CoT is Not the Chain of Truth: An Empirical Internal Analysis of Reasoning LLMs for Fake News Generation

From generating headlines to fabricating news, the Large Language Models (LLMs) are typically assessed by their final outputs, under the safety assumption that a refusal response signifies safe reasoning throughout the entire process. Challenging this assumption, our study reveals that during fake news generation, even when a model rejects a harmful request, its Chain-of-Thought (CoT) reasoning may still internally contain and propagate unsafe narratives. To analyze this phenomenon, we introduce a unified safety-analysis framework that systematically deconstructs CoT generation across model layers and evaluates the role of individual attention heads through Jacobian-based spectral metrics. Within this framework, we introduce three interpretable measures: stability, geometry, and energy to quantify how specific attention heads respond or embed deceptive reasoning patterns. Extensive experiments on multiple reasoning-oriented LLMs show that the generation risk rise significantly when the thinking mode is activated, where the critical routing decisions concentrated in only a few contiguous mid-depth layers. By precisely identifying the attention heads responsible for this divergence, our work challenges the assumption that refusal implies safety and provides a new understanding perspective for mitigating latent reasoning risks.

Artificial intelligence models are often considered safe if they simply refuse harmful requests. However, this assumes that their internal “thinking process” is also secure. Our research challenges this assumption: we found that even when an AI politely declines a dangerous prompt—such as generating fake news—its step-by-step reasoning can still internally develop and circulate unsafe ideas. To uncover this hidden risk, we developed a new analytical framework that tracks how information flows through the model’s internal decision-making components. By introducing three measurable indicators, we can pinpoint exactly when and where the AI’s reasoning begins to lean toward deception. Our experiments show that enabling the AI’s “thinking mode” significantly amplifies these latent risks, with critical unsafe decisions concentrated in just a few middle processing stages. Ultimately, this work demonstrates that a surface-level refusal does not guarantee true safety. By mapping where unsafe reasoning emerges, we provide developers with a practical roadmap to identify and correct hidden vulnerabilities, paving the way for AI systems that are safe not just in their answers, but in their entire thought process.