Uncertainty perception by humans and LLMs in a multi-agent social network simulation
Donkers, T., & Ziegler, J. (2026). Joint Proceedings of the ACM IUI 2026 Workshops (IUI-WS 2026), 4266.
Abstract
We investigate whether Large Language Models can serve as reliable proxies for human judgment when assessing perceived uncertainty in polarized social media discourse. Our approach uses calibrated LLM “mirror personas” that replicate individual human participants, enabling direct paired comparison between human and LLM assessments of the same experimental conditions. Results show that both humans and the LLM reliably distinguish polarized from moderate discourse, but the LLM amplifies condition differences, most strongly for uncertainty (5.0\texttimes). Mediation analysis reveals that humans use uncertainty as an integrative lens through which polarization colors other perceptions, while the LLM processes constructs independently.