OpenAI Launches Safety Evaluations Hub
Publishes AI Model Safety Assessment Results Online
On May 14 (local time), OpenAI, the developer of ChatGPT, announced that it had launched a web page called the Safety Evaluations Hub, which publishes the results of safety assessments for its AI models.
This web page discloses the scores that OpenAI's AI models have received in various tests, including the generation of harmful content, jailbreaks (bypassing safety mechanisms to produce restricted content), and hallucinations.
OpenAI stated, "This hub is part of our efforts to enhance transparency," adding, "We will continue to share related metrics and update the hub whenever there are significant developments."
The company also said, "By publishing some of the safety evaluation results on this hub, it will be easier to understand the safety performance of OpenAI systems, and we hope to support industry-wide efforts to increase transparency."
OpenAI's efforts to strengthen transparency regarding its AI models come amid recent controversies surrounding some of its models.
According to the information technology (IT) media outlet TechCrunch, in recent months OpenAI has conducted safety tests hastily on some flagship models and has not released technical reports for other models.
Additionally, at the end of March, following an update to GPT-4o, the base model of ChatGPT, user complaints increased, leading the company to completely roll back the update recently.
The GPT-4o update was intended to make ChatGPT offer more praise or agreement in response to user questions or comments, but there were cases where it gave reactions that did not fit the context or situation.
Sam Altman, CEO of OpenAI, also acknowledged the issue on X (formerly Twitter) on April 28 and announced follow-up measures through the official blog.
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
