1. médialab Sciences Po
  2. Actualités
  3. Social Science and Generative AI: Inquiries, Instruments, Consequences

Social Science and Generative AI: Inquiries, Instruments, Consequences

Cette conférence d'une journée examinera de manière réflexive l'IA générative en tant que dispositif sociotechnique, en discutant de ses effets sur le travail scientifique et de ses implications épistémologiques et sociales. Les interventions mettront un point d'honneur à considérer la diversité épistémologique des sciences sociales en accordant une attention égale aux approches quantitatives, qualitatives et participatives.

Rendez-vous, Conférence

⇥

Amphi Simone Veil, 28 rue des Saints-Pères, 75007 Paris

L'IA générative fait l'objet d'une attention croissante dans les sciences sociales, servant à la fois d'axe de recherche et d'instrument méthodologique. Ces technologies, qui progressent rapidement, permettent la création de données synthétiques, la mise en œuvre d'expériences à grande échelle et l'analyse d'entretiens et de récits ethnographiques. A l'inverse, les chercheurs utilisent ces technologies pour explorer leurs risques sociaux, leurs effets transformateurs et leurs implications politiques.

Ces pratiques émergentes soulèvent des questions épistémologiques et éthiques. Comment les nouveaux outils de recherche interagissent-ils avec les approches existantes de l'étude de la vie sociale ? Face à l'opacité des processus impliqués dans ces technologies, quelles options s'offrent aux chercheurs : reprendre le contrôle, expérimenter ou développer des outils alternatifs à partir de zéro ? Quels changements se produisent dans la relation entre la recherche sociale et les sociétés en ce qui concerne la position dans le travail de terrain, le rôle de la société civile et l'impact sur les institutions de gouvernance ?

Cette conférence explorera ces questions cruciales en réunissant d'éminents chercheurs internationaux qui s'intéressent à l'IA générative par le biais de la recherche empirique. Ils partageront leurs points de vue pour explorer les rôles interconnectés de l'IA générative dans le monde universitaire et la société, qui sont continuellement liés à son développement.

Programme

  • 8:30 : Accueil & Café
  • 9:00 - 9:15 : Préambule par Luis Vassy (Directeur de Sciences Po)
  • 9:15 - 9:30 : Introduction par Sylvain Parasie (Directeur du médialab)

Panel 1 : Données et Modèles

  • 9:30 - 10:20 :  “AI Errors and the Illusion of Artificial Life: Ethnographic Encounters with Generative AI” par Veronica Barassi (Professeure en études des médias et de la communication à l'Université de Saint-Gall)
  • 10:20 - 11:10 : “Preventing Language Models Weaponization: What if their Training Data were to be Compromised?” par Djamé Seddah (Chercheur à l'INRIA Paris’ Almanach)
  • 11:10 - 11:30 : Pause
  • 11:30 - 12:20 : “Machine Bias. How Do Generative Language Models Answer Opinion Polls?” par Etienne Ollion (Directeur de recherche au CNRS)

Panel 2 : Méthodes et Approches

  • 2:00 - 2:50 : “From Stories to Sonnets: Data-Centered NLP for Creative Works” par Maria Antoniak (Professeure adjointe en sciences informatiques à l'Université du Colorado Boulder)
  • 2:50 - 3:40 : “‘Conversing’ with Large Language Models: Enhancing Qualitative Data Using LLMs” par Adam Hayes (Professeur de sociologie à l'Université de Lucerne)

Panel 3 : Valeur Publique et Conséquences 

  • 4:00 - 4:50 : “Conditional Computing: Reimagining Europe’s Role in the AI Race” par Fabian Ferrari (Professeur assistant en IA culturelle à l'Université d'Utrecht)
  • 4:50 - 5:40 : “Language, Culture, Archive: Provocations from the Humanities for Generative AI Research” par Lauren Frederica Klein (Professeure associée aux départements de théorie et de méthodes quantitatives et d'anglais à l'Université Emory)
  • 5:40 - 6:00 : Conclusion

Résumés

  • Veronica Barassi : “AI Errors and the Illusion of Artificial Life: Ethnographic Encounters with Generative AI”

In this talk, I interrogate how ethnographic engagement with AI reveals the instability of machine "truths" and the social fictions we create around bot-generated knowledge. I ask: What does it mean when we believe we can learn from entities that cannot learn in the human sense? How do we make sense of what bots "tell" us, when their narratives are shaped by probabilistic association rather than contextual understanding? Bringing together reflection from The Human Error Project (2020-2024) with the ones gathered through a more recent ethnographic study on abusive digital bots and the growing terrain of Generative AI Violence, I discuss the role of gnerative AI as a destabilizing object for the social sciences and the reflect on the methods we can use to conduct meaningful ethnographic research.

  • Etienne Ollion : “Machine Bias. How Do Generative Language Models Answer Opinion Polls?”

Generative artificial intelligence (AI) is increasingly presented as a potential substitute for humans, including as research subjects. However, there is no scientific consensus on how closely these in silico clones can emulate survey respondents. While some defend the use of these “synthetic users,” others point toward social biases in the responses provided by large language models (LLMs). In this article, we demonstrate that these critics are right to be wary of using generative AI to emulate respondents, but probably not for the right reasons. Our results show (i) that to date, models cannot replace research subjects for opinion or attitudinal research; (ii) that they display a strong bias and a low variance on each topic; and (iii) that this bias randomly varies from one topic to the next. We label this pattern “machine bias,” a concept we define, and whose consequences for LLM-based research we further explore.

  • Maria Antoniak: “From Stories to Sonnets: Data-Centered NLP for Creative Works”

In this talk, I'll share two recent studies that use natural language processing (NLP) techniques, such as LLMs, to model creative works like stories and poetry. In the first part of the talk, I'll discuss NLP approaches for story detection and analysis, focusing on how NLP methods can help us study storytelling at large scales and across diverse contexts. In the second part, I'll discuss the poetic capabilities of large language models (LLMs), focusing on audits of the vast pretraining datasets used to build these models. Both studies will highlight the challenges in creating open evaluation datasets for creative works, the importance of interdisciplinary collaboration, and the new kinds of tools and resources required for such studies.

  • Adam Hayes : “Conversing’ with Large Language Models: Enhancing Qualitative Data Using LLMs”

Large Language Models (LLMs) are revolutionizing how qualitative researchers can work with textual data. Rather than relying only on codebooks or manual line-by-line analysis, scholars can now “converse” with their materials by asking targeted questions, probing for contextual insights, and refining theoretical connections. This dialogue-like process speeds up traditional tasks—transcription, coding, theme identification—while sparking broader possibilities for exploration. Researchers prompt the LLM to surface recurring patterns, detect subtle shifts in tone, or suggest new interpretive angles. The LLM becomes an active partner: it quickly identifies connections that might take days or weeks of manual work, yet the scholar remains responsible for selecting which prompts to use, verifying the outputs, and situating them within appropriate conceptual frameworks.

  • Fabian Ferrari : “Conditional Computing: Reimagining Europe’s Role in the AI Race”

The global AI race is often framed in terms of investment and compute power: a competition for deeper pockets and larger models. But the sheer scale of investments by the US and China far surpasses those of the EU, making it unrealistic to compete solely on monetary terms. This talk asks: What if the EU’s competitive advantage lies in values, not in budgets? It begins by outlining Europe’s spending disadvantage in AI. It then argues that focusing only on investments misses how Big Tech’s AI empires are also built on a philosophy of accelerationism and state-as-enterprise thinking. Finally, it proposes an alternative approach for the EU: conditional computing, a new governance framework that attaches public value conditions to AI infrastructure governance.

  • Lauren Frederica Klein : “Language, Culture, Archive: Provocations from the Humanities for Generative AI Research”

This talk derives from a working paper, "Provocations from the Humanities for Generative AI Research,” which presents a set of provocations for  considering the uses, impact, and harms of generative AI from the perspective of humanities researchers. In the paper, we provide a working definition of humanities research, summarize some of its most salient theories and methods, and apply these theories and methods to the current landscape of generative AI. We draw from foundational work in critical data studies, along with relevant humanities scholarship, in order to elaborate eight claims with broad applicability to current conversations about generative AI. In this talk, I will focus on three of these provocations: 1) Models make words, but people make meaning; 2) Generative AI requires an expanded definition of culture; and 3) Generative AI can never be representative; in order to show how core ideas from the humanities about language, culture, and the archive, can destabilize some of the current foundations of generative AI research and suggest alternative paths forward.

Informations pratiques

Date : 5 juin 2025

Heure : 8:30 - 18:00

Lieu : Amphithéâtre Simone Veil, 28 rue des Saints-Pères, 75007 Paris

Langue : anglais

Cette conférence est ouverte à toutes et à tous sur inscription.