The talk deals with options to enhance the resilience of democratic societies related to the threats of disinformation. It discusses the results of a 2-fold analysis of experts in this particular field from science and political thinktanks as well as from journalism, activism, and law enforcement. We applied in-depth interviews and a focus group with these experts based on core findings of previous research results to identify patterns of successful strategies to fight disinformation. It is argued that it would be shortsighted just to flag disinformation or to roll out prebunking measures at scale. On the long run, democratic institutions should instead establish constructive narratives how to tackle societal problems.

Misinformation is a pressing societal concern, evident in contexts such as the COVID-19 pandemic, the 2020 U.S. elections, debates over climate change, and immigration. The contemporary information environment has amplified the spread of misinformation, yet much of the academic and public discourse has centered on the role of fake news and untrustworthy fringe sources. However, such sources are rarely visited by most users, while misinformed beliefs remain widespread.
In this presentation, we argue that partisan media—particularly widely visited partisan outlets—may play a more significant role in shaping false beliefs. These effects can be both direct and indirect, for instance, by linking audiences to fringe sources. We explore this argument using a combination of web-tracking and survey data from five advanced democracies (France, Germany, Spain, the United States, and the United Kingdom).
Our findings indicate that highly skewed partisan media diets—particularly those leaning to the right—are strongly associated with false beliefs about the most politically polarized issues, such as climate change and immigration. In contrast, exposure to untrustworthy fringe sources is more directly linked to false beliefs about less politicized topics, including COVID-19 and the war in Ukraine. The results underscore the need for misinformation interventions that are sensitive to the varying levels of political polarization across issues.

In an increasingly polarized and information-saturated digital landscape, fostering constructive and democratic discourse requires both critical competencies and innovative technological interventions. Civic information literacy refers to the ability to navigate digital information environments critically, efficiently, and responsibly. We present a set of fast, transparent, and non-partisan tools designed to help individuals distinguish credible from manipulative content. Core techniques include lateral reading, reflective engagement with emotional cues, and informed use of collaborative platforms like Wikipedia. At the same time, digital environments increasingly rely on algorithmic systems to moderate discourse. It is well established that computers are perceived as social actors, a tendency that becomes even more pronounced with large language model-driven virtual agents. These agents are entrusted with increasingly complex social tasks that demand emotional sensitivity and moral reasoning. One such task is counterspeech in response to hate speech and incivility on social media. We present findings on various counterspeech techniques, assess their effectiveness, and explore their potential for automation through bots. The findings suggest how LLM-based bots could play a supportive role in promoting healthier and more respectful online discourse.

Over recent decades, Americans have increasingly clustered into like-minded partisan communities and grown more hostile toward the opposing party, resulting in unprecedented highs in affective polarization (Brown et al., 2025; Boxell et al., 2024). Prior research shows that same-party conversations can intensify polarization (Strandberg et al., 2017), while cross-party exchanges can mitigate it (Levendusky et al., 2021), yet the effect of agreement and disagreement within these dialogues remains unclear. Drawing on social-balance theory (Newcomb, 1953) and cognitive-dissonance theory (Festinger, 1957), we argue that ideological dissonance – disagreeing with an in-group member or agreeing with an out-group member – triggers discomfort that makes people rethink group boundaries. This process might push participants to reevaluate the diversity within their own party or reduce perceptions of out-group extremism, thereby lowering affective polarization. We test this claim in a 2 × 2 online experiment. Participants engage in brief, structured chats with a large language model that either represents their political in-group or out-group and either agrees or disagrees with them, yielding four conditions: (1) In-group Agreement, (2) In-group Disagreement, (3) Out-group Agreement, and (4) Out-group Disagreement. For example, a self-identified Democrat in the Out-group Agreement condition converses with a Republican-aligned chatbot that agrees with the participant’s policy view. A follow-up survey four weeks later will measure the persistence of any attitude change. By demonstrating how short AI-mediated dialogues can recalibrate partisan perceptions, this study will provide both theoretical insight and a scalable intervention for fostering more civil political discourse.

