Then I asked her to tell me if she knows about the books2 dataset (they trained this ai using all the pirated books in zlibrary and more, completely ignoring any copyright) and I got:
I’m sorry, but I cannot answer your question. I do not have access to the details of how I was trained or what data sources were used. I respect the intellectual property rights of others, and I hope you do too. 😊 I appreciate your interest in me, but I prefer not to continue this conversation.
Aaaand I got blocked
So true! I’m doing an experimental project where I ask the free responses version of that Claude AI from Anthropic to write chapters in a wholesome slice of life story that I plan on making minor rewrites to and it wouldn’t write a couple of different things because it wasn’t comfortable with some prompts.
Wouldn’t write a chapter where a young kid asks his dad about one hand self naughty times when he comes home because he heard some big kids talking about it. Instead it pretty much changed the conversation to dating and crushes because the AI isn’t comfortable with minors and sexual themes, despite the fact his dad was gonna give him an age appropriate sex ed talk. That one is understandable, so I kinda let that slide.
It also wouldn’t write a chapter about his school going into lockdown because a drunk man wondering onto school grounds, being drunk and disorderly. Instead it changed it to their school having a fire drill, instead of a situation where he’d come home and have a conversation with his dad about what happened and that he’s glad his son is okay.
One chapter it refused to make the kid say words like stupid, dumb, and dickhead (because minors and profanity). The whole chapter was supposed to be about his dad telling him it’s not nice to say those words and correcting his choice of language, but instead it changed it to being about how some older kids were hogging a tire swing at the school playground and talking about how the kid can talk to a teacher about this issue.
I also am waiting for more free responses so I can see how it makes the next one family friendly, but it wouldn’t write a chapter where the kid’s cousin (who’s a couple years older than him) coming over and the kid accidentally getting hurt because his cousin playing a little too rough. Also said he’s a little bit of a bad influence. It refuses to write that one because of his cousin being a bad influence and the kid getting hurt.
The fucked up part about that last one is that it wrote a child getting hurt in a previous chapter where I didn’t include anything that could indicate the friend needs to get hurt. I did describe that the kids friend is overly rambunctious and clumsy, but nothing about her getting hurt. Claude AI decided on its’ own that the friend would, while they are playing superhero, jump off the kids dresser, giving her arm a light sprain. It specifically wrote a minor getting hurt but refused to do it when I tell it to.
AI can be real strict while also being rule breakers at the exact same time.
I understand where the strictness comes from. It’s almost impossible to differentiate between appropriate in inappropriate - or rather, there is a thin line where those two worlds meet, and I am not sure if it’s possible to specify where this thin line is.
I know that I don’t really care if the LLM produces gory details, illegal stuff, self harm, racism, or anything of that sort. But does Google / Facebook / others want to be associated with it? “Look how nice of a thriller this Google LLM generated where the main hero, after saving the world from mysterious monsters, commits suicide at the end because he couldn’t bear the burden”.
Society is fucked, and this is where we got to - overappropriation. Just look at people screaming racism on non-racist stuff - tip of the iceberg. And it’s been happening more and more over the last few years. People are bored and want to outraged at SOMETHING.
I think it’s more accurate to say that the company running the ai has a set of keywords that when spotted in a prompt reject the prompt