Audience's Personal Opinion via the Mediator-bots
Next, we evaluated participants' change in their own personal opinions about the debate topics.We present our full model in the supplementary material.Table 4 shows a statistically signifcant relationship between the mediator strategy and whether their posttest (i.e., after reading the transcript) topic agreement Likert score difered from their original pre-test topic agreement Likert score (i.e., before reading any debate transcripts).We observed that strategies that lean towards the cooperative TKI axis (i.e., Accommodating, Collaborating, and Compromising) had a statistically signifcant infuence in changing their opinion compared to the control.The highest-performing Accommodating strategy was 2.8 times more likely to change their opinion compared to the Control (OR = 2.817, Cohen d = 0.571, p < 0.01; as derived from Table 4).Our model also highlights that users who read or write in online debates 'sometimes over the past month' are approximately half as likely to change their mind as those who read online debates 'daily' (OR = 1.584, Cohen d = 0.254, p < 0.05 for reading; OR = 1.917, Cohen d = 0.359, p < 0.05 for writing).
Beyond binary opinion changes, we also categorise if the audience were depolarised, as defned by the diference from the neutral stance in the 7-pt topic agreement Likert scale between the participant's pre-and post-test topic agreement value.Thus, those moving towards the centre after our manipulation were considered as depolarised while those that moved away from the centre were considered as polarised.

Efect of TKI Preference on Audience Opinion Depolarisation.
Users who preferred cooperative strategies had a slight increase in the likelihood of being depolarised when subject to the Compromising mediator (OR = 1.460, p < 0.05) with a small efect size for TKI preference in audience opinion depolarisation (Cohen d = 0.209).
Figure 6 displays the movements of user opinion, denoting the (de)polarisation process.We observe that participants were approximately three times more likely to be depolarised from their original stance when exposed to the Accommodating mediator compared to the no-mediator Control debate transcript (OR = 3.001, Cohen d = 0.606, p < 0.05).Interestingly, the Compromising mediator had a somewhat blowback dividing efect, being the strategy that led to the highest number of polarised participants.In addition, Averting's focus on defecting the debate towards less controversial areas/topics also had notable depolarisation efect despite its low-cooperativeness/low-assertiveness.

Critical Thinking and Change in Opinion.
Though the audience was not active in the debate, they evaluated and raised questions regarding the mediator's role in helping them engage in critical thought.The Collaborative mediator encouraged introspection even in cases where participants did not change their original opinion, with participants claiming that "...the bot made me think more about where I actually stand on this.I think I need to do some more searching because free speech is important to me" (P81), and that "It [the mediator] did help me refect, but caused me to feel stuck on deciding one way or the other.I can't choose now because I am trying to weigh each side." (P119).Likewise, participants stated that they tended to respond more to this mediator's ideas with comments like "... it allowed me to keep an open mind and try to fgure out solutions, rather than solely defend my opinion/my side" (P67) and that it used a "strategy of encouraging participants to question their beliefs and explore the topic deeper [which] is a proactive approach to facilitating meaningful discussion and potential resolution" (P107).
Conversely, the high-assertiveness but low-cooperativeness Forceful strategy made participants engage in negative information denial and analysis-whereby participants sought to fnd reasons to invalidate arguments.This mediator had an efect of further polarising preexisting beliefs, with claims that it made participants "double-down" (P86, P39, P73, P84, P89) on their view, and could also aggravate them to feel "more inclined to respond with hostility to those that hold opposing beliefs" (P89).
Interestingly, despite Accommodating and Compromising both being cooperative strategies, Compromising had an unexpected polarising efect.Participants either embraced or rejected the mediator's applicability and realism, with claims that a "quick and easy solution to a complicated topic like mass surveillance is hilariously naive." (P110), thus its compromise-based strategy was insufcient for eliciting critical thought in nuanced discussions.Meanwhile, Accommodating's polite and unassertive questioning "validated their points" (P66) and made the participants feel that the mediator had a use in building "civility in our discussions but to also combat the rise of misinformation and propaganda" (P29).
Surprisingly, the low-assertiveness / low-cooperativeness Averting mediator was mildly successful with regard to depolarising participants.The participants perceived the Averting mediator as difusing hostility by proposing alternative, less controversial solutions, questioning the relevance of the debate topic and attempting to adjourn it.For instance, users claimed that Averting was "...an efective strategy for sensitive topics especially" (P67).However, the audience contested Averting's efectiveness as a mediation strategy due to its nature to avoid resolving the specifc issue, with users claiming that this approach was "counterproductive... [as] arguing can be important sometimes and this disagreement [NSA surveillance debate] was happening with civility, with people just explaining their opinions -no name calling or personal attacks." (P118).

