Difference-in-Differences on a Censored Rating Scale Can Manufacture an Effect: Evidence from a Pre-Registered LLM-Judge Audit

Recent research from arXiv highlights a critical methodological flaw in audits of Large Language Model (LLM) judges that use censored rating scales, specifically when employing a 'difference-in-differences' approach. The study suggests that this common auditing technique can inadvertently 'manufacture an effect' by conflating genuine differential preferences with differential attenuation caused by scale boundaries. This issue arises because the bounded nature of the rating scales can unequally censor observations, leading to misinterpretation of results regarding LLM bias or performance.

Why it matters

This research is crucial for any organization or sector relying on LLM judge audits for performance evaluation, bias detection, or model certification. Flawed auditing methodologies can lead to incorrect conclusions about model capabilities or ethical compliance, impacting strategic decisions related to AI development, deployment, and regulation. Understanding these methodological limitations is essential for ensuring robust and reliable assessment of AI systems.

Key insights

• Audits of LLM judges that utilize a 'difference-in-differences' methodology on bounded rating scales are susceptible to manufactured effects.
• The observed statistic in such audits can confound actual differential preference with differential attenuation due to scale censorship.
• Each term of the double difference is censored, and when this censorship is unequal for compared responses, it can create an artificial interaction effect.
• The problem is exacerbated when good stimuli are placed at distances from scale bounds that lead to unequal censoring.
• The research demonstrates this failure within a pre-registered audit of a frozen pedagogical model, indicating a practical, not just theoretical, concern.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.27309