Experiencing recent malfunctions with ChatGPT-4 and is seeking assistance or explanations

For the past six months, I’ve been using ChatGPT-4 as an integral assistant for my work, particularly in story development. This has involved tasks such as proofreading, obtaining thoughts and opinions, discussing themes, character arcs, and so forth. Up until now, it’s been an invaluable tool and the best money I spend each month.

However, starting yesterday, I’ve encountered some significant problems:

  1. Deviation from Prompts: Even when I use prompts that I’ve copy-pasted from previous successful interactions, ChatGPT-4’s responses are now well off the mark.
  2. Short-term Memory Issues: The system seems unable to remember details from just a message prior, which is rather problematic.
  3. False Flags: An increasing number of my inputs are being flagged as potentially harmful, which hasn’t been the case before.

Troubleshooting steps I’ve taken include logging out and back in, but the issue persists. Interestingly, when I switch to ChatGPT-3, these problems vanish.

Has there been a recent update that could explain these malfunctions? Since I’m a paying customer, is there a dedicated channel or representative I can speak with to address these concerns?

Appreciate any insights or help.

  1. That’s called evidence of change. Yes, the AI model changes as it gets constant tweaking. A replay of particular old conversation sessions will often not get you the satisfaction you had before. However, ChatGPT never responds the same way twice, so you also have to consider the natural variations in output when it does understand the goal similarly.

  2. Yes, that is the “system” part of ChatGPT, the web product: conversation management. It uses its own techniques to let the independent AI input executions know what you were recently talking about, by extracting the most relevant recent chat from your full conversation in that session. Memory can’t go on forever, but also, providing the AI the maximum memory possible also can consume computing resources.

  3. Report the false flagging when your output turns orange. You don’t have control over the AI’s refusals otherwise except to give it a downvote.

All of what you detail is ongoing concern, but it is a useful report that you perceive significant daily changes - if someone was listening. All factors I describe are unlikely to change all at once.

Thank you for your response.

To give you a clearer picture: I’ve been encountering comprehension and short-term memory challenges with ChatGPT-4 that were not present earlier. For example, I often provide the model with a sequence of instructions related to two separate samples. I might say, “Please read Sample A, and once you’ve processed it, I’ll share Sample B. I’d like you to refrain from giving any feedback until you’ve read both samples.” Ideally, once Sample A is processed, the model should indicate its readiness for Sample B. But, in many cases, it either misunderstands this simple directive or provides feedback prematurely.

More concerning is the issue of memory retention. After presenting both samples and then asking about specifics from Sample A, the model behaves as though it lacks any recall or context of that sample, even though only a short interval passed since its introduction. This sudden memory lapse wasn’t an issue I previously faced, marking what feels like a significant regression in the model’s functionality.

To emphasize, while I anticipate variability in AI outputs, the divergence I’ve noticed in the model’s basic comprehension and its ability to hold onto short-term information is stark. This shift has made my interactions more tedious and less efficient.

I hope this more detailed scenario offers a clearer understanding of my concerns. Your assistance in this matter is greatly appreciated.

This forum is fellow users. I can only sympathize and get similarly annoyed when experiencing undesired behavior changes.

For tasks where you know the memory will be unreliable or unpredictable, you can compose your entire question, including all documentation or prior useful responses, into a new question ( often easier with an accompanying text editor). A question of similar quality as if you were putting it into a brand new conversation. Then paste.

(pro tip: shift-enter lets you type new lines instead of sending, so you can make more structured queries to the AI)

Hey @MrAndrew if you are able to share (using the ChatGPT share UI) a few older (good) and newer (bad) examples that are comparable we can def take a look.


Hi, @joetaylor thanks for your response. I had never experienced any issues prior to last week (so just assume they were all normal). My issues are pretty self explanatory once you view them, so please see my interaction below of the sort of issues I have now been facing.

imgur . com/a/lCGf1JU

The link has 5 images, step by step detailing one interaction

ps. It’s not letting me include any links, nor attaching more than 1 image, so I had to separate the link - I’m sure you know how to put it back together :sweat_smile: