Evaluating LLM Personas as Proxies for Human Populations

Abstract: This talk presents ongoing work on whether and how LLMs can be guided through personas to simulate human perspectives in support of policy analysis. Recent experiments, including our own, show that persona prompting allows models to reproduce average response patterns. However, these experiments also expose significant limitations: LLMs tend to flatten demographic nuance, produce exaggerated archetypes rather than authentic variation, and systematically collapse the diversity of viewpoints found in real human populations. Results from this project are presented alongside findings from related literature, with particular attention to how these “Artificial Hivemind” behaviors emerge from current training and alignment paradigms. The talk concludes by presenting ongoing research aimed at better aligning LLM outputs with observed human response distributions.

Short bio: Axel Abels is a Postdoctoral Researcher at the Université Libre de Bruxelles’ Machine Learning Group. His research focuses on collective intelligence, particularly on methods for improving the wisdom of crowds. His recent work explores hybrid collective intelligence systems that combine human groups and large language models, leading to current research on the capabilities and limits of LLMs in representing human diversity when used as proxies for human populations.

Presenter: Axel Abels (Université Libre de Bruxelles’)

Date: 2026-02-12 11:00 (CET)

Location: Oficinas ELLIS Alicante, Muelle Pte., 5 – Edificio A, Alicante 03001, Alicante ES

Add to Google Calendar, Outlook Calendar