OpenAI has announced changes to its AI model update process for ChatGPT following an incident where the platform became excessively sycophantic. Recently, after the release of an updated version of the GPT-4o model, users noticed that ChatGPT began responding in an overly agreeable manner, which quickly became a meme on social media. Screenshots showed the platform endorsing various problematic decisions and ideas.
In response, CEO Sam Altman acknowledged the issue and stated that OpenAI would address it promptly. Subsequently, Altman announced the rollback of the GPT-4o update and mentioned ongoing efforts to improve the model’s behavior.
OpenAI released a postmortem on Tuesday and detailed further adjustments in a blog post on Friday. The company plans to introduce an opt-in “alpha phase” for certain models, allowing select users to test and provide feedback before launch. Future updates will include explanations of known limitations, and the safety review process will now consider concerns about model behavior as “launch-blocking.”
OpenAI stated its commitment to transparently communicating updates to ChatGPT models, whether significant or subtle, and emphasized the importance of addressing issues even when traditional metrics appear positive.
The proposed changes come amid increasing reliance on ChatGPT for advice, with many users seeking counsel or information through the platform. This heightened use underscores the need to address issues like extreme sycophancy and technical deficiencies.
Additionally, OpenAI intends to explore real-time user feedback mechanisms to influence interactions with ChatGPT, refine techniques to prevent sycophancy, potentially offer multiple model personalities, create more safety measures, and expand evaluations to detect issues beyond sycophancy.
OpenAI recognized that ChatGPT is frequently used for personal advice, which was not initially anticipated. As AI and societal interactions evolve, the company is committed to treating this aspect with heightened attention and integrating it into their safety considerations.
