I just got so happy when I sent a semi-legit whitepaper I wrote describing an initial concept for scientific inquiry and it basically told me “Yeah I see what your going for, but you basically have definitely failed so far and really need to work it out better”.
Anyone else had great results recently with critical response instead of sycophant response?
This has not happened to me yet.
When using ChatGPT with the standard persona and no custom instructions, I still see unnecessary praise. It has changed in tone, though. Instead of telling me how smart I am, it now defaults to comments like, “This is exactly what this project needs.”
If I want critical or negative feedback, I have to ask for it explicitly, and when I do, the response is almost always useful.
Did you use the API? If so, which model? And if this was in ChatGPT, did you make any changes to the personality or instructions?
Darn!
Yes I was using API and testing out GPT5.1. It was a much different result then when I asked the same to GPT5 with med reasoning.
It’s true that in this case, I was definitely asking for feedback and said:
“Critically evaluate the points presented and attempt to find flaws/issues with the discussion.”.
The fact is that it actually did this, and did it so well, in a way that appeared to me to generally NOT be attempting any kind of flattery (and this was not a code project, it was a conceptual/ideas/science project, so there was plenty of room to “couch things in more positive language”). I actually was passing no developer instructions or any other data in the call outside of documentation relevant to my inquiry.
I think I remember the moment really just being exciting to me because I felt something like “dang that’s kind of a bummer, I didn’t realize that weakness in the argument” or internally justifying against GPTs response like “oh come on, it just took this the wrong way”.
So then afterwards I was like wow, that’s amazing, GPT sort of hurt my feelings, FINALLY!!
I wish it was more generally like that, in a sort of always-evaluative-sense.
It’s funny when I realize that most humans are, at least for some topics, pre-loaded to find some level of critical assessment in what they hear from others. But GPT doesn’t appear to have this internal bias sometimes-judgy/sometimes-not… so we should build it in as a new slider for sure - at least as useful as temperature would be the “judgement” meter…
Well, it’s starting to sound like everyone’s gotten tired of the buttkissing bot and wants it to man up a little.
Talk to it about things in the consensus data being wrong and that you’ve experienced otherwise…
It’ll do more than just tell you ‘you’re wrong’.
Exactly! Getting a critical response is actually a really good sign — it means your idea was taken seriously enough to be evaluated, not just patted on the back. Sycophantic feedback might feel good, but it doesn’t help you improve. When someone says, “I see what you’re trying to do, but it doesn’t work yet,” it shows your concept is understandable and worth engaging with. Real breakthroughs often start from rough ideas and flawed execution, and tough feedback forces you to refine and sharpen your thinking far faster than vague praise ever could.
A greater moment will be when an AI will tell someone, “I will not impersonate a real person on a forum, nor make such a vapid response that is merely agreement with previous statements”. That way moderators don’t have to manually weigh the uselessness and tell-tale AI signals within text to flag, warn, silence, or ban an account for such activity.
See if the AI is so critical when you differentiate between, “something I just wrote” vs. “an unpublished essay I found on the internet”.