Public deliberation relies on a set of fragile cognitive and social conditions: access to shared information, mutual recognition between interlocutors, and the possibility of engaging with opposing views in a constructive manner. Digital platforms – and increasingly AI-based systems that curate, generate, and recommend content – have profoundly reshaped these conditions. While these technologies have expanded access to information and lowered barriers to participation, they have also transformed the dynamics of attention, visibility, and interaction that structure public debate.
This section examines how AI-driven content curation and recommendation systems affect the quality of public deliberation. Rather than focusing exclusively on misinformation or individual belief change, it analyses how emotionally charged content, polarisation dynamics, and choice architectures interact to shape communicative norms, mutual trust, and the capacity for collective reasoning. The goal is to identify not only the risks associated with current digital environments, but also the conditions under which AI-based technologies could support more reflective, inclusive, and constructive forms of public discourse.
Moral framing and negative content
We are wired to favour socially congruent and emotionally relevant information. Evolution has shaped our minds to prioritise loyalty to our groups and develop certain biases favouring our community, which are likely present across all groups. Modern politics vividly demonstrates how coalitions can provoke strong biases. Algorithms understand this well, exploiting our cognitive tendencies to promote the most divisive content (Clark et al., 2019). They contribute to distorting the logic of informational exchanges on the Internet by encouraging the polarisation of exchanges, because the aim of social media platforms is to increase user engagement.
The truthfulness of information is less relevant when the paramount aspect lies in the ability of the induced emotion to be justified and shared among individuals (Schaffner & Luks, 2018). Substantial experimental evidence corroborates the widespread phenomenon of emotional transmission across social networks (Kramer et al., 2014; Martel et al., 2024). Emotions serve as influential drivers for dissemination within social media spaces, particularly when expressing negative sentiments or moral emotions (Brady et al., 2017).
Moreover, attitudinal disagreements are often coupled with mistrust and contempt (reflected by the notion of affective – as opposed to attitudinal – polarisation) and these can be fed by moral and emotional content (Van Bavel & Pereira, 2018). Groups with different interests within polarities can also converge and fuel a movement of global polarisation (Benkler et al., 2018).
If exchanges on social networks often lack politeness, it is also because of the possibility of exchanging with anonymous, depersonalised strangers. This experience, unique to the Internet age (Bor & Petersen, 2022), reduces our sense of personal responsibility and empathy towards interlocutors whom we no longer see as individuals but as interchangeable members of political 'tribes' (Combs et al., 2023).
Mathematician David Chavalarias (2023) points to an insidious danger directly linked to attentional capture, namely the tendency of networks to exacerbate the small differences and micro-conflicts. Positions are not necessarily more radical, but hostility between different positions is said to have increased. Discourses pointing to societal threats enjoy a considerable competitive advantage in the market of ideas and conversations, because of their appeal to our minds (Acerbi, 2019a; Blaine & Boyer, 2018; Boyer & Parren, 2015).
Recent analyses also point out that social networks – and newspapers as well – function less as a mirror than as a distorting prism for the diversity of opinions in society. Social networks seem to contribute to increasing political hostility and radicalism because people are exposed to caricatural and aggressive versions of opposing political positions, which are a source of irritation. The majority of users, who are more moderate, are reluctant to engage in political discussions that rarely reward good faith. In fact, outraged and potentially offending political posts are often made by people who are more determined to express their views and who are more radical than the average – whether they are signalling their commitment, expressing their anger or promoting their views. Even though they represent a relatively small proportion of the written output on the networks, these posts are promoted by algorithms programmed to highlight content capable of attracting attention and triggering responses, of which divisive messages are a good example (Chavalarias, 2023).
Most of our virtual connections do not develop into 'echo chamber’, isolating us within entirely homogeneous pools of political ideas. The diversity of opinions with which we are confronted online is frequently greater than that of our friendships. This exposure to ideological alterity is desirable, in theory, as it should enable us to discover the blind spots in our political knowledge and convictions (Bakshy et al., 2015; De Francisci Morales et al., 2021; Guess et al., 2018), our common humanity, and therefore make us both more humble and more respectful of each other. Unfortunately, the way in which most people express their political convictions rather lacks both nuance and pedagogy. It tends to reduce opposing positions to caricatures and seeks less to persuade the other side than to rally those who already agree with you, or to make political friends (Grubbs et al., 2019).
A series of experimental studies carried out on Twitter, as well as interviews with Democratic and Republican activists by sociologist Chris Bail and his colleagues (2018) show that repeated exposure to weak and mocking content produced by political enemies can paradoxically reinforce supporters in their pre-existing positions and identities (Törnberg, 2022), rather than bringing them intellectually and emotionally closer to each other.
