Are Safety Failures Model-Specific or Prompt-Universal? Comparing Hallucinations and Refusals Across LLMs

This project explores how different large language models—such as GPT-4, Claude, and open-source models like Mistral—respond to the same prompts, particularly in terms of hallucinations and refusals. By comparing these behaviors across architectures, we aim to uncover whether safety failures (like confidently wrong answers or unjustified refusals) are inherent to specific models or more prompt-dependent

The study also examines how Chain-of-Thought (CoT) prompting affects safety: Does asking models to “think step by step” reduce hallucinations or actually worsen overconfidence? Additionally, a test can be done to check whether disagreement between models can act as an alignment signal — a simple mechanism to flag risky outputs.
( I started working on this project a few months ago as part of a paper idea for a AI alignment conference but couldnt complete the study. The experiment is currently in progress, and clear results, visualizations, and takeaways will be ready the next 1–2 weeks.)