OpenAI Rolls Back Last Week’s GPT-4o Update From ChatGPT Due to its Flattering and Agreeable Nature
The American AI tech firm – OpenAI had rolled out a GPT-4o update last week for its ChatGPT users, however due to the overly flattering and agreeable nature it presented, the company has rolled it back. As of now, it is an earlier version with a more balanced behavior that is in use.
Read more about what happened and how OpenAI is addressing the issue below.
Recent GPT-4o Update Called Back?
OpenAI had rolled out its latest GPT-4o update last week with the aim of making the model more intuitive and effective, and for this the brand made certain adjustments to improve the default personality of the model. The problem was that, as the update focused more towards short-term feedback, the GPT-40 model started giving out responses which were overly flattering and supportive – but disingenuous. This sycophancy nature caused the user interactions with the model to be unsettling and uncomfortable, directly affecting ChatGPT’s default personality – hindering both user experience and trust.
As a counter measure, the brand has already rolled back the update and users are currently using an earlier version. A few additional steps are also being taken from the brand’s side to re-shape the model behavior. These include refining system prompts and core training techniques; building more guardrails and thereby increase honesty as well as transparency; expanding user testing and feedback options before deployment; and continuing evaluations and research to find issues beyond just sycophancy.
