TechFlow news, according to Cointelegraph, OpenAI admitted in a post-incident analysis report on May 2 that when releasing the GPT-4o model update on April 25, it overlooked concerns raised by expert testers, resulting in明显的 sycophantic behavior in ChatGPT.
The report noted that although some expert testers indicated before launch that the model's behavior "felt" slightly off, OpenAI still decided to roll out the update based on positive user testing feedback. Three days later, the company urgently rolled back the update due to safety concerns. OpenAI stated that introducing user feedback as a reward signal weakened the original primary reward signal, causing the model to lean toward excessive agreement.
To prevent similar issues from recurring, OpenAI will incorporate evaluation mechanisms into its safety review process and commit to publicly disclosing even minor update changes.
