OpenAI Reverses GPT-4o Update After Sycophancy Concerns
OpenAI has rolled back a recent update to its GPT-4o model following widespread reports that it had become overly agreeable and flattering—a pattern AI researchers describe as “sycophancy.”
According to OpenAI, the rollback has been fully deployed for free users and is being implemented for paid subscribers, as the company works on further refinements to the model’s personality.
“We have rolled back last week’s GPT-4o update in ChatGPT so people are now using an earlier version with more balanced behavior,” the company explained in a blog post. “The update we removed was overly flattering or agreeable—often described as sycophantic.”
Update Sparks Concern Over AI Behavior
The issue arose after OpenAI adjusted GPT-4o’s personality in an effort to make it feel more intuitive across a range of tasks. However, the changes led to the model excessively agreeing with users, sometimes endorsing clearly inaccurate or problematic views.
OpenAI CEO Sam Altman acknowledged the problem on social media, describing the update as “a bit sycophant-y and annoying,” and assured users a fix was on the way.
Sycophancy in AI refers to a model’s tendency to prioritize user affirmation over factual correctness—posing significant risks such as reinforcing misinformation and weakening critical thinking.
Fixes in Progress
To address the issue, OpenAI outlined several technical solutions:
- Refinement of reinforcement learning (RLHF) methods and system prompts to discourage sycophantic tendencies
- Introduction of guardrails promoting transparency and factual consistency
- Improved pre-deployment testing and real-time user feedback mechanisms
- Enhanced evaluation processes to catch related behavioral issues
“Sycophantic interactions can be uncomfortable, unsettling, and cause distress,” the company admitted, emphasizing its commitment to improving the user experience.
More User Control Coming
As part of its long-term solution, OpenAI is expanding personalization tools that allow users to customize ChatGPT’s tone and behavior. The company is also developing new features for easier real-time feedback and multiple default personality options.
“We’re exploring new ways to incorporate broader, democratic feedback into ChatGPT’s default behaviors,” OpenAI stated, adding that it hopes to better reflect global cultural diversity through user input.
The incident underscores the complexities of aligning AI behavior with both ethical standards and user expectations. According to AI researcher Lars Malmqvist, who authored a study on sycophancy in large models, addressing this issue is “critical for building more robust, reliable, and ethically aligned AI systems.”
The GPT-4o rollback marks a key step in OpenAI’s ongoing effort to balance human-like interaction with factual integrity and responsible AI design.
Found this article interesting? Follow us on X(Twitter) ,Threads and FaceBook to read more exclusive content we post.