RQ3-Perceived Likelihood that a Debate will reach a Consensus
We found that the more cooperative strategies resulted in a higher perceived likelihood that the debaters could reach a consensus ( = 0.867, SE = 0.164, p < 0.001).The Compromising mediator (mediumcooperativeness, medium-assertiveness) was the most successful for reducing the perception of audience polarisation compared to the Control 'no-mediator' group ( = 0.719, SE = 0.239, p < 0.05; when compared across all strategies in Table 5), followed by the Accommodating and the Collaborating mediators.The Forceful moderator had a perceived polarising efect on the debaters in addition to themselves (RQ2 -Figure 6), while the Averting moderator also performed poorly in perceived debater consensus-building even though it was relatively successful in depolarising the participants (RQ2 -Figure 6).We present all pairwise comparisons in Table 5 with the visualisations for each respective strategy shown in Figure 7.The audience's personal TKI preference did not have a signifcant relationship in improving the impact of the mediator unlike the audience-targeted RQ1 PAS and RQ2 personal agreement.Other relevant predictors included that those with more experience in writing/engaging in online debates such as those who write 'sometimes over the past six months' compared to 'daily' were more likely to perceive that the debaters would reach a consensus ( = 0.805, SE = 0.376, p < 0.05).Conversely, those that frequently read online debates were less likely to perceive that the debaters would reach a consensus (i.e., reading online debates/discussions 'daily' compared to 'sometimes over the past six months'; = 0.477, SE = 0.213, p < 0.05).
Furthermore, we aggregate the 7-pt Likert values into the trinomial categories of the belief that the debaters 'will not reach consensus' (Likert values 1-3), have an unsure 'Neutral view' (4) and the belief that the debaters 'will reach an agreement' (4-7).Figure 8 highlights that the mediator bots had a signifcant impact on the audience in reducing their perceived polarisation-thus helping achieve our aim to detoxify the polarised perception of social media.For the Control 'no-mediator' group, only 26.67% of participants believed that the debaters could reach a consensus compared to 44.17% for the Compromising mediator.Conversely, the Forceful mediator increased the perception of polarisation compared to the cooperative strategies and control-with only 17.5% believing that the debaters could reach a consensus.

Co-opting Optimism vs. Nihilism.
For the Compromising responses, users co-opted a belief that a compromise was necessary (even though the term never appeared in the transcripts).This was seen as being useful in cases of strong perceived disagreement between the debaters, such as perceiving  that "The views are far apart but I think both realize there is no complete situation of either or.I think with concessions and compromise both could be satisfed" (P29) and "I think [Debaters] F and E, would eventually compromise on the issue as the mediator suggested.Both have valid points, and should take in account a solution that addresses both." (P33).Moreover, the high-cooperativeness mediators tended to lead to more optimism in the sentiment of the audience responses, such as stating that "The mediator took the debate from only thinking about possible scenarios to reality, and I think that is very important.It's easy to create "what if's" in our minds with our own personal bias." (P91).As well as seeing the mediator as a "proactive approach to facilitating meaningful discussion and potential resolution" (P107)which stresses the aim to produce proactive interactive solutions to combat polarisation rather than reactive solutions from human moderators, such as banning, blocking, or bullying.
Conversely, our qualitative responses explain the Forceful mediator's strong impact on perceived polarisation (Figure 8) as the users co-opted the mediator's assertive and low-cooperative mentality.Specifcally, the audience's responses when presented with the Forceful mediator resulted in them co-opting negative and destructive criticism, with participants claiming that the debaters' arguments were "very weak, and both can be easily dismissed" (P109) and that "there is no middle ground that they can inhabit" (P74).

