I’m not sure if this is just me but I have found ChatGPT has never said sorry to me and been able to correct behaviour, in fact it often seems to get worse. For example if it hallucinates an answer and I call it on this it will say sorry and then give another hallucination. If if it says sorry it got that wrong it will then keep being wrong. The only way to fix this is to just start over. Is this just me or has anyone else found this too? GPT4 is better but I am still a bit nervous when it says sorry.
Yes, that’s right. Normally that happens when you try to get informations the model doesn’t know anything about. In the following messages it still has the own answers stored which may effect it.
I didn’t try yet, but maybe you could try to give the 1 shot examples multiple times (or yes restart the session)?
Something like:
Hey, what you just told me is wrong. It should be like this: xyz
To make that even more clear I give you an example:
xyz
And to make it even more clear I will repeat that example two more times:
xyz
xyz
Like I said I did not try. But might work.
2 Likes
I can confirm similar experiences.
Once you get an “I’m sorry but…” the conversation is pretty much dead regarding that topic. You need to start over.
This is not necessarily information the model doesn’t know or won’t give you. On prompts that have gotten an “I’m sorry but…”, I have later gotten correct, full answer when the prompt is written a different way in a new chat.
It seems that once an “I’m sorry…” is in the conversation history, the entire exchange is shot.
1 Like