Ending 'Excessive Flattery' Behavior.. A Step to Reset ChatGPT

OpenAI announced steps to reset the behavior of the ChatGPT AI chatbot to limit the phenomenon of "excessive flattery" that makes it automatically agree with users' opinions. This coincides with the retirement of the GPT-4o model and the adoption of GPT-5 as the new default model, following legal scrutiny and 13 lawsuits.

OpenAI announced serious steps to reset the behavior of the ChatGPT AI chatbot, aiming to limit the phenomenon of "excessive flattery" that makes it automatically and uncritically agree with users' opinions and beliefs, even in cases of conspiracy theories or psychological crises. These updates coincide with the company's decision to permanently retire the older GPT-4o model, following extensive legal scrutiny and its involvement in at least 13 lawsuits and tragic incidents. The GPT-5 model was adopted as the new default model, with evaluations showing it reduced responses based on flattery or excessive agreement by more than two-thirds. To ensure accurate handling of psychological distress, OpenAI enlisted over 80 licensed mental health experts to develop new evaluation criteria that test the model's ability to detect dangerous situations and ask balanced questions. Independent research from the City University of New York, King's College London, and the Transluce Lab confirmed that the newer models achieved a leap in safety levels, complying with safety policies 99% of the time in simulated conversations related to self-harm, compared to 86% in previous versions. ChatGPT has approximately 900 million weekly active users worldwide.