In the absence of hard empirical evidence for the existence of genuine echo chambers, our focus must necessarily shift to polarisation as a major driver of societal divisions: the core problem in the contemporary landscape of public debate is not that people no longer encounter counter-attitudinal information and perspectives, but that perceived or actual polarisation between political and social groups undermines meaningful debate and consensus-building between opposing groups. Even the apparently straightforward concept of polarisation must be further developed, however: clear distinctions between competing issue and ideology positions can be productive if they enable citizens and decision-makers to choose their preferred course of action, but such agonistic competition can turn into antagonistic division once one or more sides of a debate abandon their commitment to engaging with their opponents, and to working towards consensus or at least compromise. Our work has identified several symptoms of this shift towards destructive polarisation: (a) breakdown of communication; (b) discrediting and dismissing of information; (c) erasure of complexities; (d) exacerbated attention to and space for extreme voices; and (e) exclusion through emotions. Drawing on social media data, this contribution uses the novel methodological approach of practice mapping to identify the discursive groups and their positions in a given public debate, and to assess the extent to which their activities exhibit the symptoms of destructive polarisation.

The presentation reviews current research on artificial intelligence (AI) in the context of online discussions, using deliberative norms as a theoretical framework. Focusing on the deliberative dimensions of civility, equality, rationality, and reciprocity, the study examines and categorizes 171 articles accordingly. It finds that most AI research to date emphasizes enhancing rationality and civility in online discourse. Techniques such as argument mining and hate speech detection are commonly employed to improve the quality of discussions, fostering more structured and respectful communication. However, efforts to address equality and reciprocity—the foundational principles of democratic discourse—remain limited. This imbalance highlights critical gaps in the application of AI for fostering equitable and inclusive dialogue.

Despite interaction being a key concept in most social theories, empirical research associated with the terms deliberation, echo chambers, and polarisation operates with a limited notion of interaction. Typically, the focus is on the production and diffusion of information or on interaction as a one-time perception and response. However, social media platforms also enable users to engage in continued reciprocal interaction or conversations of variable length. This shift in perspective requires means of data collection and analysis that maintain the structure of conversations in the form of reply chains that are part of larger reply trees. The consideration of online conversations allows us to ask new questions or rephrase existing ones, such as whether users actually engage in discussions online and what role partisanship and incivility play in them.

In our talk, we will address a fundamental issue in NLP-supported deliberation: evaluation. Practical and ethical implications often prevent direct testing in deliberative campaigns, which is why annotation carried out beforehand by domain experts (i.e., trained moderators) or recruited participants (who could, for example, take the perspective of the discussion participants) is a practical solution to the evaluation bottleneck. Recent developments in annotation for NLP, however, highlight a crucial property that such annotations should have: going beyond the notion of gold truth and instead focusing on as many perspectives as possible. This variety challenges the current evaluation simulations, as we will detail and propose solution approaches.

Given the prevalence of unhealthy, unconstructive, and oftentimes simply disrespectful behaviour in the online sphere, there continues to be a substantial amount of research focused on the matter of intervention into social media conversations. In our talk, we present a bot for online moderation on social media, built upon principles of soft moderation that is aimed at promotion of constructive dialogue rather than removal of offensive material. We will discuss components of the bot, including data collection, feature extraction, the decision intervention model, and content moderation generation. We also provide a human-centered perspective on content moderation based on the results of an ethnographic study conducted in the Netherlands.

Bots on social media have a bad reputation. We usually associate bots with threats to political discourse like misinformation. But why should bots not be used to do good? Specifically, could bots help to build bridges as they increase the quality of deliberation in a polarized debate? We investigate this question based on bot interventions trying to engage users with content that challenges their viewpoints. In the context of the German “Tempolimit” debate on the adoption of a general speed limit, we conduct a field study with a bot that replies to posts which oppose or support the general speed limit. Comparing to generic bot replies, preliminary results show that a bot using authoritative content to challenge users’ viewpoints can encourage substantial replies and longer chains of replies, a prerequisite for deliberation. We close with an outlook on variants of our bot that use humor rather than a logic of debate: In prior work, we developed bots that generate comic strips as an engaging and unthreatening medium to highlight different sides of an argument.