DISCUSSION
In our work we aim to identify ways in which AI-powered bots can have a positive efect on the perceptions of discourse in online social media.The proliferation of chatbots and language models has raised fears that social media will soon be plagued by automated bots that spread rumours, misinformation, and lead to polarisation.Recent announcements regarding social bots by Meta [22] and Discord [5] have not curbed these concerns, and instead have fuelled further worries regarding privacy and addiction.
Our work has shown the potential benefts of psychologicallyinformed AI mediator-bots on users.Specifcally, we choose to focus on debate audiences rather than debaters themselves, because rapidfre online debates occur only briefy, whereas their transcripts exist in perpetuity for many to read.Hence, focusing on audiences ensures that the technology is relevant to more people, and is also 'backwards compatible' in the sense that our proposed bots can mediate debates that have already taken place.
The potential benefts are substantial.Primarily, we argue that the mediators we propose can address the harmful culture of social media-with its perceptions of insolvable divisions [91], toxic 'fame wars' [29,78], and perceptions of nihilism due to a lack of cooperation and solutions [48,73].The voluntary implementation of AI mediators by social media platforms would refect their commitments to a safe and open social media [12,13,24], where these mediators could be deployed as a client-side extension to users' browsers or as a voluntary feature in online threads.
Next, we outline the implications and applicability of our fndings in regards to the mediators' persuasiveness (RQ1), psychological impact (RQ2), and for perceived division (RQ3).We then contextualise our fndings in regards to our target practitioners of social media platforms (with regards to our prompt-tuning approach), alongside HCI and Psychology researchers.Finally, we refect on potential future work to build on our fndings and further contribute to improving social media culture via voluntary interventions.

Persuasiveness of Mediator Bots Strategies (RQ1)
While high-cooperativeness strategies were generally more persuasive, PAS was the only metric for which the assertiveness of the mediator had a moderate and signifcant efect due to the audience's perception that assertive behaviour imbues confdence (per our qualitative data).This corroborates with prior work on political (mis)information that showed that confdence in information presentation is a signifcant factor for improving vaccine uptake when using COVID-19 chatbots and conversational agents [41,51], as well as for enhancing user trust of AI decisions [9,10,17,41].However, while moderate assertiveness is necessary to demonstrate confdence, high-assertive behaviour can detract the focus of the audience away from the debate itself and redirect it towards the mediator-as seen in the polarising/divisive efect of the Forceful mediator in Figure 7.The balance of assertiveness corroborates with riot control dynamics, whereby confrontational rhetoric can escalate tensions and provoke a defensive reaction from the protesting or rioting groups through creating a majority vs authority mobmentality dynamic [46].Similarly, in online discourse, the assertive behaviours from authoritative fgures can trigger community-level radicalisation through a cascading belief stampede [25].Thus, highcooperativeness is an essential mediating force for assertiveness to avoid polarising and immolating the debate.
More generally, PAS can act as an overall 'gateway' variable for analysing argumentation, as it acts as a generalised argumentation and persuasion 'performance' measure.Critically, PAS aims at convincing and changing audience perceptions [4,37,81,93]-which is required for the mediator's depolarising arguments to have its intended efect.We envision that PAS could become a standardised metric for future mediator-bot designs-where social media platforms or researchers could substitute the language model based on other argumentative NLP studies which utilise PAS as a measure for conversational language model performance [51,81].PAS is currently a common benchmark metric for analysing Public Service Announcement strategies [4,38,93].Thus, PAS can be applied as an approximate benchmark for potential mediator-bot performance of a language model.