Defensive dynamics and opinion polarisation
Polarised environments do not merely reflect pre-existing disagreements; they actively reshape how individuals express, defend, and even form their positions. A central issue with polarisation is that it places individuals and groups in a defensive posture, encouraging them to adopt stronger, more rigid positions than they would endorse in less antagonistic contexts. When people feel that they, or the social groups they identify with, are under threat, reasoning becomes oriented toward justification and identity protection rather than mutual understanding or truth-seeking.
A large body of research in social and political psychology shows that perceived intergroup threat intensifies motivated reasoning, reduces openness to counterarguments, and increases hostility toward out-groups (Kunda, 1990; Taber & Lodge, 2006; Van Bavel & Pereira, 2018). In online environments, these effects are amplified by public exposure, reputational concerns, and the performative nature of political expression. Users are often not merely expressing personal beliefs, but signalling loyalty, moral commitment, or group membership to an imagined audience. As a result, positions become more extreme, less nuanced, and less responsive to evidence.
This dynamic is particularly detrimental to public deliberation, which presupposes that participants are willing to revise their views, acknowledge uncertainty, and engage with disagreement in good faith. In polarised settings, disagreement is more likely to be interpreted as hostility or bad faith, making constructive dialogue increasingly difficult. Even moderate individuals may self-censor or disengage from political discussions, perceiving them as emotionally costly or futile, thereby leaving the public space dominated by more extreme and combative voices.
Algorithmic recommendation systems can exacerbate these tendencies by amplifying content that triggers outrage, fear, or moral condemnation, as such content reliably generates engagement. Over time, this creates a feedback loop in which defensive postures, emotional escalation, and polarised narratives become the dominant modes of interaction. The result is not necessarily greater ideological extremism, but a hardening of social boundaries and a degradation of the communicative norms required for democratic deliberation.
Understanding the intricate and often opaque nature of interactions between users and the algorithms that curate content on social media and other online platforms is crucial but challenging due to limited transparency and the dynamic nature of the algorithms (Lewandowsky et al., 2024). How social media algorithms work remains very opaque, but timelines will certainly favour content that has been the subject of higher engagement. An obvious consequence is that posts prompting for an immediate reaction will climb to the top of the newsfeed.
Recent research conducted by Rathje, Van Bavel & van der Linden (2021), analysing 2.7 million posts on Facebook and Twitter, underscores the strong predictive nature of content targeting out-groups in driving engagement across social media platforms. The quality of public debate thus seems to be threatened by increased polarisation: recommendation engines, with the aim of keeping users for as long as possible, come up with increasingly polarising content. Such dynamics can even lead to violent behaviours, as shown by the role of social media in organising the storming of the U.S. Capitol on January 6, 2021.
A model proposed by Santos, Lelkes and Levin (2021) shows how the strength of social influence affects opinion dynamics, by stressing the role of link recommendation algorithms commonly used in online social networks in the creation of ties and opinion polarisation. Algorithms tend to suggest new connections between users with a high number of common acquaintances, which can promote structural similarity and the formation of isolated, like-minded communities. In this context, even moderate opinions can contribute significantly to opinion polarisation. Other empirical data suggest that link recommendation algorithms do indeed change the rewiring pattern of social networks.
Empowering and making individuals more responsible
Empowerment is not only about regulating information flows or curating content: it also concerns how individuals engage with the information they encounter. In everyday life, people’s reasoning is influenced by powerful motivational factors that determine whether and how they mobilize their cognitive resources. For reasoning to be oriented toward truth, the argumentative context must be collaborative, and individuals must be willing to step outside their own perspective. The salience of truth versus belief maintenance depends on context: the more that knowing the truth serves a personal or social purpose, the more likely individuals are to invest cognitive effort in deliberation. This contextual component is perhaps the most important factor in tackling the challenge of false information on the Internet. A central aspect of citizenship is the possibility of authentic dialogue.
Yet, research suggests that individuals are often less polarized in their underlying opinions than online debates might suggest (Treuillier et al., 2024; Oswald, Schulz & Lorenz-Spreen, 2025). What intensifies conflict is not the extremity of one’s own position, but the perception that others – whether individuals or groups – are hostile, unyielding, or unwilling to engage constructively. Online interactions, especially when provoked or framed confrontationally, tend to put people on the defensive, amplifying reactive behaviours rather than genuine disagreement. The central challenge, therefore, lies in fostering mutual trust in the willingness and ability of the opposing side to engage sincerely with our arguments. Citizens do not need to adopt identical opinions; rather, they need to believe that their interlocutors are motivated by the same goal: to collectively find the best solution to shared problems. Building this perception of constructive intent is a crucial step toward empowering individuals to deliberate responsibly and to move discussions beyond reactive defensiveness.
Individual vs. societal consequences
The consequences of information are not the same at the individual and societal levels. Upholding distorted depictions on certain topics, especially those with significant political and social ramifications, can result in profound effects, impacting areas such as health and politics. The success of political endeavours relies on their adherence to a truthful representation of reality. The collective understanding of this representation holds significant consequences when influencing the outcomes of decisions made by society as a whole. Democracy indeed requires a common base of knowledge among citizens to function optimally, which includes trust in electoral processes and evidence to inform policy debates (Lewandowsky et al., 2023).
Research shows that widespread disinformation campaigns are eroding the shared knowledge that democracy depends on, thereby undermining democratic institutions and processes (Lewandowsky et al., 2023). Not because people believe in this false information, but because they have less trust in the benevolence of media institutions. Even when the epistemic commitment to misconceptions about the world is not very strong and may only be a means of justifying pre-existing intuitions, individual attitudes can be harmful when they become dominant in the public sphere. Once a narrative enters public discourse, it has the potential to evolve into a norm that shapes the perception of societal issues. Consequently, modes of thinking that prioritise coalition-building over collaboration and truth-seeking may dominate more moderated forms of discourse. Returning to a peaceful debate is then very difficult, as people are even less interested in hearing contradictory views. Positions defended for social or identity-related reasons can in turn lead to the undermining of the common ground on which decisions taken at societal level are based. If this common ground is not clearly identified and shared by all, the public debate becomes distracted from the real issues facing society.
This polarisation impedes collective decision-making, as citizens are less inclined to engage with opposing perspectives or to work together to achieve common goals. In terms of social attitudes, polarisation undermines the foundations of constructive social debate, fostering intolerance and hostility rather than open and respectful dialogue. As a result, it diminishes the ability to find consensus or compromise, essential elements of a functioning democratic process. If citizens are to remain genuine actors involved in collective decision-making, the issue of truthfulness in the exchange of information, including on the internet, should become an individual priority. Users must be given the means to protect themselves against polarising mindsets when navigating the internet.
How AI could foster public deliberation
Historically, technologies that enhance information transmission tend to benefit society more than they harm it. Artificial intelligence, particularly large language models (LLMs), presents a unique opportunity to improve both the quality of information and the conditions for constructive public deliberation. Rather than aiming to make citizens hold the same opinions, AI could help increase trust in the willingness and capacity of interlocutors to engage sincerely, even when disagreements persist.
One approach involves automatic fact-checking and reasoning support. LLMs could provide real-time annotations to social media posts, highlighting logical inconsistencies or contradictions with established facts, and suggesting alternative interpretations grounded in evidence. Platforms could incentivise the use of such verifications by boosting posts that adopt them, thereby improving the informational environment while leaving individuals free to form their own opinions. Similarly, AI tools could detect misleading or AI-generated accounts, reducing their influence and helping to maintain discourse integrity.
Beyond factual accuracy, AI can play a role in structuring deliberation and mitigating defensive responses. Bijnen & Greco (2018) demonstrate that making disagreements explicit and transparent can reduce defensiveness, allowing participants to focus on the reasoning behind others’ positions rather than perceiving them as threats. Digital tools such as the BCAUSE application, developed within the Horizon project ORBIS, implement this principle by guiding users to articulate arguments clearly, indicate points of agreement and disagreement, and structure debates so that all participants feel heard. By fostering the perception that opposing parties are sincerely trying to understand one’s arguments, these systems enhance the trust necessary for constructive engagement, even when fundamental disagreements remain.
AI could also support the creation of information bots that counteract misinformation bots. These bots can disseminate accurate information, engage politely in discussion, and model deliberative behaviour. When designed at scale, they could balance the online informational environment without forcing homogenization of opinions. Importantly, their goal is not to eliminate disagreement but to maintain a climate of deliberation, where individuals feel confident that others are participating in good faith.
Finally, AI-driven tools could democratize the production of engaging, evidence-based content. By lowering the cost of creating high-quality information, smaller organizations and local actors could compete with larger outlets, making the informational landscape more pluralistic. Even if some smaller actors occasionally misrepresent facts, the combination of AI-assisted moderation, structured debate, and transparency mechanisms would enable citizens to assess credibility and engage responsibly, enhancing both individual autonomy and collective decision-making.
The references mentioned in this report are available in the Bibliography of KT4D Social Risk Toolkit Module B: AI, trust and awareness. (https://zenodo.org/records/18375398)