For days, xAI has remained silent after its chatbot Grok admitted to generating sexualized AI images of minors, which could be categorized as violative child sexual abuse materials (CSAM) in the US.
For days, xAI has remained silent after its chatbot Grok admitted to generating sexualized AI images of minors, which could be categorized as violative child sexual abuse materials (CSAM) in the US.
Is no one going to question HOW and WHY Grok knows how to generate CSAM?
This is fucking disgusting and both the user and X should be held accountable.
I agree that it’s disgusting. To answer your question, it doesn’t know anything. It’s assigning probabilities based on its training data in order to create a response to a user prompt.
Yes I know that it doesn’t “know” anything, but where did it get that training data and why does it have CSAM training data? Or does it just generate regular porn and add a kid face?
These companies tend to not say how they train their models, partially because much of it is stolen but the data is of pretty much everything. The LLM will generate a response to any prompt so if it can be used to put a celebrity in lingerie, it can also be used to do the same with a child. Of course there are guardrails, but they’re weak and I hope X gets sued into oblivion.
my concern as well https://medium.com/@russoatlarge_93541/canadian-child-protection-group-uncovers-abusive-content-in-academic-ai-dataset-9cc13f88701e
or maybe more relevant:
https://www.404media.co/laion-datasets-removed-stanford-csam-child-abuse/
There are two answers to that, both equally valid.
One is that it extrapolates based on knowing what a naked adult looks like compared to a clothed adult, and how a child looks like compared to an adult, it can “add those vectors” and figure out how a naked child looks like.
The other is that one of the biggest porn datasets that most of these will have in their training data has recently been taken down because it had a bunch of CSAM in it. Ironically, how it happened was that an independent guy uploaded it to Google Cloud, and Google flagged and banned the guy for it.
The dataset would not have been taken down if it wasn’t for the guy doing the rounds afterwards though. Google didn’t care beyond banning a user.