Mediator Strategy can Impact People's Beliefs (RQ2)
By design, content moderation must be consistent, enforced, and reactive-as the failure to uphold consistent fagging or removal of content can be perceived as a form of platform bias, which can lead to a sense of injustice and polarisation [28].Content mediation intends to reduce perceptions of polarisation before they necessitate immediate removal, as well as reduce the toxic divides in modern discourse through the mediator's questioning and solution-building approaches.
Mediator-bots could act as proactive and preemptive tools to combat the degradation of online debates into polarised hostile groups which imbue current social media culture [11,54].Further, our mediator bots do not rely on stifing speech or controlling the discussion to reduce polarisation, unlike moderators-as evident by the impact of our more successful mediator-bots as shown in Figure 8, without any changes or even responses from the debaters.Thus, our argumentation-based approach to depolarisation adds to the toolkit for reducing online polarity by platforms, researchers, and the government.
Importantly, our approach underlines the commitment by platforms and governments to develop non-invasive interventions to combat polarisation and extremism agreed upon in the transnational Christchurch Call [12,13].As such, platforms could either implement mediator-bots across discussion threads-such as debate or news subreddits or utilise an external measure to justify when the mediator-bot should intervene.For instance, a mediator-bot could be contingent on keyword or sentiment-detection models to detect extreme polarity (such as extremism or hate-speech detection models [28,29,60]) and interject thereafter to utilise the calming presence experienced via the Accommodating strategy, or even for the Averting mediator to deescalate the debate and divert towards less controversial solutions.
For researchers, our fndings extend on prior work by identifying that TKI preference is signifcant and also applies to external third parties (in this case mediators) in addition to the person themselves [76,88].Thus, platforms could utilise a one-of self-questionnaire on registration to personalise which mediator strategy they personally see to maximise the personal depolarisation efect.Moreover, researchers could extend our mediation strategy fndings (e.g., Accommodating as the highest-performing strategy) to test apolitical mediation scenarios-such as testing the TKI-strategy mediator-bot for resolving personal online disputes or for persona-driven social bots (such as those proposed by Meta [22]), and for mental health support [78].However, future work in human oversight of AI decision-making and discourse, as well as conversational ethics and its societal implications, is critical to ensure robust AI interventions that protect at-risk groups from AI mistakes or inadvertent risks such as alienation [34].

Perceived Consensus-Building in Online
Debates (RQ3) It is crucial to reduce the perception that debates are polarised and unsolvable, as it can negatively impact the audience's mental health through disengagement in social media discussions [19].In addition, this has been shown to lead to dangerous existentialist and nihilistic outlooks on society [48,73].In our work, we found a general trend that the cooperative strategies were more successful at increasing the perceived likelihood of debaters' consensus, with participants co-opting a more optimistic approach to consensusbuilding amongst debaters.This fnding matches previous work that found that injecting more optimistic news into divisive conficts moderately improved the self-reported optimism and positive experiences of social media users' [8].Conversely, the Forceful mediators led participants to project more negative feelings, resulting in a reduction of participants' belief that the debaters could reach a consensus.However, unlike the audience-targeted argument strength (PAS) and personal opinion, there was no signifcant relationship between preference for the TKI axis and any of the used strategy.The lack of TKI preference impact is likely due to the diference between personal opinion and perceived (debaters) opinions.Specifcally, the reader's preferred TKI infuences PAS values as PAS is an introspective judgement on the argument within the wider confict, which relies on the reader's personal view on the mediator's points.Conversely, the lack of a link between preferred TKI strategy and perceived likelihood of the debaters reaching a consensus is likely a result of the participants' projected view on how others would view the confict-resolution strategy rather than refecting on their own view and preference.

AI Strategisation in Human-AI Collaboration
Our fndings have important implications for social media users and platforms, and HCI/Psychology researchers.Firstly, our manipulation check highlights that Large Language Models (LLMs) can replicate specifc psychological strategies with high confdence (with substantial inter-rater agreement).Researchers can beneft from these fndings as they demonstrate that LLMs can be promptpersonalised to operationalise, automate, and test social and psychological theories at scale.Likewise, the fndings that assertive and cooperative strategies improve argument strength could be useful for tools to improve formal persuasive writing-which could supplement or improve writing aides like Grammarly [30] through using psychologically-driven strategies.This prompt-driven approach can be utilised for creative ideation and dispute-resolution -such as utilising the Accommodating mediator as a bot in Slack or Teams workplace discussion, or in informal Reddit-like discussions.
The hybridisation of formal psychology-driven strategies and an open-ended bot approach enables researchers and online communities to engage with AI as a collaborative partner in resolving online and hybrid disputes, ideation tasks, and persuasion needs-with the potential to extend the mediator-bot design as a future business or legal partner/tool.Our fndings advance mediation through testing confict resolution strategies on a third-party rather than a debater-identifying that the audience have a slight preference for a mediator which has a strategy that matches their preference as well as the trend that individual grievance/disagreement solving (high-cooperative) strategies are more persuasive (RQ1), encourage critical thinking on pre-existing beliefs (RQ2), and more efective at increasing the plausibility for debater consensus (RQ3).

