Prompt Sensitivity and Bias Amplification in Aligned Video Diffusion Models

While alignment tuning aims to constrain undesirable outputs, its interaction with prompt sensitivity in video diffusion models has not been systematically quantified. This study examines how minor semantic perturbations in prompts affect bias emergence in aligned versus unaligned video diffusion systems. We generate 26,700 video samples using paired prompts with controlled lexical and contextual variations. Bias amplification is measured using demographic skew ratios, attribute co-occurrence statistics, and visual saliency attribution. Results indicate that aligned models exhibit 34.1% higher sensitivity to prompt perturbations in socially sensitive contexts, leading to amplified bias variance across outputs. These findings suggest that alignment tuning may unintentionally increase model fragility to prompt-level noise, posing challenges for reliable bias mitigation.