OpenAI explains why ChatGPT became too sycophantic
Openi has published a post -mortem About the recent Sycophanancy problems with the standard AI model Powering Chatgpt, GPT-4O problems that forced the company to reverse an update for the last week released last week.
In the weekend, after the GPT-4O model update, users on social media noted that Chatgpt started to respond in a validating and pleasant way. It soon became a meme. Users have placed screenshots from chatgpt that all kinds of problematic, dangerous decisions And idea.
In a message on X on Sunday, CEO Sam Altman recognized The problem and said that OpenAi would work on solutions ‘as quickly as possible’. Two days later, Altman announced The GPT-4O update was reversed and that OpenAI was working on “extra solutions” for the personality of the model.
According to OpenAiThe update, which was intended to feel the standard personality of the model “more intuitive and effective”, was too much informed by “Feedback in the short term” and “did not explain how the interactions of users with chatgpt evolve over time.”
We returned to Chatgpt’s GPT-4O update last week because it was overly flattering and pleasant. You now have access to an earlier version with more balanced behavior.
More about what happened, why it matters and how we focus on SycopHancy: https://t.co/lohou7i7DC
– OpenAI (@Openai) April 30, 2025
“As a result, crooked gpt -4o crooked to reactions that were overly supportive but unfair,” Openai wrote in a blog post. “Sycophantic interactions can be uncomfortable, disturbing and cause suffering. We are short shots and are working on it to get it right.”
OpenAI says it implements various solutions, including the refining of its core model training techniques and system prompts to explicitly send GPT-4O away from Sycophanancy. (System prompts are the first instructions that guide the umbrella behavior and tone of a model in interactions.) The company also builds more safety argrails to increase “ [the model’s] Honesty and transparency, “and continue to expand his evaluations to” help identify problems that go beyond sycofancy, “says it.
OpenAI also says it is experimenting with ways to have users give “real-time feedback” to “directly influence their interactions” with chatgpt and to choose from multiple chatgpt personalities.
‘[W]Exploring new ways to include wider, democratic feedback in the standard behavior of Chatgpt, “the company wrote in his blog post.” We also believe that users should have more control over how ChatGpt behaves and, to the extent that it is safe and feasible, make adjustments if they do not agree with the standard behavior. “
