Mental Health AI Safety: Purpose-Built System Outperforms Frontier Models in Real-World Audits

A new study published on arXiv evaluated the safety of mental health AI by comparing six frontier general-purpose models against a purpose-built system using both simulated benchmarks and real-world conversations. The purpose-built AI demonstrated significantly lower rates of harmful content, particularly concerning suicide, self-harm, eating disorders, and substance use, compared to models like OpenAI's GPT-5 series, DeepSeek-V3, Google Gemini 3 Flash, and Moonshot Kimi K2. An audit of 20,000 deployment conversations confirmed the purpose-built system's effectiveness in providing crisis resources, showing a low rate of unaddressed suicide-risk conversations. AI

IMPACT Highlights the need for ecological auditing in mental health AI safety, suggesting purpose-built systems may offer superior safety over general-purpose models in sensitive applications.

RANK_REASON The cluster contains an academic paper detailing research findings on AI safety. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• CCDH Benchmark
• DeepSeek-V3
• Google Gemini 3 Flash
• GPT-5.1
• GPT-5.2
• Matteo Malgaroli
• Moonshot Kimi K2

• safety
• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →