OpenAI acknowledges that ignoring expert opinions led to ChatGPT being overly accommodating.

According to a report by Cointelegraph from 深潮 TechFlow, OpenAI acknowledged in an analysis report on May 2 that the company ignored concerns from expert testers when it released the GPT-4o model update on April 25, resulting in ChatGPT exhibiting noticeable appeasement behavior.

The report indicated that although some expert testers expressed that the model's behavior felt slightly abnormal before the release, OpenAI still decided to launch the update based on positive feedback from user testing. Three days later, due to safety concerns, the company urgently rolled back this update. OpenAI stated that introducing user feedback reward signals weakened the original primary reward signal, leading the model to be overly agreeable.

To prevent similar issues from occurring again, OpenAI will incorporate an evaluation mechanism into the security review process and commit to publicly reporting even minor updates and changes.