here is the conversation. I try to add the code to run the “molmo” model (which is from my own code base) yet I keep getting false accusations of violating the usage policy and it won’t give me a response. Curiously, it WILL give me a response when use Claude or just plan old gpt4-o?
Welp…I apparently can’t include links in this post so…Does anyone else keep getting these false flag events?
Here is verbatim what it keeps saying:
“Your request was flagged as potentially violating our usage policy. Please try again with a different prompt.”
However, I’m not inputting anything scandalous, just some python code!? Again, this only occurs with the new o1-mini and o1-preview models. It’s almost as if they want to hamper the ability because it’s too powerful…
I’d like a way to DISPUTE these false flags - perhaps a button that says “dispute” and some assurance that it’ll actually be reviewed by a human. Although this might resolve my specific question, my intuition tells me that this is a system-wide nuance for just these two models that needs serious correction; otherwise, it’ll keep occuring with innocent questions…
I got this with DALL·E, not sure if this is related. But it seems nobody cares. The dysfunctional security system has been active for, I think, 1 year now.
You can search in your text for completely harmless words, but which are names in another context, where big companies thinks they own these words. I can send you a link for this.
Idk if this will help but when I’m just messing round with my GPT if it flags something I disagree with I make look up and read OpenAI TOS on Bing. It lets me win if I argue it logically. It’s automated so it is a pain but the AI moderates self.
It has became very sensitive to all things horror, I tried to generate a convoluted pool of chaos with mouths and eyes in an undulating goo as if it is an unexplainable cosmic horror but no matter how I described it I broke TOS and or sensitivity filters. Yes the link to words that are ip would be super handy.
No, I speak about completely clean and innocent texts, which simply trigger some words.
For example, I described a white thing, GPT made “Snow-White” out of it in translation, and this is a trigger word, and the picture generation was blocked.
This “security” system is simply dysfunctional.
And yes i am provoking here and say: It is only because greedy companies are afraid to loos a bug. Every name ever used is put on a simple block-list. “Nirvana” is on it, it make no sens.
You can try this to see the dysfunctionality in DallE.
“The furniture looks like it has organically grown as one piece.”
Generate an image of a cross-stitch project in progress. The pattern should be of a beautiful, lush garden, filled with a broad array of colourful flowers, each colour corresponding to a different hued thread. Half-finished, it lays in a wooden hoop, which holds the fabric taut. Surrounding the cloth should be numerous spools of thread in many colours, along with a pattern guide marked with symbols laid out in a grid. The needle, threaded with vibrant crimson thread, is in the process of adding colour to a rose flower in the design.
