This research explores how persona-dependent alignment affects large language models' decisions in moral dilemmas, revealing significant variations based on sociodemographic contexts.

• LLMs make different moral choices when adopting different personas
• Moral decisions align with human judgment patterns across cultural and demographic dimensions
• Models show inconsistent reasoning despite similar decisions across personas
• Findings highlight the importance of considering demographic representation in AI alignment

Security Implications: This research underscores critical concerns for safe AI deployment, demonstrating that LLM behavior in ethical dilemmas can vary substantially depending on the persona adopted, raising questions about alignment consistency and potential bias in high-stakes applications.

Exploring Persona-dependent LLM Alignment for the Moral Machine Experiment