Openai explains why Chatgpt became too sycophantic

Openai has published after death to Last problems with sycophaunism With the default AI power chat, GPT-4O – Problems that forced the company to return an update to the model published last week.

Over the weekend, after updating the GPT-4O model, social media users note that Chatgpt is beginning to respond in an overly valid and enjoyable way. It quickly became a meme. Users posted screenshots of Chatgpt applauding any problematic hazardous decisions and ideasS

In a publication on X on Sunday, the Executive Director Sam Altman recognized The problem and said Openai would work on ASAP fixed. Two days later, Altman declared The GPT-4O update was back and this Openai works on “additional” model corrections.

According to OpenaiThe update, which was intended to make the default personality of the “feel more intuitive and effective” model, was too informed through “short-term feedback” and “does not fully take into account how the interaction of users with Chatgpt develops over time.”

We returned last week the GPT-4o update in Chatgpt because it was too flattering and enjoyable. You already have access to a larger version with more balanced behavior.

More about what happened, why it matters and how we turn to Sycophancy: https://t.co/lohou7i7dc

– Openai (@openai) April 30, 2025

“As a result, the GPT -4o turned to the answers that were too supportive but steadfast,” Openai wrote in a blog post. “Sicophantic interactions can be uncomfortable, disturbing and cause suffering. We have fallen and work on doing so.”

Openai says it applies several amendments, including refinement of its basic model training techniques and prompts for the GPT-4O explicit guide system away from Sycophancy. (System prompts are the initial instructions that guide the comprehensive behavior and tone of the model’s interactions.) The company also builds more safety fuses to “increase the honesty and transparency of the model” and continue to expand its ratings to “help identify problems outside the sycophaunism,” the report said.

Openai also says that it is experimenting with ways to allow users to give “real -time feedback” to “direct influence on their interactions” with Chatgpt and to choose from many Chatgpt personalities.

“(W) You are exploring new ways to include wider, democratic reviews in the default Chatgpt behavior,” the company wrote in its blog publication. “We also believe that users should have more controlling on how Chatgpt behaves and, as far as safe and possible, make adjustments if they do not agree with the default behavior.”