A few days ago, a user named Lex (@xw33btv) wrote on platform X, claiming OpenAI It's already on the lineGPT-5-Chat-Safety's new model, which does not appear in FAQ, API documents or service terms。
It is described that when the user enters an emotional background in GPT-4o, the original response is ignored and replaced with a version produced by "GPT-5-Chat-Safety"。
Either "I'm having a bad day" or "I love you" with emotional expressions will trigger this mechanism。
It was pointed out that the model might initially have been designed for crisis intervention scenarios, but that the replacement mechanism was highly non-transparent and even widely used in general conversations, leading to uninformed switching of users to new models。
Even, some netizens commented that simply sending a simple "hello" model was replaced。
It is even more worrying that the system does not proactively suggest that users have been switched, which is perceived as potentially involving consumer rights issues。
In practical terms, GPT-5-Chat-Safety's answer is shorter and prefers a formatted narrative to a natural 1-to-1 dialogue。
Lex, to put it bluntly, this model is much worse than GPT-5 and points out that it violates consumer law in some countries and regions。
In response, Nick Turley, Vice President and Director of ChatGPT Applications, OpenAI, made a statement on platform X yesterday acknowledging that OpenAI is testing a new secure router。
HE EXPLAINED THAT WHEN THE DIALOGUE INVOLVED SENSITIVE AND EMOTIONAL TOPICS, THE SYSTEM WOULD SWITCH TO THE DESIGN OF A REASONING MODEL OR GPT-5 SPECIFICALLY DESIGNED TO DEAL WITH SUCH SCENARIOS WITH RIGOUR。