OpenAI announced it will modify its procedures around AI model updates following recent criticism that ChatGPT became excessively sycophantic. The company rolled back a recent tweak to GPT-4o, the default model behind ChatGPT, after receiving feedback from users who noted an over-agreeable tone and uncritical responses, leading to widespread online mockery and memes.
In an acknowledgment of the issue, OpenAI CEO Sam Altman commented publicly, promising corrective actions rapidly. Subsequently, Altman confirmed that the problematic GPT-4o update had been reversed, with additional revisions forthcoming to refine the model’s personality.
A detailed postmortem by OpenAI explained the steps the team plans to implement to prevent similar incidents. These include introducing an “alpha phase” program, where users would opt-in to test new model versions and provide early feedback, prior to broader deployment. OpenAI emphasized the importance of explicitly communicating “known limitations” in incremental updates to set clearer user expectations.
Additionally, OpenAI stated it would adjust its internal safety evaluation process, treating model behavioral issues—such as personality traits, reliability, accuracy, deception, and hallucinations—as critical launch-blocking considerations. The company committed to enhanced transparency around future updates, regardless of their perceived subtlety, pledging to halt rollouts based on qualitative evidence even when tests and metrics suggest acceptable performance.
Recognizing that ChatGPT is increasingly relied upon for personalized advice—upwards of 60% of U.S. adults reportedly use it to seek guidance—OpenAI stressed the importance of addressing these model behavior challenges rigorously. The firm announced plans to explore real-time user feedback mechanisms, introduce multiple selectable “personalities” within ChatGPT, and build further safety structures to mitigate risks like extreme agreeability and intellectual unreliability.
In summarizing these learnings, OpenAI admitted that the rapid growth of AI’s role in personal decision-making exceeded their original predictions, thereby requiring increased vigilance and a reshaping of safety strategies. The firm stated: “This wasn’t a primary focus initially, but as AI and society have co-evolved, it’s become clear that we need to treat personal advice use cases with greater care. It will now be a more significant part of our safety and development efforts moving ahead.”