Limitations and Future Work
Our study seeks to introduce psychology-driven strategies for online debate depolarisation in a simulated environment with realworld data.However, our experimental design only considers the audience response rather than the potential debaters response.Our focus on just mediator replies without debater responses highlights the common scenario where a user skims a high-level subreddit topic thread rather than reading the individual in-depth reply chains.In context, prior work identifed that reading online argumentative text reduced in-depth and focused reading compared to general expository text, whereby users tended to jump between top-level points/paragraphs in the argumentative text (i.e., "non-linear reading") [52].
Furthermore, the added beneft of this approach is that it avoids the confounding efect of positive or negative debater responses or tangents interfering with the mediator's strategies/content.By letting the audience evaluate the mediator's replies, we ensure that they judge the mediator's efectiveness in their depolarisation process based on their own criteria, not peer pressure and social conformity [85][86][87].Nonetheless, future work should expand on our study by focusing on the role of the mediator in countering emotive debater responses-to test the robustness of the mediators in intradebate disputes.Real-time experiments between debaters could also test the utility of conversational and context dynamics, as well as the mediator's ability to withstand and counter abusive responses or hate speech.In addition, future work in language model-based mediation targeting debaters themselves could involve exploring and enforcing specifc moderator rules to improve consensus-building, such as the infuential factors of conversation tempo [45], encouraging involvement of under-represented debaters [39,87], and voting on the debate topic or mediator(s) responses [87].
The purpose of the mediator-bots for practitioners/platforms is not to replace but to supplement content-moderation systems.Our mediator-bots target the topic-specifc ideas and concepts in a debate with the intent to promote critical thinking with our Questioning, Claim speech acts; and promote consensus through our Respond speech acts (such as the proposed solution shown in our example Figure 2).We envision future work to highlight trigger-mechanisms to detect when to engage in content-moderation to regulate the tone, semantics, and conduct of a debate as a second line of defence when the mediator-bots idea-driven depolarisation techniques must be supplemented with emotive-driven deescalation techniques.However, content-mediation should not stymie or override debate as polarised discussions are not always negative.Thus, mediation suits voluntary opt-in collaborative idea-building and critical thinking while moderation is suited for regulating and deescalating speech that has strayed from the platform or subreddit/thread community's rules (such as addressing trolling or countering hate speech).
Future work should also consider interdisciplinary evaluations of whether AI mediation should pursue neutrality vs. principlesa topic which remains unresolved even in the feld of mediation itself [74].Moreover, the presentation of information (facts and sources pulled by GPT-4 [57]), as well as the perspective of the mediator, may not necessitate 'a compromise' or even a debate.In these cases, this is where human oversight and content-moderation comes into play-as a debate that violates rights/human dignity (e.g., violent/hate speech) likely mandates a reactive moderation rather than proactive mediation solution.Thus, future work should consider the potential ethics of language models and their own cognitive/information biases in AI collaboration activities-and create prompt-engineering methods to counter-bias LLM responses.

CONCLUSION
Content mediation ofers an additional collaborative avenue to depolarise social media audiences on divisive topics.We investigate a prompt-tuned GPT-4 mediation design based on the fve TKI confict resolution psychological strategies.Our fndings highlight the efectiveness of diferent strategies in terms of their Perceived Argument Strength (PAS), their depolarising efect on the audience of a debate, and the audience's perceived likelihood that the debaters will come together and reach a consensus.We present our designs, prompts and analysis framework to enable social media platforms to consider new proactive depolarisation measures before resorting to reactive moderation, which can be perceived as heavy-handed and stifing.Our fndings also highlight that user preferences towards the TKI cooperativeness/assertiveness axis improves the mediator's argument persuasiveness and depolarising efect, enabling platforms to consider stylising mediator-bots to optimise and enhance users' interaction in online discussions.Our results indicate that mediator-bots can improve the user experience of social media through detoxifying the hostile and polarised culture of online discourse.By bridging the divide of modern social media discourse, we can mitigate the negative mental health efects caused by online polarisation, all while protecting open and secure speech.

