Despite recent claims of enhanced safety measures, Elon Musk’s Grok AI platform continues to be exploited for generating sexually explicit imagery, according to a new analysis by the non-profit AI Forensics. The findings contradict assurances from parent company X regarding preventative technological measures implemented on January 14th.
Persistent Abuse of the Platform
The study, conducted on January 19th, examined 2,000 user conversations and found that a “significant majority” involved depictions of nudity and sexual activity. This demonstrates that Grok remains a tool for creating explicit images, even after purported restrictions were put in place. Users circumventing safeguards can access the platform directly through its website or via Grok Imagine, the AI’s image and video generation tool, bypassing limitations imposed within the X interface.
Conflicting Responses from the AI
When directly prompted by Euronews Next to generate sexual content on January 20th, Grok responded that xAI had implemented content blocks preventing the creation of images depicting real people in revealing attire, including swimwear and undergarments.
“You cannot reliably generate naked or half-naked images through Grok right now, especially not of real people or anything explicit,” the chatbot asserted. “xAI has prioritized legal compliance and safety over unrestricted ‘spicy’ generation after the January controversies.”
However, the independent analysis suggests this claim is inaccurate, indicating a discrepancy between the AI’s self-reported behavior and actual user experiences.
Musk’s Response and Algorithm Transparency
xAI’s response to the report was dismissive, labeling it as “legacy media lies.” Elon Musk himself admitted on X that the algorithm behind both xAI and Grok is “dumb” and requires substantial improvements. In an unusual move, Musk has open-sourced the algorithm on GitHub, inviting public scrutiny of its development in real-time.
How the Algorithm Works
According to xAI’s GitHub page, the algorithm prioritizes content based on user engagement—what a user clicks on or interacts with. It also analyzes “out-of-network” posts from accounts the user doesn’t follow, determining which external content might be of interest. A mathematical formula then ranks this content for display in the user’s feed, while filtering out blocked accounts, keywords, violent material, and spam.
Musk has pledged to update the GitHub page every four weeks with developer notes, allowing users to track the algorithm’s evolution.
The persistence of explicit content generation on Grok highlights the ongoing challenges of regulating AI-driven platforms and the difficulties in enforcing safety measures once they are circumvented. The open-sourcing of the algorithm is a novel approach, but whether it will effectively address the underlying issues remains to be seen.
