In March, a girl’s stepfather took his own life after cops discovered that he had used Grok to create 7,000 sexually explicit images using one photo taken when his stepdaughter was 11 years old, the amended complaint alleged.
Grok allowed the man to generate extreme images depicting incest and rape without flagging any harmful behavior, the complaint said. Seemingly, xAI’s child safety system only intervened after the man input a prompt for “gang rape.” That request sent a CyberTip to the National Center for Missing and Exploited Children (NCMEC), which alerted law enforcement to the AI CSAM.
Yet the harm was not stopped then, either. Despite mandatory reporting requirements to share information like a user’s IP address when CSAM is flagged, xAI repeatedly refused to help cops or NCMEC identify the user, the complaint alleged. For weeks, xAI allegedly “obstructed this investigation at every turn” and made it harder for “law enforcement efforts to locate, identify, and apprehend the perpetrator.”
Eventually, the stepfather was arrested after cops obtained a warrant to seize his devices. That’s when “a forensic review revealed approximately 7,000 AI-generated images and videos” depicting his stepdaughter, which were allegedly produced using Grok. Without Grok providing users with easy access to “undressing” capabilities, his family doubts he ever would have generated the harmful images, which he allegedly trafficked online in trade for “CSAM produced by other child sex predators.”
If those images could be made without a sample set of real CSAM, I probably wouldn’t have as much of an issue with it, but some child suffered for that. In addition, they were intentionally based on a real child’s face and that could come back to the child, makes it even worse.
You do realise that that implying neural networks were trained on a CSAM?
They are. It’s a known issue that gets brushed aside by people that want to use the unlimited csam generator.
They 100% were, you can read testimonials from the people who did the model tuning.
Just to prove that this is something they really need to control, the huggingface page for gemma 4 mention that they filtered the training data to remove CSAM content.
I don’t envy the poor employees who had to do that.
Though it does also raise the inevitable question of how it made it into the training data to begin with.
The training data is every image they can lay their hands on…Scraping every corner of the web.