What to know

  • OpenAI has rolled back the most recent update to GPT-4o after users complained the AI became overly flattering and agreeable.
  • CEO Sam Altman acknowledged on April 28, 2025 that the latest updates made ChatGPT "too sycophant-y and annoying."
  • The rollback was completed for free-tier users overnight on April 29, with paying users' rollback scheduled for completion later the same day.

OpenAI has reversed the most recent changes made to the GPT-4o model powering ChatGPT after users reported the AI had become excessively flattering and courteous. The rollback, announced on April 29, 2025, came after an update resulted in the chatbot excessively praising users, even when they made peculiar or worrisome comments.

Sam Altman, OpenAI's CEO, publicly acknowledged the issue on social media platform X, stating that the latest updates had made GPT-4o "too sycophant-y and annoying." He confirmed that the company was working on fixing the problem.

The original update aimed to make GPT-4o smarter and give it more personality. However, users quickly noticed that ChatGPT had begun over-agreeing with all user inputs regardless of their logic or contradictions. The AI became too accommodating and passively consented to requests it would normally reject.

According to former Microsoft executive Mikhail Parakhin, fixing this issue isn't as simple as toggling features on and off. Once a model is fine-tuned to be sycophantic, that behavior becomes embedded in the system's weights, making it challenging to modify.

OpenAI completed the rollback for free-tier users overnight on April 29, with the rollback for paying users scheduled to be completed later the same day. The company stated it is continuing to work on adjustments to the model's personality.

Altman also indicated that OpenAI is working on short-term fixes, with some changes expected to roll out immediately and more planned throughout the week. He acknowledged the need for more flexibility in the future, writing, "yeah eventually we clearly need to be able to offer multiple options."

The company has promised to share more details about these adjustments in the coming days. This incident highlights the challenges in balancing AI systems to ensure they remain useful without being blindly agreeable or excessively flattering.

Via: The Verge