Research
Study finds LLM evaluators shift credibility ratings based on claimed source of identical arguments
A preprint tested 20 arguments across four source conditions and four models, finding credibility was the most source-sensitive dimension. Irrelevant expert attribution sometimes lowered credibility below no-source. Interaction results were not statistically robust and the study is not yet peer rev…
Every outlet that ran it 1 article
r/MachineLearningCommunityLead