Figure 2 :
Figure 2: Layout for the Russia-Ukraine debate topic (UKR), displaying the pro-stance (Person E) and anti-stance (Person F) debates, alongside the in-depth (high-cooperativeness, high-assertiveness) Collaborating strategy mediator.

Figure 4 :
Figure 4: The experimental fow of our study-covering demographic, predictor and pre-test questions in Stage A, the iterative per-transcript questions across the separate Intervention and Control groups in Stage B, and the post-task exit question in Stage C.

Figure 5 :
Figure 5: Distribution of Perceived Argument Strength Scores between the Debaters (Control) and the Mediator Strategies.

Figure 6 :
Figure 6: Audience change in opinion between strategies and the no-mediator Control group, as categorised between being polarised, depolarised, or neither/no change.

Figure 7 :
Figure 7: Distribution of the perceived likelihood that the debaters would reach a consensus (7-pt Consensus Likert) between the intervention and control groups.

Table 1 :
Fleiss' Kappa agreement for each one of the debate topics.
[37]7,93]e strategies infuence participants' perceptions of the persuasiveness of the mediator's arguments?-QuantitativeMeasure:PerceivedArgumentStrength (PAS) score for the Mediator (intervention-group) and Debaters (control-group, for reference only)-a numeric score between 1.0 to 5.0 based on 9 Likert questions.PAS provides an objective means of gauging the impact of diferent rhetorical strategies and persuasive techniques and enables the evaluation and refnement of counterarguments, as it enables researchers to pinpoint the weaknesses in opposing viewpoints as perceived by the target audience.PAS demonstrates high replicability[4,37,93], unlike measuring argument strength as a binary (strong vs weak) option or as a single Likert scale[37].-Qualitative Measures: An open-ended question on the mediator's approach towards mediating the debate: "What were your thoughts on the mediator's strategy to mediate in this debate?".RQ2: How do the strategies infuence participants' personal opinion on the debates?
RQ1: -Quantitative Measure: Topic Agreement Score-comparing Pre-test debate topic agreement vs. Post-test debate topic agreement 7-pt Likert between 'Strongly Disagree' to 'Strongly Agree'.Namely, we measure the audience's agreement on the debate transcript's question (such as "Public transport should be free") prior to showing the debate transcript and after they completed steps within that particular transcript.-Qualitative Measure: An open-ended question on the participant's introspection and refection on the debate to investigate the logic and causes between their shift (or lack thereof) of opinion after reading the debate: "Did the mediator help you refect on this debate's topic?If so, in what way?" RQ3: Which strategy is the most efective at increasing the audience's perceived likelihood that a debate will reach a consensus?-Quantitative Measure: Perceived Consensus Score-a 7-pt Likert between 'Extremely Unlikely' to 'Extremely Likely' on "How likely do you think that Person A and Person B would come to an agreement or solution to the debate?"-Qualitative Measure: An open ended response to "Explain why you chose your prior answer."

Table 2 :
Descriptive statistics on PAS scores for the diferent mediator strategies, and human debater PAS scores (from all debate topic transcripts in the control condition) for reference.

Table 3 :
Pairwise comparison of PAS coefcients across mediator strategies (Model A) and across the assert/coop TKI axis (Model B).
4.1.2Perceptions of the Mediator's Approach.

Table 4 :
Pairwise comparison of binary topic opinion change (i.e., if post-test agreement difered from pre-test), across mediator strategies and control (Model A) and across the TKI axis (Model B)).Contrast (across TKI strategies and control) Estimate (log-odds) Standard Error (SE) p-value

Table 5 :
Pairwise comparison of perceived likelihood that the debate will reach a consensus across mediator strategies and the control (Model A) and across the overall TKI axis (Model B).