Current evaluation of LLM-generated debate largely focuses on argumentative quality, leaving the value structure of generated arguments underexamined. This article introduces a diagnostic metric framework for comparing how values are expressed and organized in human-reference and AI-generated debates. The framework separates value expression, the range of recognizable values a debate invokes, from value organization, how those values are distributed and sustained across turns. Grounded in Schwartz's Basic-10 value theory, it represents debates as multi-level value distributions over argumentative units, utterances, stances, and full debates, enabling metrics for value coverage, alignment, diffusion, and cross-turn dynamics.
Separating Value Expression from Value Organization
The core innovation of this framework is its decoupling of two dimensions that have been conflated in prior work. Value expression captures the breadth of values a debate activates—whether it touches on benevolence, universalism, conformity, or other Schwartz Basic-10 values. Value organization tracks how those values are distributed and sustained across argumentative turns. By representing debates as multi-level value distributions, the framework can measure not just which values appear, but how persistently they are maintained, how they shift from turn to turn, and whether they cluster around particular stances or argumentative units.
AI-Generated Debates Activate Comparable Value Dimensions
Across multiple generation systems, AI-generated debates are broadly value-expressive, often activating as many or more value dimensions than matched human references. This finding challenges the assumption that models inevitably produce narrower or more stereotyped value ranges. The framework's metrics for value coverage reveal that models can span the full Schwartz Basic-10 set, even when their argumentative quality or coherence is judged differently. However, the distribution of those values across utterances and stances differs systematically from human debates.
Differences in Value Organization
The framework's organization metrics show that AI-generated utterances are more locally value-diffuse. In human debates, values tend to be anchored to specific stances or persistent positions taken by participants. In contrast, model-generated debates spread values more thinly across individual turns, without the same sustained commitment to particular value clusters. Source-blind judges in the study were able to distinguish value clarity—the perceived definiteness with which a value is expressed—from value focus—the extent to which a debate maintains a coherent value trajectory. This distinction explains why models can score well on value coverage metrics while still producing debates that feel unfocused or meandering to human evaluators.
Implications for Debate Evaluation
These findings indicate that evaluation of AI-generated debate should consider not only value presence but also value organization. A debate that activates many values but distributes them diffusely across turns may score poorly on organization metrics even as it scores well on coverage. The framework provides concrete metrics that can be incorporated into existing evaluation pipelines, allowing researchers and developers to diagnose whether shortcomings in generated debates stem from limited value range or from poor value sustainment and alignment across turns. Code and supplementary materials are available at the project's https URL.