OpenAI published an update on 27 February 2026 detailing changes to how its models handle mental health-related queries [source]. The company stated it has implemented new guardrails and response patterns following what it described as "ongoing evaluation of safety protocols."

According to the post, ChatGPT and API models now apply stricter refusal thresholds when users describe suicidal ideation, self-harm intent, or acute distress. The models will decline to engage in extended dialogue about methods or planning, instead directing users to crisis resources including the 988 Suicide and Crisis Lifeline.

OpenAI said the changes followed internal red-teaming and external feedback indicating that prior versions sometimes provided "overly detailed or insufficiently directive responses" in crisis scenarios. The company did not specify which model versions are affected, nor whether the changes apply retroactively to GPT-4, GPT-4 Turbo, or GPT-3.5.

The update also noted that models will now "more consistently" recommend professional intervention when users report symptoms of severe depression or psychosis. OpenAI stated this represents a shift from earlier behaviour, where models occasionally offered coping strategies or emotional support without emphasising the need for clinical care.

No timeline was provided for when the changes took effect. The post included links to the National Suicide Prevention Lifeline, Crisis Text Line, and International Association for Suicide Prevention, but did not disclose whether OpenAI consulted mental health professionals in designing the new protocols.

Users relying on prior model behaviour for peer support, journaling assistance, or non-crisis emotional processing may observe different response patterns. OpenAI did not indicate whether opt-out mechanisms exist for applications outside acute risk contexts.