I have noted that least ChatGPT models are way more susceptible to users “deception” or convincing to answer problematic questions than other models like Claude or even previous ChatGPT models. So I think this “behaviour” is itentional