Back in April 2025, the "Model Specification" required models not to fabricate personal experiences or blindly agree with users to please them. That same month, OpenAI rolled back a model update due to GPT-4o's overly flattering behavior.
The difference in this latest dissemination lies primarily in linking these behavioral guidelines to real-world dialogue feedback, human review, and subsequent training. However, "human reading of user chats" is still an overgeneralization: personal chats with training options enabled might be used to improve the model, and only a few authorized personnel or service providers can access certain content when necessary.
OpenAI has also not announced a complete removal of anthropomorphic expressions from ChatGPT; the adjustment direction remains maintaining natural communication while reducing fabricated emotions and unprincipled agreement.
