This project explores the transformative potential of Artificial Intelligence (AI) through an interdisciplinary approach that integrates computer science, network science, economics, and sociology. On one hand, we study the limitations and potential of AI to understand social, technological, and economic dynamics; on the other, we apply these models to conduct innovative investigations into complex issues such as the evolution of labor, the emergence of innovation trajectories, and the role of moral values in online debates. Activities include developing new continual and creative learning algorithms (Dreaming Learning, Lyapunov Learning), performing quantitative analyses of conceptual representations in language models, and employing LLMs to predict emerging trajectories of technological innovation. The goal is to foster a responsible adoption of AI capable of supporting strategic planning, sustainability, and collective progress.
Within the scope of this project, we propose to analyze the transformative potential of Artificial Intelligence (AI) through a markedly interdisciplinary approach that integrates Computer Science, Network and Complexity Science, economics, and sociology. The adopted perspective is twofold: first, we will investigate the potential and limitations of AI as an analytical tool for social, technological, and economic dynamics; second, we will utilize these methodologies to conduct innovative studies in these domains, moving beyond the predominant use of qualitative approaches in favor of a rigorous quantitative analysis, which is essential for addressing their intrinsic complexity.
Activities include developing new training methodologies for neural networks and conducting empirical studies on economically and socially relevant issues, such as the emergence of innovation trajectories (through patent analysis), the dynamics of AI substitution and complementarity in the labor market, and the evolution of opinions within online communities. Integrating quantitative and theoretical foundations with a literature review from diverse disciplines constitutes the essential element for rigorously addressing the outlined challenges.
The work within this project is divided into two macro-areas, schematically described below.
1. Potential and Limitations of AI
Use of LLMs to study and predict innovation dynamics. Specific applications to sustainability themes (industrial pollution, critical raw materials, energy transition).
The use of open-source Large Language Models (LLMs) enables the automated analysis of vast volumes of unstructured data (patents, publications, code) for the geographical and semantic mapping of technological skills. This approach is employed to forecast innovation trajectories through new combinations (recombinant innovation) and to connect policy recommendations with local territorial capabilities, particularly for environmental policies (e.g., critical raw materials and industrial pollution reduction).
Use of LLMs for the study of social phenomena within online communities.
We explore the feasibility of using LLMs to gather signals and map social dynamics (such as social unrest and inequality) from publicly accessible discourses on social networking platforms.
Theory of Artificial Creativity.
The research focuses on Continual Learning for the progressive integration of new information while avoiding Catastrophic Forgetting. Key contributions include Dreaming Learning (published at NeurIPS 2024), which mitigates Model Collapse and enables the controlled exploration of new information configurations. In support of this, the theoretical work First-Extinction Law for Resampling Processes provides an analytical description of Model Collapse. Furthermore, we develop Lyapunov Learning (published at ICML 2025), based on the Edge of Chaos, to extend the capabilities of Dreaming Learning to complex multidimensional variables (e.g., climate change). Practical applications include the Duetto project, an advanced system that utilizes Transformer architectures and vector quantization for the real-time analysis and anticipation of human movement in dance.
2. Socio-Economic Impact of AI
Impact on the labor market.
The objective is to provide a quantitative perspective on the impact of AI on labor, structured into three tasks:
Measuring Occupational Exposure through the analysis of AI startup pitches;
Identifying Emerging Skills by analyzing job postings and code repositories (GitHub);
Studying Public Opinion and Ethics to understand the influence of social trust on AI adoption.
Methodologically, LLMs will be used to link startup descriptions (from datasets like Crunchbase) and code repository content (from readme files) with occupational activity descriptions (from O*NET) in order to measure research, development, and funding activities in the AI field aimed at automating specific tasks.
Human Values and AI Values.
This area aims to develop “Responsible Artificial Intelligence” models that reflect local moral values, with the goal of curbing harmful content and fostering ethical interactions. The research analyzes how moral values influence the evolution of opinions in online communities and the extent to which the “irreducible core elements” of morality are represented and understood by AI algorithms, including BOTs.
