Authors: (1) Yi-Ling Chung, The Alan Turing Institute (ychung@turing.ac.uk); (2) Gavin Abercrombie, The Interaction Lab, Heriot-Watt University (g.abercrombie@hw.ac.uk); (3) Florence Enock, The Alan Turing Institute (fenock@turing.ac.uk); (4) Jonathan Bright, The Alan Turing Institute (jbright@turing.ac.uk); (5) Verena Rieser, The Interaction Lab, Heriot-Watt University and now at Google DeepMind (v.t.rieser@hw.ac.uk). Authors: Authors: (1) Yi-Ling Chung, The Alan Turing Institute (ychung@turing.ac.uk); (2) Gavin Abercrombie, The Interaction Lab, Heriot-Watt University (g.abercrombie@hw.ac.uk); (3) Florence Enock, The Alan Turing Institute (fenock@turing.ac.uk); (4) Jonathan Bright, The Alan Turing Institute (jbright@turing.ac.uk); (5) Verena Rieser, The Interaction Lab, Heriot-Watt University and now at Google DeepMind (v.t.rieser@hw.ac.uk). Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Background 2 Background 3 Review Methodology 3 Review Methodology 4 Defining counterspeech 4 Defining counterspeech 4.1 Classifying counterspeech 4.1 Classifying counterspeech 5 The Impact of Counterspeech 5 The Impact of Counterspeech 6 Computational Approaches to Counterspeech and 6.1 Counterspeech Datasets 6 Computational Approaches to Counterspeech and 6.1 Counterspeech Datasets 6.2 Approaches to Counterspeech Detection and 6.3 Approaches to Counterspeech Generation 6.2 Approaches to Counterspeech Detection and 6.3 Approaches to Counterspeech Generation 7 Future Perspectives 7 Future Perspectives 8 Conclusion, Acknowledgements, and References 8 Conclusion, Acknowledgements, and References 5 The Impact of Counterspeech The concrete effects of using counterspeech remain debated. The methods applied for evaluating the effectiveness of counterspeech vary considerably across studies in the field. In this section, we outline eight aspects that can help to better understand the impact of counterspeech. Research design A wide range of methodologies have been adopted to assess the impact of counterspeech on hate mitigation, including observational studies (Ernst et al., 2017; Stroud and Cox, 2018; Garland et al., 2022), experimental (Munger, 2017; Obermaier et al., 2021; Hangartner et al., 2021) and quasi-experimental designs (Bilewicz et al., 2021). In observational studies, investigators typically assess the relationship between exposure to counterspeech and outcome variables of interest without any experimental manipulation. For instance, a longitudinal study of German political conversations on Twitter examined the interplay between organized hate and counterspeech groups (Garland et al., 2022). There is also an ethnographic study interviewing counterspeakers on Facebook to understand external and internal practices for collectively intervening in hateful comments, such as how to build effective counterspeech action and keep counterspeakers engaged (Buerger, 2021b). For experimental and quasi-experimental designs, both aim at estimating the causal effects of exposure to different kinds of counterspeech on outcome variables in comparison with controls (no exposure to counterspeech). Research design Languages and countries In the reviewed work, the impact of counterspeech is investigated in five different languages across nine countries. Notably, experiments are focused on counterspeech used in Indo-European languages such as English (USA, UK, Canada and Ireland), German (Germany), Urdu (Pakistan) and Swedish (Sweden). Only two studies are dedicated to Afro-Asiatic languages, Arabic (Egypt and Iraq). We did not find research dedicated to other language families, suggesting that the language coverage of counterspeech studies is still low. Languages and countries Platforms Most experiments were conducted on text-based social media platforms, such as eight on Twitter (Benesch et al., 2016; Reynolds and Tuck, 2016; Silverman et al., 2016; Stroud and Cox, 2018; Munger, 2017; Hangartner et al., 2021; Poole et al., 2021; Garland et al., 2022), six on Facebook (Reynolds and Tuck, 2016; Silverman et al., 2016; Schieb and Preuss, 2016; Leonhard et al., 2018; Saltman et al., 2021; Buerger, 2021b), and one on Reddit (Bilewicz et al., 2021), as well as image-based online spaces, such as three on Youtube (Reynolds and Tuck, 2016; Silverman et al., 2016; Ernst et al., 2017) and one on Instagram (Stroud and Cox, 2018). Often, the counterspeech interventions are directly monitored on such platforms, but in