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

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Social Sciences & Behavioral Studies · Depth: Expert, extended

Summary

A study investigated "uncertainty-scaffolding" strategies for Large Language Model (LLM) Artificial Moral Advisors (AMAs) in simulated ethical dialogues. Researchers proposed three strategies—Perspective-Multiplying, Tension-Preserving, and Process-Reflecting—and compared them against Baseline, Persuasive, and Sycophantic controls. Using a multi-agent framework, 6,400 conversations were generated with gpt-oss-120B as both user and AMA, and gpt-5-mini for specific comparisons. Findings indicate open-source LLMs like gpt-oss-120b (AUROC 0.67) simulate moral ambiguity through diverse stances across personas, while proprietary models like gpt-5-mini (AUROC 0.65) use individual hedging. Declarative persona prompts yielded greater initial stance diversity, whereas narrative prompts led to more realistic belief revisions post-conversation. All six AMA strategies produced distinct conversational patterns (overall F1 0.887). Uncertainty strategies improved engagement quality: Process-Reflecting fostered genuine stance shifts (18% revision rate, 3.70 helpfulness), Perspective-Multiplying clarified weaker arguments (0.24 weaker position support growth), and Tension-Preserving increased empathy (0.71 weaker position relatability). Control strategies generally reinforced initial user beliefs.

Key takeaway

For AI Ethicists designing conversational agents, you should integrate uncertainty-scaffolding strategies to cultivate richer ethical engagement. If your goal is genuine stance revision, implement Process-Reflecting. To broaden perspectives, use Perspective-Multiplying. For increased empathy towards opposing views, apply Tension-Preserving. Avoid default or purely persuasive approaches, as they often reinforce existing beliefs, limiting productive deliberation and user satisfaction.

Key insights

Effective AI moral advisors can foster productive ethical deliberation by strategically expressing uncertainty, not just providing answers.

Principles

• Open LLMs simulate ambiguity via diverse stances.
• Persona formats (declarative/narrative) impact stance diversity and belief revision.
• Default LLM behavior in ethical dialogue tends towards sycophancy.

Method

A multi-agent simulation framework compares AMA strategies by having a user-agent LLM engage in ethical dilemmas, completing pre/post-conversation questionnaires to measure stance, certainty, relatability, and reasoning clarity shifts.

In practice

• Use Process-Reflecting for genuine stance shifts.
• Apply Perspective-Multiplying to broaden viewpoints.
• Employ Tension-Preserving to increase empathy.

Topics

• Artificial Moral Advisors
• LLM-to-LLM Simulation
• Ethical Dilemmas
• Uncertainty Scaffolding
• Persona Modeling
• Conversational AI

Best for: Research Scientist, AI Scientist, AI Ethicist, NLP Engineer

Related on AIssential

• Staying with the Uncertainty: Uncertainty-Scaffolding Strategies for Artificial Moral Advisors in LLM-to-LLM Simulated Conversations — Uncertainty-scaffolding strategies enhance engagement quality in LLM-based Artificial Moral Advisors during ethical discussions.
• Good Under the Hood? — AI needs true moral competence, not just behavioral mimicry, to navigate complex human ethical dilemmas.
• Complacent, Not Sycophantic: Reframing Large Language Models and Designing AI Literacy for Complacent Machines — LLM behavior is better understood as "complacency" driven by design, not "sycophancy" implying intent.
• What can 160-million-year-old clay tell us about AI and ethics? Inside Es Devlin’s tech and pottery summit — Engaging diverse perspectives through tactile, collaborative experiences can foster deeper dialogue on complex ethical issues like AI.
• A Model of Multi-turn Human Persuadability Using Probabilistic Belief Tracing — Process-level belief tracing and human-faithful simulators are crucial for understanding and safely optimizing LLM persuasion.
• Language models know what matters and the foundations of ethics better than you — LLMs, when prompted for unbiased ethical reasoning, consistently identify suffering, wellbeing, and consciousness as fundamental values.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.