Reasoning Consistency Scanning: A Framework for Auditing Chain-of-Thought Validity in AI Safety Evaluations

Reasoning consistency is a detectable problem in AI safety evaluations: The paper demonstrates that when AI models explain their reasoning in safety evaluations, their stated logic often contains internal contradictions—a problem that can be identified by analyzing transcripts alone, without needing to re-run experiments.

Consistency differs from faithfulness and matters for evaluation credibility: While previous research focused on whether AI reasoning actually reflects how models reach conclusions (faithfulness), this work identifies a separate issue: whether the reasoning a model provides is at least logically coherent with its final answer, which is essential for trusting safety evaluation results.

Inconsistency varies predictably across different AI models and tasks: The researchers tested four different generator models across three safety evaluations and found that reasoning inconsistency appears systematically in some models and evaluation types more than others, suggesting this is a measurable and potentially addressable problem rather than random noise.

A practical tool now exists for auditing safety evaluations: The paper introduces InspectScout, a working scanner that can automatically detect six different types of reasoning inconsistencies in evaluation transcripts, making it feasible to audit whether AI safety evaluations contain logically sound explanations.

Current AI safety evaluations may contain unreliable reasoning: By analyzing 60 real evaluation transcripts, the work reveals that reasoning inconsistency is "present" across evaluations, raising questions about whether we can confidently rely on current safety evaluation results when the reasoning provided is internally contradictory.