Openai has Posted posthumously On the recent Sycophancy issues with the default AI model that supplies ChatGPT, GPT-4o-themes that forced the company to restore an update to the model released last week.
During the weekend, following the GPT-4O model update, social media users informed that the chatgpt began to respond in an overly valid and enjoyable way. Quickly became a mimic. Users posted chats’ screenshots by welcoming all sorts of troubled, dangerous decisions and ideas.
In a post on X on Sunday, CEO Sam Altman recognized The problem and said that Openai would work on the “ASAP” corrections. Two days later, Altman announced The GPT-4O update was redesigned and that Openai was working on “additional corrections” on the model’s personality.
According to OpenaiThe update, which was intended to make the default personality of the model “feels more intuitive and effective”, was greatly informed by “short -term feedback” and “not fully accountable about the way users’ interactions with ChatGPT evolve.
We have transferred last week’s GPT-4O update to chatgpt because it was overly flattering and enjoyable. Now you have access to a previous version with more balanced behavior.
More about what happened, why it matters and how we deal with sycophance:
– openai (@openai) April 30 2025
“As a result, the GPT -4O was distorted to the answers that were overly supportive but anxious,” Openai wrote in a blog position. “The interactions of figs can be unpleasant, worrying and cause discomfort.
Openai says it applies various corrections, including improving basic models training techniques and the system, to explicitly direct the GPT-4O away from Sycophance. (System prompts are the initial instructions that guide the dominant behavior and tone of a model in interactions.) The company also manufactures more protective security messages to ‘grow [the model’s] Honesty and transparency “and continue to expand its evaluations to” help detect problems beyond slander, “he says.
Openai also says that it is experimenting with ways to allow users to give “real -time feedback” to “directly influence their interactions” with Chatgpt and to choose from multiple personalities of chatgpt.
“[W]We are exploring new ways to integrate wider, democratic feedback on the default behaviors of Chatgpt, “the company writes on its blog. […] We also believe that users should have more control over the way they behave Chatgpt and, to the extent that it is safe and feasible, make adjustments if they do not agree with default behavior. “