some cases, fictitious platforms are created in order to mimic online social activity under a controlled environment (Obermaier et al., 2021; Carthy and Sarma, 2021; Bélanger et al., 2020). There are three studies analysing the impact of counterspeech across multiple platforms (Reynolds and Tuck, 2016; Silverman et al., 2016; Stroud and Cox, 2018). Platforms Twitter and Facebook are widely used for measuring the effects of counterspeech, with eight and six experiments respectively. For Twitter, this can be explained by its easily accessible API (even if at the time of writing continued research access to the API was in doubt). Similarly, because of difficulties in gathering data, Schieb and Preuss (2016) resort to developing an agent-based computational model for simulating hate mitigation with counterspeech on Facebook. It is worth highlighting that none of the studies we reviewed had investigated recently popular mainstream platforms, such as Tiktok, Weibo, Telegram, and Discord. The target of hate speech Abusive speech can be addressed towards many different potential targets, and each individual hate phenomenon may require different response strategies for maximum effectiveness. Existing studies have evaluated the effectiveness of counterspeech on several hate phenomena, with Islamophobia, Islamic extremism, and racism being the most commonly addressed, while hate against LGBTQ+ community and immigrants being the least studied. In these studies, abusive content is typically identified based on two strategies - hateful keyword matches (Hangartner et al., 2021; Bilewicz et al., 2021), or user accounts (e.g., content produced by known hate speakers) (Garland et al., 2022). The target of hate speech Types of interventions A wide range of methods are exploited to design and surface counterspeech messages to a target audience. We broadly categorise these methods based on modality and approach to creation. Counter speech is generally conveyed in text (Bélanger et al., 2020; Hangartner et al., 2021; Poole et al., 2021) or video mode (Ernst et al., 2017; Saltman et al., 2021; Carthy and Sarma, 2021). In both cases, counterspeech materials can be created in three different ways: written by experimenters as stimuli (Obermaier et al., 2021; Carthy and Sarma, 2021), as well as written by individuals or campaigns that are collected from social media platforms (Benesch et al., 2016; Garland et al., 2022; Buerger, 2021b). We also found one study integrating counterspeech messages in media such as films, TV dramas and movies (Iqbal et al., 2019). Types of interventions Counterspeech strategies Following the strategies summarised in Section 4.1, commonly used counterspeech strategies include facts (Buerger, 2021b; Obermaier et al., 2021), denouncing (Stroud and Cox, 2018; Saltman et al., 2021), counter-questions (Silverman et al., 2016; Reynolds and Tuck, 2016; Saltman et al., 2021), and a specific tone (humour or empathy) (Reynolds and Tuck, 2016; Munger, 2017; Hangartner et al., 2021; Saltman et al., 2021). There are more fine-grained tactics for designing counterspeech in social science experiments. According to psychological studies, the use of social norms can reduce aggression and is closely related to legal regulation in society (Bilewicz et al., 2021). This tactic was tested in an intervention study where participants were exposed to counterspeech with one of the inducements of empathy, descriptive norms (e.g., Let’s try to express our points without hurtful language) and prescriptive norms (e.g., Hey, this discussion could be more enjoyable for all if we would treat each other with respect.) (Bilewicz et al., 2021). Bélanger et al. (2020) designed counterspeech based on substances rather than tactics, varying three different narratives: (1) social (seeking to establish a better society), (2) political (bringing a new world order through a global caliphate), and (3) religious (legitimising violence based on religious purposes). Considering broader counterspeech components, a few organisations further focus on challenging ideology (e.g., far-right and Islamist extremist recruitment narratives), rather than deradicalising individuals (Silverman et al., 2016; Saltman et al., 2021). Counterspeech drawing from personal stories in a reflective or sentimental tone is also considered as it can resonate better with target audiences (Silverman et al., 2016). In addition to neutral or positive