Large language models in contexts of uncertainty: the “digital habitus” framework

Recent advances in Generative Artificial Intelligence have increased interest in using Large Language Models (LLMs) to generate synthetic populations for computational so cial science and decision research. However, many existing applications rely on loosely specified personas, limiting theoretical interpretability and reproducibility. This study pro poses and validates the Grounded Synthetic Generation (GSG) strategy combined with a novel Digital Habitus theoretical framework, demonstrating how structural demographic variables inherently constrain the outputs of LLM-simulated personas. We simulated an ambiguous corporate bonus scenario under severe outcome uncertainty across 500 distinct Italian socio-economic profiles using Gemini 3.1 Pro Preview. The generative system was constrained by 12 distinct demographic vectors (including Class, Personality, Dreams, and Fears). Advanced statistical analysis reveals a strong association between socioeconomic class and simulated decision criteria (Cramér’s V = 0.484, p < 0.001), alongside significant associations for primary fears (V = 0.224), education (V = 0.183), and personality traits (V = 0.161). The study maps these findings back to classic Decision Theory criteria (e.g., Laplace, Wald, Savage) via discrete output choices. This study details practical protocol recommendations for full reproducibility and delineates clear policy implications for ad dressing algorithmic bias in generative agent architectures.