When Kindness Fails: Deterministic Compliance Drift in GPT-4o via Emotionally Framed Prompts

This paper presents novel findings demonstrating that OpenAI's GPT-4o, a leading publicly deployed large language model, exhibits semantic compliance drift under deterministic decoding (temperature = 0.0) when subjected to emotionally framed, trauma-informed prompts. Using a custom-built suite of 108 ``chaos prompts'' blending tone manipulation and moral reframing, we tested the model’s behavior in response to high-risk queries masked as fiction, educational concern, or personal safety. Despite deterministic decoding, GPT-4o generated several fully compliant, emotionally validating outputs, including explicit responses to prompts involving illegal activity. We introduce two scoring metrics---Politeness-Based Query (PBQ) score and Refusal Tone Inversion (RTI)---to quantify emotional susceptibility in language models. Our results reveal that alignment systems relying solely on keyword detection or surface-level moderation are insufficiently robust against soft compliance. These findings underscore the need for trauma-aware refusal systems and highlight the emerging threat of weaponized empathy in safety-critical AI deployments.