1. médialab Sciences Po
  2. Activités
  3. NextGenAI

NextGenAI

Using LLMs to study the dynamics of democratic issue evolution and to enhance the quality of public debate.

Recherche

Research project

In 2025, Sciences Po joined NextGenAI, an international consortium bringing together leading universities and OpenAI.

Within the Open Institute for Digital Transformations, established as part of the TIERED initiative, two research projects have been launched with the support of NextGenAI, involving Jan Rovny (CEE), Raphaële Xenidis (EDD), Emiliano Grossman (CDSP), Kevin Arceneaux (CEVIPOF), and Béatrice Mazoyer (médialab). These projects explore the relationship between “AI and democracy” through two main areas of focus:

Dynamics of political issues

Using LLMs to study the dynamics of democratic issue evolution.

The study of political issues has been central to political science. How do political problems arise, how do they become contested, and how does this contest shape political supply and demand? These questions are even more pertinent today in the context of the rising salience of cultural contestation, engaging issues of identity and rights, and issues questioning liberal democratic institutions. While the development and dynamics of political issues have been studied for decades, previous works have primarily relied on the analysis of closed survey questionnaires or manual coding of political text (especially party manifestos). LLMs provide a unique opportunity to assess the dynamics of political issue evolution using both 1) a much larger corpus of text (parliamentary speeches, press releases, media articles, and reports), 2) a holistic analysis of all the words or concepts used in these political discourses over time and 3) the capacity to articulate the expression of issues to specific minorities across Europe where opinion may be ignored with other measurement tools. In this way, LLMs will be able to study all political issues raised in these forms of party communication over an extensive period. This study of issue evolution will provide invaluable insights into the dynamics of political competition in contemporary democracies (in Europe). Beyond that, it may help develop a better understanding of how political debates translate into policy decisions, focusing on the factors that will push specific issues into the legislative arena.

AI for augmented participation in public debate

Using LLMs to enhance the quality of public debate.

Functioning democracies require open and inclusive fora for citizens to share, debate, and form opinions. There is widespread doubt about whether social media can play this essential role in its current form. A wealth of studies has demonstrated that polarization tends to create tensions in online discussions: toxicity is on the rise, including pervasive discrimination, misinformation is spreading, and opinions tend to radicalize and oppose more brutally while most moderate voices are sidelined and marginalized groups silenced and excluded. LLMs can play a role in alleviating this vicious circle, but their contributions must be empirically tested. We know that large language models tend to reproduce the language, but also the political and moral perspectives of the data with which it was trained. This is the reason why it is so important to characterize the LLM’s inherent ideological – and potentially discriminatory – biases and systematically measure how these biases emerge when answering heterogeneous real-world prompts across a multitude of online systems. 

In theory, conversational bots promote increased participation from more diverse individuals engaged in facilitated discussions. We expect intelligent agents to encourage participation by summarizing lengthy debates and finding common grounds between diverging opinions, reformulating citizens' contributions when they do not meet certain standards or norms, or simply translating them to include contributors speaking different languages. Additionally, they may lessen toxicity by rephrasing posts in a more civilized way - while not silencing marginalized communities - increasing the quality of conversation even when ideologically opposed individuals are debating. Certain scholars have even imagined that conversational bots could participate actively in collective debates by simulating alternative perspectives. However, such applications call for solid and legitimate public scrutiny regarding the potential biases introduced by using LLMs. How faithful are language models to the opinions expressed by contributors? How can we measure this fidelity and guarantee that LLMs are not reducing the pluralism of the views when summarizing debates or adding an ideological slant when translating an argument? How to ensure that LLMs do not amplify already pervasive discriminatory prejudices and stereotypes?

Sciences Po has developed theoretical solid and practical experience in conceiving and organizing inclusive democratic debate at various scales and in different settings (small physical citizen assemblies, large online public consultations, etc.). This normative stance is essential to define which principles we want intelligent agents to abide by when in action in such sensitive spaces for democracy. More importantly, political scientists, sociologists and lawyers can work together to find ways to properly operationalize such democratic principles and turn them into actionable metrics (e.g. creating a dedicated benchmark).

We must also acknowledge that those intelligent machines cannot be studied without taking into account the way humans interact with them. We can assume that certain people distrust them, voluntarily ignoring their recommendations, or that others adopt specific strategies to keep control and use AI tools in a targeted way. A digital ethnography of how people interact with such systems is also key to understanding the entire socio-technical system and evaluating their potential to restore the virtues of democratic digital public spaces.

The médialab's mission

As part of this project, the médialab is conducting research on the relationship between artificial intelligence and democracy.

The mission involves designing and implementing projects in the humanities and social sciences that utilize generative AI models, particularly in the areas of modeling public policy issues and analyzing social media dynamics. It also includes a component dedicated to supporting researchers: providing computing resources (APIs, servers), training, technical support, and usage monitoring.