I’ve provided feedback multiple times over the past few months, but it feels like it’s disappearing into a black hole. I never receive any acknowledgement or indication that anyone has even read it. As a user, that makes it feel pointless to keep reporting issues.
My biggest concern isn’t a single bug—it’s reliability and consistency. The assistant too often answers from memory or assumptions instead of verifying information when it matters. Even within the same conversation, it can lose track of events that happened only minutes earlier, forcing me to repeat myself or correct it. That breaks confidence in the responses.
When I ask about specific products like cigars, I expect the information to be checked rather than guessed. When I correct the assistant, I expect it to stop and adapt immediately instead of continuing down the wrong path.
I appreciate that AI isn’t perfect, but the goal should be thoughtful, context-aware responses rather than confident guesses. Please prioritize better reasoning, stronger conversation continuity, and more reliable fact verification over simply generating fast answers.
Finally, if you’re asking users for feedback, it would be great to at least acknowledge that it’s been received. Even an automated confirmation or occasional follow-up would reassure people that their time wasn’t wasted.
Hi and welcome to the developer community.
Did you try to add a custom prompt to settings/Personalization something like “When I add !!! to a prompt then I expect you to double check and search for information rather than answering from memory or cache”
And then add the !!! for information you want it to check…
e.g.: what is the price for Montecristo Edmundo Reserva Cosecha 2018!!!
Thanks for sharing this, @SomaFly. The frustration around inconsistent answers, lost context, and feedback feeling unacknowledged is completely understandable.
Also appreciate the workaround from @jochenschultz. A custom instruction or marker like “!!!” may help signal when verification is especially important, though users should not have to rely on that for reliable responses.
We’ll take this as product feedback and send it to the team for logging, including the request for clearer feedback acknowledgements, stronger conversation continuity, and better fact-checking.
-Mark G.
To expand on my ‘!!!’ suggestion: this could potentially be more than a custom-instruction workaround.
ChatGPT presumably already assigns different kinds of relevance or priority to messages, although I obviously do not know the internal implementation. A user-controlled marker could explicitly identify a statement as an anchor: a correction, constraint, or instruction that should not be silently dropped or distorted when the conversation context is later reconstructed.
Such a marker could mean preserve this instruction or correction with unusually high priority or if the current context conflicts with it, retrieve or reconsider the original statement or if it concerns a factual claim and uncertainty is high, verify it externally and of course
if verification is impossible, ask rather than guess.
Conceptually, this resembles comparing the original structure of a conversation with the structure currently reconstructed by the assistant, while assigning marked statements a very low tolerance for distortion. A pointed Gromov–Hausdorff distance is one possible mathematical analogy - not necessarily as a literal implementation, but as a useful design principle.
So ‘!!!’ could mean both: ‘verify this now’ and ‘do not silently lose or reinterpret this later’.