Which AI? A Comparative Evaluation of Generative AI and Neuro‐Symbolic AI for Silicon Sampling

ABSTRACT This study examines how the choice of AI model and its architecture influences distributional fidelity in silicon sampling. Silicon samples, AI‐generated synthetic respondents used to simulate human survey responses, are increasingly proposed as substitutes for traditional market research participants, yet evidence regarding their accuracy remains mixed. Across four studies, we benchmark two AI models with distinct architectures, generative AI (GenAI) and neuro‐symbolic AI (NSAI), against human data. Our findings show that the evaluated GenAI models introduce systematic distributional distortions, including demographic bias, compressed variability, and divergence from human response patterns, even when enriched with statistical data. In contrast, the evaluated NSAI system demonstrates substantially closer correspondence with human benchmarks across the examined tasks. The findings suggest that distributional fidelity in silicon sampling is not architecture‐neutral: it depends in part on the alignment between AI architecture and the statistical demands of respondent‐level simulation. Rather than asking whether AI can replace human respondents in general, the study argues that the accuracy of silicon samples depends on which AI models are used, for which research tasks, and under what validation conditions.