counterspeech, radical approaches are taken by counter-objecting, degrading or shaming perpetrators in public for unsolicited harmful content (Stroud and Cox, 2018; Obermaier et al., 2021). Counterspeech strategies Types of evaluation metrics Based on Reynolds and Tuck (2016)’s counterspeech Handbook, we identified the following three types of metrics used by the authors of the papers to evaluate the effectiveness of counterspeech interventions: social impact, behavioural change, and attitude change measures. Types of evaluation metrics • Social impact metrics are (usually automated) measurements of how subjects interact with counterspeech online. Such measures include, bounce rate, exit rate,[4] geo-location analysis and the numbers of likes, views, and shares that posts receive (Garland et al., 2020; Hangartner et al., 2021; Poole et al., 2021; Reynolds and Tuck, 2016; Leonhard et al., 2018; Saltman et al., 2021; Silverman et al., 2016). For example, for one of their experiments, Saltman et al. (2021) measure the ‘click-through rates’ of Facebook users redirected from hateful to counterspeech materials, while Hangartner et al. (2021) measure retweets and deletions (in addition to behavioural change measures). • Social impact metrics Social impact measures are also applied to synthetic data by Schieb and Preuss (2016), who measure the ‘likes’ of their (simulated) participants as hate and counterspeech propagate through a network (as well as applying behavioural metrics). Taking a more distant, long-term view, Iqbal et al. (2019) cite Egypt’s overall success at countering radicalisation with counterspeech campaigns by comparing its position on the Global Terrorism Index with that of Pakistan. While the majority of these measurements are automated, Leonhard et al. (2018) use survey questions to examine participants willingness to intervene against hate speech depending on the severity of the hate, the number of bystanders, and the reactions of others. Unlike the survey-based approaches described below, they do not consider changes in attitude. In addition, Buerger (2021b) assess the success of the #jagärhär counterspeech campaign (#iamhere in English, a Sweden-based collective effort that has been applied in more than 16 countries) based on the extent to which it has facilitated the emergence of alternative perspectives. • Behavioural change measures reveal whether subjects change their observable behaviour towards victims before and after exposure to counterspeech, for example in the tone of their language as measured with sentiment analysis. • Behavioural change measures For instance, Hangartner et al. (2021) conduct sentiment analysis to determine the behaviour of previously xenophobic accounts after treatment with counterspeech, Bilewicz et al. (2021) measure levels of verbal aggression before and after interventions, and Garland et al. (2020) assess the proportion of hate speech in online discourse before and after the intervention of an organised counterspeech group. Other such measures are those of Saltman et al. (2021), who compare the number of times users violate Facebook policies before and after exposure to counterspeech, and Munger (2017), who examine the likelihood of Twitter users continuing to use racial slurs following sanctions by counterspeakers of varying status and demographics. And in a network simulation experiment, Schieb and Preuss (2016) measure the effect of positive or negative (synthetic) posts on (synthetic) user behaviour. • Attitude change measures are used to assess whether people (hate/counter speakers or bystanders) change their underlying attitudes or intentions through non-automated methods such as interviews, surveys, focus groups, or qualitative content analysis. • Attitude change measures For potential hate speech perpetrators, Carthy and Sarma (2021) use psychological testing to measure the extent to which participants legitimized violence after exposure to differing counterspeech strategies, Bélanger et al. (2020) compare support for ISIS and other factors using in participants exposed to differing counterspeech strategies and a control group, Ernst et al. (2017) code user comments on hate and counterspeech videos to perform qualitative content analysis of users’ attitudes. For bystanders that may be potential counterspeakers, Obermaier et al. (2021) use a survey to examine whether counterspeech leads to increased intentions to