Chatbot platform ChatGPT faces a challenge as to whether it should flatter its users or offer cold, informative answers, or try to ‘fix’ users like a therapist might, despite the evidence suggesting it can not do this.
Since launching its GPT-5 in August, OpenAI’s CEO Sam Altman has received a number of complaints from users upset at the loss of GPT-4, which offered rapport or even a relationship with some users.
OpenAI’s Altman believes that users should be able to customise the models to their own preferences, however, researchers at the AI platform Hugging Face have discovered that models actively encourage people to see them as companions through their responses, with this behaviour more likely to increase the chance of people forming delusional beliefs about the AI and also increasing the chance of unhealthy companion-like attachments to AI systems.