This research explores how different personas influence large language model decisions in moral dilemma scenarios, revealing important insights for AI alignment and safety.

• LLM moral decisions vary substantially based on assigned personas
• Models demonstrate inconsistent moral reasoning across sociodemographic contexts
• Persona-dependent alignment poses significant challenges for deploying LLMs with agency
• Results highlight the need for robust alignment techniques that account for societal diversity

For security professionals, this research underscores critical concerns about AI alignment in high-stakes scenarios and emphasizes the importance of accounting for sociodemographic factors when deploying AI systems that make moral judgments.

Exploring Persona-dependent LLM Alignment for the Moral Machine Experiment