OpenAI explains why ChatGPT became too sycophantic

Openai has Postmortem was published on Last typofancia problems With the default AI model Powering Chatgpt, GPT-4O – Issues that make an update again in a model broadcast last week.

After the weekend, after the GPP-4o model update, users in social media, Chatgptin began to respond to an extremely confirmed and satisfied way. It was a quick memorial. Users published ChatGPT screenshots, welcome all kinds of problems, dangerous decision and thought.

Sunday in an article in X, CEO Sam Altman recognizes Problem and Openai ‘ASAP’ will work on the idea. Two days later, Altman declared The GPT-4O update was withdrew and the identity of the Openai model was “additional adjustments”.

According to OpenaiThe standard identity of the model is provided for “Short-term feedback” to make the “more intuitive and effective”, and “short-term opinion” and did not think that the users were established with ChatGPT. “

We returned the last week’s GPT-4O update again as flattering and satisfied in Chatgpt. Now you have access to a previous version with a more balanced behavior.

Why are it important why Sycofany is important because of what happens: https://t.co/lohou7i7dc

– Openai (oopenai) April 30, 2025

“As a result, GPT-4O, overwhelmed, but flawed, but flawless answers,” Openai wrote in the blog post. “Sycofantic interactions can cause anxiety, disturbing and distress. We are short and try to get it properly.”

Openai says that it is implementing several adjustments, including the main model, which is removed from the dismantle of the system, applies the steering wheel signals behind the wheel. (System proposals are initial instructions that direct the basic behavior and tone of a model. [the model’s] Honesty and transparency “and the assessments continue to help identify the problems on the edge of the typofania,” he says.

Openai is also practicing with ways of issuing “real-time feedback” with the ways of the users “Real-Time feedback”.