A new preprint compares human and AI debates by the values they invoke and how those values persist across turns. Generated debates can mention as many value dimensions as human examples while organizing them differently. The framework uses Chinese competitive-debate transcripts and separates value expression from factual accuracy, persuasiveness and a model’s internal beliefs.
A new arXiv preprint proposes a framework for analysing the values expressed in AI-generated debates. It uses Schwartz's theory, a classification of ten basic values. Each argumentative unit joins a stance-specific claim to the passage giving its supporting premise. The system assigns a distribution across values to each unit and aggregates those distributions for utterances, stances and the whole exchange. The findings come from a single preprint.[1]
The comparison pairs Chinese competitive-debate transcripts with generated debates containing the same number of turns. It includes general-purpose language models and R-Debater, a debate system. The researchers find that the number of value dimensions invoked by generated debates can match or exceed the human examples. Their arrangement differs within an utterance and through the course of a stance. Separate metrics measure coverage, alignment with the human example, local diffusion, stability and drift.[1]
Human evaluators assess value clarity and focus with the origin of each text concealed. Human labels provide a check on automatically extracted premises and assigned values, though implicit and culturally specific appeals can still produce errors. A diffuse distribution can reflect several perspectives. The framework describes values expressed in text, rather than a model's internal beliefs, and its fixed-turn setup simplifies interruptions and audience responses found in live debate.[1]
