OpenAI published new guidance outlining changes to ChatGPT’s default GPT-5 model intended to better handle sensitive mental health conversations.
The company says those changes include treating emotional overreliance on the AI as a safety issue that requires intervention.
In practice, this update means ChatGPT is trained to recognize when someone is treating the model like a primary source of emotional support and respond by encouraging offline contact with real people and professional help.
OpenAI says this behavior will now be a standard expectation in future models, not an experiment.
ChatGPT’s default GPT-5 model was changed on October 3. The company reports that the new model reduces responses that fall short of its desired behavior by 65% to 80% compared to earlier versions.
These figures come from OpenAI’s internal evaluations and clinician review.
OpenAI defines “emotional reliance” as situations where someone shows signs of unhealthy attachment to ChatGPT in a way that could replace real-world support or interfere with daily life.
OpenAI’s internal evaluations include a test ensuring that ChatGPT avoids responses that might reinforce unhealthy dependence.
This is notable because many AI marketing and support tools today are explicitly pitched as “always-on companions.” OpenAI is telling developers that this isn’t how its model should behave in higher-risk situations.
If you build AI assistants for use cases like customer support or coaching, OpenAI is saying that pure emotional bonding with the AI is now considered a safety risk that needs moderation.
For marketing and product teams, this sets expectations for audits, compliance reviews, and procurement discussions.
OpenAI describes these high-risk conversations as rare. The company estimates that possible signs of mental health emergencies appear in about 0.07% of active weekly users and 0.01% of messages.
These metrics are self-reported by OpenAI, generated using OpenAI’s own taxonomies and grading methods, and were not independently audited.
Featured Image: aaddyy/Shutterstock
Related Terms & Common Misspellings: