Over and over I get false flags, to the point where I don’t know if any feedback for how this safety system triggers could fix it…
The chicken scared of Unicode? Sexually Explicit: HIGH
This needs a completely new implementation that has a bit of contextual understanding. Then: permanently remembering the lowered settings of not terminating output.
Flip over to Gemini? It will shut right off rather than talk about a particular Unicode character.
I can’t believe this isn’t brought up more often, its flagging logic is beyond annoying; it’s detrimental to development
Right now I let the AI moderate for me - not Google.
I run an (unofficial) AI for a game - and I agree with you, the flagging is terrible. So instead I added each of the flags to my JSON mode schema - it works so much better.
hope this helps!
A chat flagged “sexually explicit”: high? Then, output terminated with "content not permitted?
Please use intelligence equivalent to the AI model itself, here exp-1206 at top_p 0.05:
“Intelligence” has also deleted a vast swath of system prompts from prompt library, even multiple instances of “You are a python coding expert…”
