OpenAI released "Strengthening ChatGPT's responses in sensitive conversations," announcing that it collaborated with over 170 clinically experienced mental health experts to update ChatGPT's default model to more reliably identify help-seeking signals, de-escalate conversations, and guide users to real-world support. According to measurements in the paper, responses with undesirable behavior in mental health-related areas decreased by approximately 65%–80%. The company also expanded its crisis hotline coverage, redirected sensitive conversations from other models to safer ones, and added gentle reminders to take a break during long conversations.
This update focuses on three scenarios: severe symptoms such as psychosis/mania, self-harm and suicide, and emotional dependence on AI. OpenAI also updated the Model Spec to clarify that models should avoid reinforcing unfounded beliefs, respect real interpersonal relationships, and pay more attention to indirect signs of self-harm and suicide. Going forward, in addition to the existing baseline for self-harm and suicide, "emotional dependence" and "non-suicidal psychological emergencies" will be included in the standardized baseline testing for future model releases.
Frequently Asked Questions
Q: Where exactly are these changes reflected?
A: Updated default model behavior, automatic redirection of sensitive conversations, wider crisis hotline links, and "break reminders" for long conversations.
Q: What priority scenarios are involved?
A: Acute symptoms such as psychosis/mania, risk of self-harm and suicide, and excessive emotional dependence on the model.
Q: How to quantify the effect?
A: Officials said that inappropriate responses in related areas have decreased by 65%-80%; and the reliability has remained at 95%+ in high-difficulty long-dialogue security assessments.
Q: Have the safety principles changed?
A: Make existing goals more explicit in the Model Spec, such as not affirming unfounded beliefs and paying attention to indirect signs of self-harm or suicide.
Q: How will the new model be evaluated in the future?
A: Add "emotional dependence" and "non-suicidal emergencies" to the baseline test as part of the release threshold along with the self-harm and suicide baseline.
