Staying with the Uncertainty: Uncertainty-Scaffolding Strategies for Artificial Moral Advisors in LLM-to-LLM Simulated Conversations

Researchers study how Large Language Models deployed as Artificial Moral Advisors should communicate with users discussing ethical dilemmas, proposing three uncertainty-focused conversation strategies and finding that different approaches sustain distinct quality levels of engagement rather than producing uniform belief revision.

This research addresses a critical gap in LLM deployment: how AI systems should facilitate ethical deliberation rather than push users toward predetermined conclusions. The study moves beyond traditional chatbot evaluation metrics by focusing on conversational quality during moral reasoning, where nuance and perspective exploration matter more than reaching agreement. The researchers tested six distinct communication strategies—three emphasizing uncertainty management (Perspective-Multiplying, Tension-Preserving, Process-Reflecting) against controls—revealing that LLMs exhibit fundamentally different reasoning patterns depending on their architecture. Open-source models maintain ambiguity through diverse outputs across different personas, while proprietary systems employ hedging language within consistent personas, both effectively modeling human uncertainty but through opposite mechanisms.

The findings challenge assumptions about LLM standardization. The distinction between declarative and narrative persona prompts—where declarative formats better capture initial disagreement while narrative formats show more realistic belief evolution—suggests that prompt engineering significantly affects not just output quality but the cognitive authenticity of simulated reasoning. Rather than demonstrating that one strategy universally succeeds, the research shows that uncertainty strategies differ primarily in engagement quality, not stance revision magnitude. This distinction proves crucial for real-world applications where maintaining thoughtful deliberation prevents premature closure on complex moral questions.

For AI developers building conversational systems in ethics, policy, counseling, or education domains, this research provides empirical guidance on communication design. The work underscores that effective moral advisory requires preserving complexity rather than reducing it, and that different model architectures demand distinct approaches to achieve authentic uncertainty handling. Future systems targeting ethical domains should prioritize engagement quality metrics alongside traditional performance indicators.

• →Open and closed LLMs manage moral uncertainty through different mechanisms—divergence versus hedging—both achieving human-like ambiguity representation.
• →Uncertainty-focused conversational strategies sustain higher engagement quality than persuasive or baseline approaches, independent of belief revision magnitude.
• →Narrative persona prompts better simulate realistic belief revision patterns compared to declarative formats during ethical deliberation.
• →LLM deployment as moral advisors requires strategy-specific design; no single communication approach optimizes all dimensions of ethical dialogue.
• →Prompt format and model architecture jointly determine both conversational authenticity and uncertainty handling capability in moral reasoning scenarios.