The study also reveals that 86.1% of bias incidents required only a single prompt, underscoring how easily AI models can still produce biased outputs despite advances in safety.