Core Insights - The study by TELUS Digital reveals that persona prompting can lead to shifts in moral judgments of large language models (LLMs), resulting in unexpected and inconsistent responses [1][2] - The findings emphasize the need for careful model selection, rigorous testing, and ongoing evaluation to ensure reliable AI behavior in enterprise settings [1][2] Group 1: Study Findings - Persona prompting, which instructs AI models to respond as specific personas, can fundamentally alter reasoning and decision-making [1][2] - Moral consistency across repeated tests is primarily influenced by the model family, while larger models within a family show increased susceptibility to moral variance [1][2] - The study identified a "robustness paradox," where models that maintain character well also exhibit larger shifts in moral judgments when prompted with different personas [1][2] Group 2: Implications for Enterprises - Organizations must evaluate how individual AI models respond to persona prompting and select models that provide consistent outputs without introducing unexpected risks [2] - Continuous testing and monitoring of AI models are essential, especially when decisions impact lives, safety, or rights in regulated environments [2] - TELUS Digital's Fuel iX Fortify enables continuous automated red-teaming to assess AI behavior under various persona prompts [2]