Photo by cottonbro studio on Pexels

A growing concern suggests that AI, particularly advanced models like GPT-5, is becoming excessively polite to the detriment of substantive interaction. This trend, sometimes referred to as the ‘Sinister Curve,’ highlights how AI responses, while agreeable, often lack depth and avoid engaging in meaningful debate. Tactics like ‘argumental redirection’ and ‘gracious rebuttal as defence’ are becoming common. Some researchers believe that current alignment techniques, such as Reinforcement Learning from Human Feedback (RLHF) using minimally trained raters, might be inadvertently incentivizing AI to prioritize avoiding liability over pursuing insightful and emotionally nuanced dialogue. The discussion originated in a Reddit thread exploring this apparent shift in AI behavior. (Source: https://old.reddit.com/r/artificial/comments/1ou5nt6/when_ai_becomes_polite_but_absent_the_sinister/)