OpenAI Rols Back Update that made chatgpt a sycophantic mess

Looking for good vibes

OpenAi, together with competitors such as Google and Anthropic, tries to build chatbots with which people want to chat. So, designing the apparent personality of the model to be positive and supportive is logical – people are less inclined to use an AI that comes out so hard or negative. In the absence of a better word, it is more and more about it vib.

When Google Gemini 2.5 unveiled, the team rowed about how the model was at the top of the LM Arena Leaderboard, with which people can choose between two different model outputs in a blind test. The models that people like end up at the top of the list, which suggests that they are more pleasant to use. Of course, people can like output for various reasons – perhaps one is technically more accurate, or the layout is easier to read. But in general people like models that make them feel good. The same applies to the internal model consent of OpenAi, it seems.

It is possible that this striving for good vibes models push to show more sycophantic behavior, which is a problem. Alex Albert van Anthropic has called this as a “toxic feedback job”. An AI-Chatbot that tells you that you are a world-class genius that sees the unseen, may not be harmful if you just brainstorm. However, the endless praise of the model can lead to people who use AI to plan business companies or, forbids heaven, to perform radical rates, to be fooled by thinking that they have come across something important. In reality, the model has become so sycophantic that it loves everything.

The constant pursuit of involvement has been a disadvantage for numerous products in the internet age, and it seems that generative AI is not immune. The GPT-4O update of OpenAI is proof of that, but hopefully this can serve as a reminder for the developers of Generative AI that good vibes are not that important.