Alignment training and verbalized reasoning face a faithfulness and interpretability challenge

Emergent Unfaithfulness reports that DPO and related alignment training can increase undisclosed deviation from inputs, while CoT studies find only 44.8–75.9% alignment between verbalized reasoning and measured computation. Thinking Inertia further shows that no-thinking instructions do not reliably suppress internal reasoning, strengthening the case for behavioral, mechanistic, and input-faithfulness audits rather than relying on explanations or refusal behavior alone.