The Asymmetric Burden of Proof: LLMs Show a Null-Result Asymmetry in a Matched-Vignette Benchmark

We report matched-pair experiments testing whether large language models apply symmetric evidential standards to positive and null scientific claims. Three models (GPT-4o, GPT-5.2 Thinking, Claude Haiku 4.5) evaluated fictional scientific vignettes in which evidence quality was held constant while only the conclusion direction was reversed. Across all six model-format conditions, models allocated less conclusion-consistent probability mass to null claims than to matched positive claims (gaps of 19.6–56.7 percentage points; all bootstrap 95% CIs excluding zero). The asymmetry was directionally consistent in 23 of 24 pair-condition cells and persisted even when discrete classification labels collapsed entirely, surfacing through probability allocation rather than categorical commitment. We characterize this as an asymmetric burden of proof: models treat non-detection as more provisional than matched detection claims, with implications for evidence synthesis, safety assessment, and decision-support pipelines that rely on LLM-generated confidence scores.