In an interesting turn of events, researchers have uncovered that large language models (LLMs) display significant favoritism toward specific character names that vary by model and version. For instance, the pairing of names like Elena Vasquez and Marcus Chen has been predominantly linked to outputs from the Claude model. This revelation emerged as a secondary finding during the development of a model-differentiation technique known as CDD, ultimately evolving into a comprehensive study on this naming phenomenon.
The research highlights that these favored names travel together in what can be described as 'correlated ensembles.' They frequently show up across various websites in roles such as volcano experts, podcast hosts, and protagonists in thrillers, as well as authors of numerous academic papers published in rapid succession. Notably, the researchers discovered a third name that consistently appears alongside the first two, leading to a fascinating collage of instances where different websites independently generated content featuring this trio, often accompanied by AI-generated stock photos.
These findings raise important questions about the influence of language models in shaping narratives and character identities in AI-generated content. The full study can be accessed in the preprint linked here.
