ChatGPT changes model for sensitive topics
OpenAI has introduced a new security measure for ChatGPT. A special AI model is designed to handle sensitive topics.
(Image: CHUAN CHUAN/Shutterstock.com)
When it comes to sensitive or emotional topics, ChatGPT will in the future be able to handle these using specially customized AI models. This means that the model can change during a conversation without the human noticing.
Nick Turley writes about this security measure at X. He is Head of ChatGPT at OpenAI. According to him, ChatGPT either switches to a reasoning model for sensitive topics or switches to a version of GPT-5, “which was developed for particularly careful handling of these contexts.” The decision as to which model answers is apparently made similarly to when the reasoning model is involved in more in-depth questions. There must therefore be some kind of internal orchestration. “Our aim is to always provide answers that match our model specification.” These model specs are defined guidelines for models.
Videos by heise
OpenAI had already announced in a blog post that such forwarding would be introduced. It was stated that in the event of “acute stress,” queries would be answered by a separate model. It is not entirely clear how sensitive these “sensitive or emotional topics” have to be to trigger a model change.
OpenAI decides on emotional conversations
There are lawsuits against OpenAI in the USA, as ChatGPT is said to have assisted a teenager in committing suicide. Specifically, the chatbot is said to have offered help with writing a suicide note, among other things. The new security function should make such a reaction impossible.
Nevertheless, it can be observed how many people have supposed relationships with chatbots. This usually includes emotional conversations. When OpenAI recently updated the chatbot to the GPT-5 model, users complained that the chatbot was behaving less emotionally. Some spoke of having lost their boyfriend or girlfriend. OpenAI then made it possible to continue using the predecessor model, GPT-4o. The question now is what boundaries OpenAI wants to draw for emotional topics and close relationships with a chatbot.
(emw)
