AI in the Public Sphere: Sciences Po hosts a 2-day conference

As a world-class university at the forefront of research and sustained engagement with both public and private actors, Sciences Po has uniquely positioned itself to analyse and challenge the forces shaping our contemporary world. Its international research network, stretching from academics to practitioners, has recently unsurprisingly turned to the field that is increasingly touching most disciplines: AI.

In this vein, the Open Institute for Digital Transformations, created as part of TIERED*, and the European Polarisation Observatory, financed by CIVICA, co-hosted the first edition of a two-day conference exploring the implications of artificial intelligence for democracy and the public sphere (May 18-19, 2026). 

Ranging from the effects of social media platforms on polarisation and misinformation to the implicit biases and power imbalances associated with AI, the research presented ultimately raised the question of what methods can be used to accurately measure and evaluate the repercussions of new technologies, and how they may consequently be effectively regulated.

Day One: Platforms, Polarisation, and Power

The first morning of the event revolved around the social, linguistic and political dynamics of social media platforms. The conference featured researchers affiliated with the European Polarisation Observatory, a CIVICA research hub: Korhan Koçak (IE University, Spain), Marton Karsai (Central European University, Austria), Elisa Omodei (Central European University, Austria), and Drew Dimmery (Hertie School, Germany). Together, they revealed the complex ways through which platforms influence their users by exploring how individuals cognitively process information online, how linguistic patterns may reveal patterns of social inequalities and marginalisation, and how online environments may foster polarisation and misinformation.

In this context, the need for methodological renewal through more iterative and experimentally-geared approaches in social science research was also reinforced. Speakers argued that critically evaluating datasets and diversifying the forms of data used may evidence more nuanced processes of belief formation, polarisation, and online behaviour.

Understanding social media platforms' socio-political impacts also calls for increasing scrutiny of the growing power and influence of search engines, social media companies, and AI-powered tools. This challenge was the focus of the first day's evening conference, "Interrogating Platform Power." The opening panel comprised both academics and industry professionals: Henri Verdier, CEO of the INRIA Foundation, Anastasia Stasenko from Pleiais, Antonio Calleja-López from Decidim, and Andrew Perrin from Johns Hopkins University. Blending their respective academic and industry experiences enabled them to discuss the theoretical and practical limits of platform sovereignty, especially in the context of the AI boom.

Jen Schradie, moderator and expert in digital sociology at Sciences Po's CRIS, recalled that the emergence of such technologies brings both the promise of liberalisation and further democratisation, as well as the threat of an increasing concentration of power within technology firms in ways that could affect individual freedom. Such risks raise questions about public services' reliance on privately owned technologies, especially as they become increasingly integrated into everyday life. Henri Verdier articulated this concern directly, stating: "Maybe we need a space that is state-owned, or at least state-backed, where you know that your data isn't used to trick you. Maybe we need to fight stronger to take back and impose something that is public."

This was followed by two student presentations. The first, by Julie Boury and Lucien Chaudron (School of Public Affairs), titled "@Grok, What Do You Think?," explored how Grok, an AI-powered chatbot, reshapes the dynamics of online deliberation. The second, by Anna Bogutska and Asger Grimberg (Paris School of International Affairs), titled "Investigating with AI: The Digital Representation of the Ukraine-Russia Killzone," presented a comparative longitudinal analysis of visual content across ten Telegram channels over four years of conflict, revealing how a shared visual language may convey opposite messages.

Ethan Zuckerman, professor at the University of Massachusetts-Amherst, delivered a final keynote address, shifting the focus from regulatory frameworks to the biases encoded within AI systems themselves. As he expressed, large language models (LLMs) embed a structural bias in their algorithmic construction—an alarming observation in the context of their globalised use. Put simply, these models are trained on digitised knowledge and writing that skews Western and Anglophone, and the values they encode reflect that cultural bias. This dynamic has been described as "digital erasure": the structural marginalisation of Global South voices and knowledge systems within AI-generated outputs. The result is that existing hegemonic cultural hierarchies are not merely replicated but actively reinforced at scale.

Yet, identifying this structural flaw also opens up a possibility to combat hegemonic bias by integrating alternative voices. In Ethan Zuckerman's words: "As we expand Large Language Models to include more knowledge, include more languages, we are expanding our cognitive toolkit." Pointing to the incommensurable investment in the industry, estimated at 2.5 trillion dollars this year alone, he added: "What I would really love to see us do is take 1% of that money and apply it towards making sure that those models are actually multilingual, multicultural, and are actually including an incredibly diverse range of voices."

Day Two: AI in the Public Sphere

The second day, "AI in the Public Sphere", examined how AI and politics mutually shape one another, structured around two complementary lines of inquiry. Featuring Sanne Kruikemeier (Wageningen University), Kevin Munger (European University Institute) and Janos Kertész (Central European University), the first session explored AI's impact on human political beliefs and behaviour. Research notably showcased how AI-generated political content, from TikTok edits to targeted ads, can influence voters’ attitudes. Understanding the interplay between AI and human behaviour also paves the way to the theorisation of 'co-evolution', whereby humans and algorithms influence each other in complex feedback loops, possibly demanding new frameworks for AI oversight and democratic governance. Those presentations were concluded by a roundtable discussion reuniting policy practitioners Carlo Santagiustina (Inria, Médialab), Vlad Bujdei-Tebeica (SNSPA), Paul-Antoine Chevalier (Tech-Ops Bureau, Viginum), to discuss platform governance and influence campaigns from a more regulatory perspective - complementing prior academic insights with actionable steps.

The second session, featuring Tanise Ceron (Bocconi University), Paul Bouchaud (CNRS-Sciences Po), Pedro Ramaciotti (CNRS-Sciences Po) and Stephan Lewandowsky (University of Bristol), focused on the political opinions encoded within AI itself, rather than those it may indirectly incite. LLMs are shown to reflect consistent political biases across languages, traceable to the composition of their training data, underscoring the need for greater transparency in data curation practices. Beyond the biases embedded in their outputs, LLMs also appear to encode structured representations of individuals themselves, retaining information on political leaning, nationality, and biographical attributes in ways that map directly onto legal criteria for identifiability under GDPR, with significant implications for how regulators should treat model weights. Reciprocally, algorithmic recommender systems have been found to encode users' political leanings in ways that blur the line between active and passive profiling, raising questions on data privacy regulation. These findings are further complicated by evidence that transparency mechanisms, such as disclosing AI-generated content, do not themselves undo AI's persuasive effects, pointing to the need for more robust regulatory frameworks that address AI's inherent capacity for manipulation. Joining researchers, Fabien Tarissan (CNIL, France), Erik Wetter (Stockholm School of Economics) and Artur Bogucki (Warsaw School of Economics) rounded off the discussion on a panel evaluating the possible mandates and practices that may be used to actively counteract and regulate those very algorithmic pitfalls. 

The conference ended on a broader reflection, led by Jon Cardoso-Silva (London School of Economics) and Kevin Munger (EUI) on AI's implications for education and academic publishing. Be it by disrupting student learning, shaping peer review processes, or accelerating submission rates beyond manageable thresholds, AI emerges here as a double-edged sword, whose consequences may not be fully understood, and certainly not fully mastered. As technologies progress further, and with them socio-political ideals and mechanisms, this conference stands as a valuable contribution to a conversation that is only just beginning. 

Article by Pia Tasso

*TIERED benefits from government funding administered by the National Research Agency (ANR) under the France 2030 program and by the European Union's NextGenerationEU program, under reference number “ANR-22-EXES-0014”.