This research introduces a systematic approach to evaluate and improve the safety and reliability of AI chatbots in mental health contexts.

• Developed a 100-question benchmark with ideal responses validated by mental health experts
• Created five guideline questions as evaluation criteria for chatbot responses
• Tested framework on a GPT-3.5-turbo-based mental health chatbot
• Provides tools to identify and mitigate risks of harmful AI responses

With mental health chatbots becoming more accessible due to their human-like interactions and 24/7 availability, this framework offers crucial guardrails to ensure they provide safe, ethical support rather than potentially harmful advice.

Building Trust in Mental Health Chatbots: Safety Metrics and LLM-Based Evaluation Tools