A preprint based on 646 adults compares AI models on video interviews, reporting a modest multimodal edge and more reliable ratings for targeted questions.