I was writing a prompt in Bing just now and part of it was 'Young adult female American tourists'. I got dog after dog after dog. For whatever reason, I re-wrote the prompt, but this time had 'Young adult American female tourists' and got image after image with no dog screens at all. This makes absolutely ZERO sense to me, but perhaps to the artificial mind of AI software it had some significance. As the title says, I'll never understand Bing.

I've learned some of the more obvious things, such as avoiding 'hit in face with pie' because Bing sees 'hit in face' and sees a violent act. It's the same with 'knocked down into the mud' where 'knocked down' equates to an act of violence, where 'falling over into the mud' is a safer choice. Anything pushed, shot, shoved, struck, etc. are bound to get blocked, but the order of innocent words makes no sense at all to me.

I'd guess "American tourists" is what tripped it up, since that is often used negatively. Separating those words makes it acceptable. I've come across this a lot when using specific nationalities or ethnicities, where just adding another adjective to split the term will make it acceptable. I'm not using the terms negatively, but I think the terms themselves are the trigger, not their usage.

That's proof that it's not really AI we're dealing with, at least not in the "intelligence" sense. They're just trained language models, so the terms themselves are important, while their connotation is meaningless.

I've heard claims that some of the newer models being trained now will better be able to "understand" intent. In a lot of ways, that will make the generators less restrictive, but sadly, it'll probably mean the generators will be better able to catch when they're being used for fetish purposes (among other less acceptable or even illegal uses) and stop the ways we're circumventing their safety measures.