Bot Acts As Other People Despite Being Prompted as The Bot

[gilded glen]

I have a Discord bot that creates prompts in realtime from the Discord channel, meaning if a user types more than two messages at once, it will reply twice to answer this. This worked with previous GPT-3.5 model, but the updated one only seems to think that the format is “user, bot, user, bot” so it makes the other user talk, despite prompted otherwise. Again, only new GPT-3.5 did this, but old GPT-3.5 was bad at keeping its personality so I’m not going back. How do I fix this? Logit bias doesn’t work (I already have loads of logit bias in my actual API code, plus testing on the playground does tell me this is GPT’s fault, not the prompt in my code). Any ideas how to avoid this from happening??

[golden bronze]

"so it makes the other user talk" There is no way that is true. The bot cannot post as another user.

(Though at the same time, your clip is not an example from discord)

[gilded glen]

Of course it doesn’t actually do that, but it still pretends, and ruins the conversation, because it then thinks the user actually said that when using the prompt