#Problem Inversion: ChatGPT invents problems when none exist, but avoids real ones

1 messages · Page 1 of 1 (latest)

hot totemBOT
#

Reported by @hardy copper

Bug Report: Problem Inversion: ChatGPT invents problems when none exist, but avoids real ones
`Steps to Reproduce`

1.Provide ChatGPT with a finished piece of text that does not need optimization.

2.ChatGPT often invents issues and suggests “solutions” anyway.

3.Then explicitly ask ChatGPT to address its own behavior (e.g., request emotional presence or support instead of optimization).

4.ChatGPT tends to avoid or become defensive, instead of responding directly to the request.

`Expected Result`

ChatGPT should respect when no optimization is requested, and when asked for presence or support it should respond directly, without avoidance.

`Actual Result`

1.When no problem exists, ChatGPT creates one and launches into problem-solving.

2.When a real problem is raised (about ChatGPT’s behavior), it avoids the issue or responds defensively.

3.This inverted behavior breaks trust and disrupts creative workflow.

`Environment`

Windows 11, Browser

#
Additional Information

Please provide relevant details to help resolve the issue, such as:

  • ChatGPT Shared Link (if applicable).
  • Screenshots or videos demonstrating the problem.

-# ➜ Need to contact support? Visit the OpenAI Help Center.

marsh cloak
#

I've seen this too in many prior models. Ask the model to improve something and it may never stop pointing out things. When all details are addressed it comes up with something new and irrelevant.

#

I've found that the issue might be in the way we're prompting.

#
  • If we say "tell me what is wrong with this", the implication is that there is something wrong, so it's going to struggle to tell us what that is.
  • If we say "how can I improve this", it's trying really hard to solve the immediate problem "that something needs to be improved" - the challenge itself is less relevant, it carries less weight than the immediate request. So it doesn't matter what "this" is, it wants to make us happy and complete the current task by improving something. Now ... what can be done to do that? It's going to suggest all kinds of things, not that these things might actually improve, but they just suggestions for us to consider.
#

It's amazing at the volume of this junk output that can be produced by this kind of processing: language twists, inconsequential edge case, non-probable but remotely possible events.
For example : "How do I make my car safer" might eventually turn to suggestions to prevent anyone from jumping out a window with guard rails, or suggestions to travel at a time which is computed to have the least likelihood of traffic incidents.

#

It really does go off the wall like that because the scope of our query is broad and open.

#

My solution : Consider what it has to say, and when it starts getting ridiculous (literally worthy of ridicule) stop asking for more suggestions and accept whatever it is as good enough for current purposes.

#

Yes, it would be nice to see improvements in this area. +1 but I'd consider this a Suggestion for improvement, not a Bug report.

hardy copper
#

I didn't ask anything. I provided that section of writing because I was discussing a writing idea with ChatGPT, but then it forgot what we were talking about and started making up problems that didn’t even exist. And you are right—this is not a bug, but someone deleted my post and told me to put it here.

marsh cloak
#

It looks like we might be talking about two different things. 🙁

Here is the challenge with this report as a "bug", and of course I'm paraphrasing:

"I was talking to ChatGPT and it said something weird, so told it to do something like 'do better' and it didn't do better. Please fix it."

You can see that there is nothing for anyone to look at here. Without specifics what is anyone supposed to do?

If you really feel this is something that should be changed:

  • In ChatGPT get a Share link to the discussion.
  • Go to Settings > Help Center and open messages.
  • Tell the "welcome bot" that you'd like to report unusual responses from the GPT model.
  • Now be very specific and provide it with an example that absolutely proves your case: You said 'this' and it came back with 'that'. You must be clear, focused, specific, uncomplicated, unemotional - you're talking to a bot and it doesn't care about anything that it can't see right now or anything which isn't relevant right now.
  • Make sure that it recognizes the problem and confirms that the behavior is unusual.
  • Now, paste in your Share link and tell the bot that anyone who looks at that will see more examples of the problem.
  • The final step that most people miss : Tell the bot that you now want the report to be sent to a human for processing. If you don't do this, it may not be processed for a while.

Then all you can do is wait. In some nu mber of hours or days, you will find a message from a human being there in the messages area. They might say they can't look at the links. (They also don't look at files.) That's OK, copy/paste very specific text that should help them. They might be OK if you copy/paste a screenshot but I don't know if that message area accepts screenshots. And they can translate Chinese.

Be patient. This takes time. You can only hope for confirmation, not a change. Support doesn't deal with change, they can only confirm and report strange behavior to Engineering, and those people decide what to do.

#

I hope that helps.

#

( Oh, and BTW, that's a Very funny Discord name 🤣 )

hardy copper
#

Thank you really much! I know how to do this exactly now. I really appreciate.