ivory cedar
This could be replicated by sharing some info about yourself as a user, deleting the user message or rewriting over it. And then asking what info does gpt know about the user. I have no simple replication to share, but I have some screenshots of the observed behavior.
This makes prompt engeneering in the playground completely unreliable and should be fixed asap. For example, the correct functioning of a prompt could be due to the trail and error still stored in the background instead of a correct fine-tuning.
I have reported this to openAI twice and haven't had any feedback, and laughed at in the bugbounty as a non-important issue. However, it's a heavy issue potentially making the use of the playground useless.