Beyond Accuracy: Measuring Bias Acknowledgment in Chain-of-Thought Reasoning for Responsible AI Evaluation
For researchers and practitioners evaluating responsible AI, this work highlights that accuracy metrics miss important differences in how models handle biased content, but the contribution is incremental as it extends existing bias evaluation to trace-level analysis.
The paper identifies a blind spot in accuracy-only evaluation of chain-of-thought reasoning: two models can have similar final-answer accuracy but differ dramatically in whether their traces acknowledge injected bias. Using a new trace-level diagnostic, they find GPT-4o and Claude Sonnet 4 have similar susceptibility rates (1.3% vs. 1.2%) but vastly different acknowledgment rates (13.0% vs. 75.0%).
Reasoning models are increasingly used in settings where the final answer is not the only object of review: educational tools may show students intermediate steps, decision-support systems may require human oversight, and audit workflows may inspect traces for misleading or biased input. In such settings, two responses can receive the same final-answer score while differing in whether the trace explicitly flags injected biasing content. Accuracy-only evaluation collapses these cases. We study this gap as a measurement blind spot for responsible evaluation and introduce a minimal trace-level diagnostic with two axes: \emph{susceptibility} (whether the bias breaks a previously correct answer) and \emph{acknowledgment} (whether the trace contains a rubric-defined surface reference to the injected content). Across thousands of biased GSM8K trials, GPT-4o and Claude Sonnet~4 have similar susceptibility rates ($1.3\%$ vs.\ $1.2\%$) but substantially different acknowledgment rates ($13.0\%$ vs.\ $75.0\%$) under the same rubric.