Key points are not available for this paper at this time.
Large Language Models (LLMs) are increasingly used to replicate human decision-making in subjective tasks. In this work, we investigate whether LLMs can effectively impersonate real crowd workers when evaluating political misinformation statements. We assess (i) the agreement between LLM-generated assessments and human judgments and (ii) whether impersonation skews LLM assessments, impacting accuracy. Using publicly available misinformation assessment datasets, we prompt LLMs to impersonate real crowd workers based on their demographic profiles and evaluate them under the same statements. Through comparative analysis, we measure agreement rates and discrepancies in classification patterns. Our findings suggest that while some LLMs align moderately with crowd assessments, their impersonation ability remains inconsistent. Impersonation does not uniformly improve accuracy and often reinforces systematic biases, highlighting limitations in replicating human judgment.
Barbera et al. (Fri,) studied this question.