Personalized alignment of large language models and its impact on moral judgment

Key Points

• The study aims to explore the impact of personalized alignment on moral judgment in large language models using the HEXACO personality model.
• Two studies tested the effectiveness of prompts reflecting HEXACO traits on language model outputs.
• Study 1 confirmed that language models can express HEXACO personality traits through specific prompts.
• Study 2 examined the alignment's effect on utilitarian tendencies in models compared to human judgments.
• High levels of honesty-humility, agreeableness, and conscientiousness prompts significantly reduced utilitarian choices by GPT-3.5, GPT-4, and ERNIE 3.5.
• The alignment framework highlights the moral implications of specific personality traits in language models.