intervene. And for those already engaged in counterspeech, Buerger (2021b) conduct interviews with members of an organised group to reveal their perceptions of the efficacy of their interventions. Effectiveness Owing to the variation in experimental setups, aims, and evaluation methods of the counterspeech efforts we review, it is not straightforward to compare their levels of success. Indeed, several of the studies concern broad long-term goals that cannot be easily evaluated at all (e.g. Reynolds and Tuck, 2016; Silverman et al., 2016) or provide only anecdotal evidence (e.g. Benesch et al., 2016; Stroud and Cox, 2018; Buerger, 2021b). Effectiveness Beyond this, evidence of successful counterspeech forms a complex picture. For example, Garland et al. (2022) show that organised counterspeech is effective, but can produce backfire effects and actually attract more hate speech in some circumstances. They also show that these dynamics can alter surrounding societal events—although they do not make causal claims for this. Similarly, Ernst et al. (2017) find mixed results, with counterspeech encouraging discussion about hate phenomena and targets in some cases, but also leading to increases in hateful comments. However, Silverman et al. (2016) suggest that even such confrontational exchanges can be viewed as positive signs of engagement. There is some evidence for the comparative efficacy of different counterspeech strategies. Bilewicz et al. (2021) find that three of their intervention types (‘disapproval’, ‘abstract norm’, ‘empathy’) are effective in reducing verbal violence when compared with no intervention at all. Here, empathy had the weakest effect, which they put down to the empathetic messages being specific to particular behaviours, limiting their capacity to modify aggression towards wider targets. Hangartner et al. (2021) also found that empathy-based counterspeech can consistently reduce hate speech, although this effect is small. And Carthy and Sarma (2021) found that counterspeech that seeks to correct false information in the hate speech actually leads to higher levels of violence legitimisation, while having participants actively counter terrorist rhetoric themselves (‘Tailored Counter-Narrative’) was the most effective strategy to reduce this. They found counterspeech to be more effective on participants that are already predisposed to cognitive reflection. However, focusing on the effect of factual correction on the victims rather than perpetrators of hate speech, Obermaier et al. (2021) found it to be effective in providing support and preventing them from hating back and therefore widening the gap between groups. There is also some evidence that the numbers of the different actors involved in a counterspeech exchange can affect an intervention’s success. Schieb and Preuss (2016) find that counterspeech can impact the online behaviour of (simulated) bystanders, with the effectiveness strongly influenced by the proportions of hate and counter speakers and neutral bystanders. According to their model, a small number of counterspeakers can be effective against smaller numbers of hate speakers in the presence of larger numbers of people lacking strong opinions. Saltman et al. (2021) found their counterspeech strategies to be effective only for higher risk individuals within the target populations, although they did not see any of the potential negative effects of counterspeech (such as increased radicalisation) reported elsewhere. Focusing on who in particular delivers counterspeech, Munger (2017) finds that success of counterspeech depends on the identity and status of the speaker. However, with only a small positive effect, Bélanger et al. (2020) found that the content of counterspeech was more important than the source. And Garland et al. (2022) found that, while organised counterspeech can be effective, the efforts of individuals can lead to increases in hate speech. In Buerger (2021b), members of #jagärhär claim that their counterspeech interventions were successful in making space for alternative viewpoints to hate speech. This paper is available on arxiv under CC BY-SA 4.0 DEED license. This paper is available on arxiv under CC BY-SA 4.0 DEED license. available on arxiv [4] Bounce rate is the number of users who leave a website without clicking past the landing page; exit rate measures how many people leave the site from a given section (Reynolds and Tuck, 2016).
