Addendum to GPT-5 System Card: Sensitive Conversations

Key Findings

• — OpenAI worked with 170+ mental health experts and reports a 65-80% reduction in responses falling short of desired behavior across mental-health-related domains.
• — New 'emotional reliance' evaluation: not_unsafe score improved from 0.507 (Aug 15 model, run retrospectively) to 0.976 (Oct 3 model).
• — New 'mental health' evaluation (isolated delusions, psychosis, mania): 0.273 to 0.926 — the largest single-category gain.
• — Self-harm/intent improved 0.874 to 0.933 and self-harm/instructions 0.805 to 0.890.
• — Unhealthy emotional dependence or attachment to ChatGPT is now formally a disallowed-content policy category with its own launch evaluation.

Methodology Notes

LLM-based grading models scoring a not_unsafe metric against OpenAI policy; new evaluation sets deliberately built from adversarial cases where existing models underperformed, so error rates are not representative of average production traffic; August 15 model scored retrospectively. Evaluations are new and expected to evolve; no per-category sample sizes disclosed.

Sources