ED Mathématiques et Informatique
Building, evaluating and understanding socio-cultural AI: leveraging concepts and methods from human sciences
by Grgur KOVAC (Institut national de recherche en informatique et en automatique - Bordeaux - Sud-Ouest)
The defense will take place at 17h00 - Amphi Edison 200 avenue de la Vieille Tour 33405 Talence
in front of the jury composed of
- Pierre-Yves OUDEYER - Directeur de recherche - INRIA, University of Bordeaux - Directeur de these
- Jan SNAJDER - Full professor - University of Zagreb - Rapporteur
- Maarten SAP - Assistant professor - Carnegie Mellon University - Rapporteur
- Clementine FOURRIER - Ingénieure de recherche - Hugging Face, Inc. - Examinateur
- Peter Ford DOMINEY - Directeur de recherche - CNRS, University of Burgundy - CoDirecteur de these
- Mehdi KHAMASSI - Directeur de recherche - CNRS, University of Sorbonne - Examinateur
- Vered SHWARTZ - Assistant professor - University of British Columbia - Examinateur
The recent evolution of artificial intelligence (AI) has led to its increasing integration into human culture. This growing presence of AI in human society raises important scientific questions - particularly those concerning the socio-cultural aspect of AI, which are crucial for better conceptualizing, evaluating, and designing future AI systems, as well as understanding their potential impact on the human society and culture. While this is a complex and nuanced topic, valuable insights can be drawn from psychology and human sciences, which have studied related aspects of socio-cultural behavior in humans and animals for decades. We explore three core scientific questions: What does an intelligent system need to enter a human culture? Drawing on developmental psychological theories by Michael Tomasello and Jerome Bruner, we outlined core socio-cognitive abilities that we believe are most relevant for current AI research. We introduced The SocialAI School - a tool for generating environments designed to support evaluation and development of these abilities in artificial agents. We demonstrated various usages of the SocialAI School with RL and LLM-based agents. Our experiments revealed limitations of standard RL agents, particularly in their inability to generalize to new contexts, and showed that while LLMs exhibit behavior somewhat consistent with correctly inferring social cues, a performance gap still remains. How can we characterize a culture encoded within an artificial system, such as a large language model (LLM)? We demonstrate that LLMs exhibit strong sensitivity to seemingly trivial context changes, which challenges the assumptions underlying many psychological questionnaires increasingly used to assess LLMs. We caution against naively using such questionnaires to draw general conclusions about LLM behavior. To investigate this, we systematically compare LLMs based their sensitivity to trivial context changes, i.e. on the stability of values expressed by simulated personas over various contexts induced by simulated conversations on different topics. We observe that some model families - Qwen, Mixtral, Mistral, GPT-3.5 - consistently exhibited higher stability in various experimental setups. Then, we construct a leaderboard by further extending this methodology. The latest results suggest that, while rank-order stability may be approaching its ceiling in this suite, a persistent gap in validation scores points to either remaining room for improvement or a fundamental limitation in applying human-centric theories to LLMs. How does a human–AI culture change and evolve over time? Using the iterative chain design (a method adapted from the field of cultural evolution) we explored how different properties of human data influence the evolution of AI-generated content. We found that higher lexical diversity and greater gaussianity in human data were associated with increased deterioration over generations, while higher semantic diversity and overall data quality with smaller deterioration. We also observe that data properties from one domain had little influence on the data generated for another domain. These findings suggest that different parts of the internet may exhibit distinct evolutionary dynamics, shaped by the properties of the underlying human data. The three questions discussed in this thesis are inherently complex and interdisciplinary. We presented only initial steps toward addressing them, leaving many open questions about the socio-cognitive capacities, internal representations, and cultural impact of AI systems. These questions are of utmost importance, especially as such systems are increasingly influencing human society and culture.