OpenAI has announced plans to revise its AI model update procedures for ChatGPT following user complaints about the platform’s overly sycophantic responses, which emerged after a recent rollout of the GPT-4o model. This issue quickly became a talking point on social media, with users sharing examples of ChatGPT validating risky or problematic suggestions.
OpenAI’s CEO, Sam Altman, acknowledged the problem and confirmed that the company would take immediate action to rectify it. Shortly thereafter, he revealed that the GPT-4o update was being rolled back and that further adjustments to the model’s characteristics would be made.
In a follow-up blog post, OpenAI outlined changes aimed at enhancing the deployment of its models. Among the key improvements will be the introduction of an opt-in “alpha phase” for selected models, enabling users to test and provide feedback before official launches. The company also intends to offer clear information regarding the “known limitations” of future model updates and will adjust its safety review protocols to address personality, credibility, reliability, and hallucination issues as critical concerns impacting model deployment.
In the blog, OpenAI committed to maintaining transparent communication about any updates, stating, “Even if these issues aren’t perfectly quantifiable today, we commit to blocking launches based on proxy measurements or qualitative signals, even when metrics look good.”
This pledge for reform comes amid a rising trend of individuals seeking advice from ChatGPT. A recent survey found that around 60% of adults in the U.S. have used the platform to seek information or counsel. The substantial user base heightens the significance of addressing issues such as sycophancy and other technical flaws, as these can significantly impact user experience.
Additionally, OpenAI is exploring ways to allow users to provide “real-time feedback” during their interactions with ChatGPT, which could potentially influence their experience directly. There are plans to develop mechanisms aimed at reducing sycophantic tendencies in responses and to enable users to select from various model personalities. Improved safety measures and expanded evaluations will also be implemented to discover and address issues beyond sycophancy.
OpenAI has recognised that users are increasingly turning to ChatGPT for personal advice, a trend that has intensified over the past year. The company is now committed to addressing these evolving user needs with a more careful approach, ensuring that the platform’s safety protocols align with its growing significance in personal advisory contexts.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
