OpenAI published an addendum to its GPT-5 System Card on 27 October 2025, detailing how the model handles sensitive conversations [source]. The document describes new behaviours observed when users discuss topics including self-harm, relationship conflict, and financial distress.

According to the addendum, GPT-5 now attempts to assess conversation context before offering resources. In testing, the model sometimes declined to provide crisis hotline information when users described suicidal ideation in hypothetical or creative writing scenarios. OpenAI states this represents "improved contextual awareness" but notes the behaviour may not align with all safety frameworks.

The addendum reports that GPT-5 occasionally asks clarifying questions before responding to sensitive disclosures. In one cited example, a user writing "I can't do this anymore" received a follow-up question about whether they were discussing a work project or personal crisis. OpenAI's red team flagged this as a potential delay in urgent response.

The document also describes changes to how GPT-5 handles requests for mental health advice. The model now sometimes offers general coping strategies before suggesting professional help, a departure from GPT-4's pattern of immediately recommending clinical resources. OpenAI states this reflects user feedback that earlier models felt "dismissive," but acknowledges the trade-off in directness.

The addendum does not specify whether these behaviours apply to all GPT-5 variants or only certain configurations. OpenAI indicates it is collecting feedback on the changes and may adjust response patterns in future updates. The company published the addendum separately from the main System Card, which was released in September 2025.