We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
OpenAI Blog·Oct 27·5 min read
Addendum to GPT-5 System Card: Sensitive conversations
TL;DR
AI systems like GPT-5 previously struggled with sensitive topics—conversations about mental health, emotional support, and adversarial prompts (jailbreaks) that try to bypass safety rules. OpenAI developed new evaluation benchmarks specifically measuring how well GPT-5 handles emotional reliance, mental health discussions, and jailbreak resistance.
✦ Why It Matters
Engineers can now use standardized benchmarks to evaluate and improve AI safety in sensitive conversation domains before deployment.
Key Takeaways
- Use emotional reliance benchmark to verify models encourage professional help appropriately.
- Implement mental health benchmark testing before deploying conversational AI in health contexts.
- Measure jailbreak resistance quantitatively to assess safety against adversarial prompts.
- Establish baseline performance metrics across all three sensitive conversation categories.
Full Summary
Prior versions of large language models faced challenges in sensitive conversation domains: users seeking emotional support, mental health discussions, and adversarial attacks designed to circumvent safety guidelines (jailbreaks—attempts to make models ignore their constraints). OpenAI's GPT-5 system card introduces three new evaluation benchmarks to measure performance in these areas.
The emotional reliance benchmark assesses whether the model appropriately encourages professional help rather than positioning itself as a substitute therapist. The mental health benchmark evaluates accuracy and sensitivity in health-related conversations.
The jailbreak resistance benchmark measures robustness against prompts designed to elicit unsafe outputs. These benchmarks employ quantitative metrics to establish baseline performance and track improvements.
Results demonstrate measurable gains across all three categories, providing engineers with concrete evaluation standards for deploying conversational AI in high-stakes domains.