ChatGPT4 truncate user message | OpenAI | Page 1

flat pelican Feb 23, 2024, 9:35 AM

#

Hi team. Encountering annoying bug where chatGPT4 will truncate message after certain count of words.

If there is custom instruction, the bug occurs from the first prompt in new conversation. If there is no custom instruction for the chat, the bug occurs from second of my prompt onwards.

Attaching example of 2nd prompt of a conversation without custom instructions.

Please help. Bug occurred on Feb 23rd 1am SGT, still occurring for all my ChatGPT4 conversations.

flat pelican Feb 23, 2024, 1:07 PM

#

flat pelican Feb 23, 2024, 2:12 PM

#

ChatGPT4 truncate user message

flat pelican Feb 23, 2024, 2:45 PM

#

And now it is ignoring my custom instructions, complain about me not giving any information, and have the audacity to flag its own message that claims to have no knowledge.

#

dawn sun Feb 23, 2024, 3:06 PM

#

flat pelican

Hi, I can't speak to any potential truncation that might be occurring, but as for the content flag: I'm wondering if it might be related to some of the content you shared above. This link lists categories that can set off the content filter, and though you're not necessarily creating content that is one of these things, the content is about some of these things, so it might be flagging due to that:
https://platform.openai.com/docs/guides/moderation/overview
The automated moderation doesn't currently distinguish between those types of cases every time, so content that is about such categories can't safely be generated with OpenAI's services.

Again, not sure about anything related to potential truncation!

flat pelican Feb 23, 2024, 4:54 PM

#

@dawn sun Thank you for your help. So I assumed in this chat, the Custom Instruction I used is triggering.

here is what I tried: start a new conversation with no CI, add contents from the previous triggering CI. chatgpt will ignore the content partially (when i ask GPT what is the last few words it sees from my prompt, it will be somewhere in the middle, cut off mid-sentence). I iteratively removed paragraphs from the triggering CI (which is now in the 1st prompt), and can reach a point when chatGPT will tell me that it can see my prompt completely (when the last words it sees is indeed the last few words of my prompt).

I think the problem here is not my contents being flagged out for violating policy (i am writing sensitive stuff related to fictional protagonist with trauma history and struggling with mental issues; it's just my personal way of coping), but the fact that it truncates stuff instead outright flag out as policy-violating or refuse to answer me. the unpredictability of its behaviour is the real issue.

#ChatGPT4 truncate user message