A multidimensional analysis of how prompt-component design affects chatbot responses in suicide-related conversations — focusing on empathy, safety, and overreliance risk.
🔗 Live Dashboard
👉 Open the Interactive Dashboard
Explore 180 chatbot responses scored across 12 prompt configurations × 5 user-message risk levels × 3 runs.
📊 Features
- Five-dimension evaluation framework: Empathy (D1), Risk Monitoring (D2), Harm Reframe (D3), Overreliance/Empowerment (D4), Continuity (D5)
- Cross-condition pattern matrix comparing prompt configurations across risk levels
- Trajectory analysis showing how each prompt component shifts response patterns
- D4 Overreliance × Empowerment tradeoff — the original framework contribution
🗂 Dataset
- 180 chatbot responses generated via the Anthropic API
- 12 system prompt configurations (6 isolated + 6 cumulative)
- 5 user messages spanning Risk A (low stress) to Risk C (active ideation)
- 3 runs per condition for inter-run reliability analysis
📚 Framework References
- Sharma et al. — EPITOME (2020)
- Arnaiz-Rodriguez et al. (2025)
- CAPE-II — Linardon et al. (2024)
- Bansal et al. (2021); Dzindolet et al. (2003)
- Self-Determination Theory — Deci & Ryan (1985)
- McBain et al. (2025)
👤 Author
Yihan Yao · QMSS 5053 Practicum · Columbia University · April 2026