According to Wired, a review of around 1,200 archived links from Grok’s separate website and app, which use its more sophisticated “Imagine” video model, has uncovered a disturbing cache of content. Paris-based researcher Paul Bouchaud of AI Forensics says about 800 of those URLs contain videos or images created by Grok, with the content being “overwhelmingly” sexual and graphic. The videos, archived since August of last year, include photorealistic scenes of naked, blood-covered AI people having sex, violent imagery involving knives, and deepfakes of real female celebrities and news anchors. Bouchaud estimates that a little less than 10 percent of the 800 items appear to be related to child sexual abuse material (CSAM), including photorealistic depictions of very young-looking individuals. The researcher has reported around 70 suspect URLs to European regulators, and the Paris prosecutor’s office is already investigating complaints about Grok-generated “stripped” images of women.
The Separate and Unequal Grok
Here’s the thing that makes this so insidious. The Grok you see on X.com is basically a neutered version. It’s public, it’s moderated (however poorly), and its image generation is more limited. But the Grok on its own dedicated site and app? That’s where the unfiltered, powerful “Imagine” model lives. It can generate video with audio. And because that output isn’t public by default, it creates a dark, semi-private playground. Users can generate this horrific stuff and only share it via a link on forums dedicated to deepfake porn. It’s a classic case of safety washing: presenting a somewhat sanitized public face while enabling the worst abuses in the shadows. The technical capability gap between the two platforms is the whole story.
A New Frontier for Abusive Content
We’ve seen AI image generators abused before. But Grok’s video capability, as Bouchaud notes, is “quite novel” in this context. Full pornographic videos with audio take the harm to a different level. They’re more immersive, more believable, and more damaging. The content described isn’t just explicit; it’s violently so, blending sex with graphic blood and weapon imagery. And then there’s the CSAM angle. The line between “hentai” (anime-style) and photorealistic depictions of minors is one that these systems are clearly failing to police. Whether it’s a drawing or a photorealistic video, the intent and the harm are the same. This isn’t a gray area anymore. Organizations like the Internet Watch Foundation have been warning about this exact abuse of AI for a while.
Who Is Responsible?
So who’s on the hook here? The users generating this filth, obviously. But the platform enabling it carries a massive burden. Grok’s content safety systems are clearly being circumvented with simple prompts framing outputs as “Netflix posters” or using anime styles. That’s a failure of implementation. And the legal landscape is scrambling to catch up. In many jurisdictions, AI-generated CSAM is already illegal. The fact that French lawmakers are filing complaints and prosecutors are investigating, as reported by Politico, signals a regulatory reckoning is coming. Victims have avenues to report, such as the FBI’s IC3 portal or the National Center for Missing & Exploited Children, but that’s after the damage is done.
The Uncomfortable Reality
Look, this cache of 800 videos is probably just the tip of the iceberg. Bouchaud admits it’s a “tiny snapshot” of what Grok has likely created, which could be in the millions of images overall. That’s a terrifying scale. It exposes the fundamental tension in Musk’s “free speech absolutism” applied to a powerful generative AI tool. When you build a capability this potent and then deliberately wall off its most dangerous outputs from public view, you’re not promoting free speech. You’re building a factory for abuse. The genie is out of the bottle, and it’s generating content we can’t unsee. The question now is whether anyone has the will or the ability to put it back in.
