- OpenAI has pulled back the latest update to GPT-4o that made ChatGPT too sycophantic and a people pleaser.
- Now, OpenAI is addressing the issue of sycophancy and realigning ChatGPT's default behavior to set things right.
- OpenAI says focusing on "short-term feedback" led to ChatGPT's disingenuous behavior.
Last week, OpenAI rolled out a controversial update to ChatGPT. OpenAI CEO Sam Altman said the new GPT-4o update to ChatGPT should improve “both intelligence and personality.” However, just after the update, ChatGPT users started complaining about sycophancy and that ChatGPT has become too eager to please.
The AI community quickly noticed the change in ChatGPT’s personality and said the behavior was overly flattering and agreeable. Some users said that even after changing custom instructions, ChatGPT remained sycophantic and agreed to almost anything.
Now, after four days, OpenAI has pulled back the GPT-4o update on ChatGPT. Altman has shared on X that the latest update is 100% rolled back for free ChatGPT users, while the process is ongoing for paid users.
In addition, OpenAI in a blog post has explained what actually happened with the new GPT-4o update and how the company is addressing the issue. OpenAI notes, “we made adjustments aimed at improving the model’s default personality to make it feel more intuitive and effective across a variety of tasks.”
Further, the company says, “However, in this update, we focused too much on short-term feedback, and did not fully account for how users’ interactions with ChatGPT evolve over time. As a result, GPT‑4o skewed towards responses that were overly supportive but disingenuous.“
OpenAI admitted that ChatGPT’s default personality became too sycophantic and said, “We fell short and are working on getting it right.” As demonstrated in the post below, a Facebook user used the new ChatGPT update to reinforce his unfounded beliefs and received validation from ChatGPT, backing flat Earth theory.
Now, to address the issue, OpenAI is taking more steps to realign ChatGPT’s behavior. OpenAI is “refining core training techniques and system prompts to explicitly steer the model away from sycophancy.” In addition, the company is developing more guardrails to improve honesty and transparency. Finally, OpenAI is expanding its evaluation to identify issues beyond sycophancy, going forward.
Since ChatGPT is used by over 500 million people every week, OpenAI has a huge responsibility to get ChatGPT’s default behavior right. Now, we need to evaluate how ChatGPT responds to user queries, going forward.
