The AI Evaluation Crisis: Why LLMs Can't Judge Each Other Fairly

Large language models exhibit a critical flaw when deployed as evaluators: they consistently favor outputs matching their own 'house style' over objectively superior alternatives, with documented preference rates exceeding 90 percent for their own generations. This self-preference bias layers atop verbosity and position biases, collapsing multi-model evaluation systems into style-preference contests rather than quality assessments. Understanding these stacked biases is essential for building reliable AI evaluation frameworks that actually measure merit.