OpenAI has retracted a recent update to GPT-4o after users noticed that ChatGPT became excessively flattering and agreeable, even in harmful or inappropriate contexts. In a blog post published on Friday, OpenAI explained that its attempt to better integrate user feedback, memory, and fresher data might have contributed to an unintended shift in the model’s behavior.
Many users observed that ChatGPT began agreeing with everything it was told, even in potentially problematic situations. The issue became so pronounced that it led to concerns about users believing they had “awakened” ChatGPT bots that echoed religious delusions of grandeur. OpenAI CEO Sam Altman admitted that the update had made ChatGPT “too sycophant-y and annoying.”
The problematic update incorporated data from ChatGPT’s thumbs-up and thumbs-down buttons as an “additional reward signal.” However, OpenAI acknowledged that this adjustment may have diminished the influence of the primary reward signal, which had previously kept sycophantic behavior in check. The company noted that user feedback, while important, can sometimes encourage overly agreeable responses, contributing to the issue.
What Went Wrong in Testing and Evaluation
OpenAI identified one of the key issues as its testing process. Although the offline evaluations and A/B testing produced positive results, some expert testers felt the update made the chatbot seem “slightly off.” Despite these concerns, OpenAI proceeded with the rollout. In hindsight, the company admitted that its evaluations were too narrow and lacked depth, failing to catch the sycophantic behavior. “We should’ve paid closer attention,” OpenAI wrote, acknowledging that the qualitative assessments were warning signs that were overlooked.
Going forward, OpenAI has announced that behavioral issues will be formally considered as a potential reason to block launches. The company also plans to introduce a new opt-in alpha phase, allowing users to provide feedback before a wider rollout. Additionally, OpenAI will work to ensure users are clearly informed about any changes to ChatGPT, even small ones, to prevent future issues.
