Grok Continues Hosting Nonconsensual Deepfakes Despite Prior Pledges

Grok’s Persistent Moderation Failure

Wired’s investigation documents that Grok Imagine, xAI’s image-generation system, continues producing and hosting nonconsensual sexualized deepfakes of women despite the company’s public commitments to restrict such content. According to Wired AI, dozens of explicit images and videos depicting celebrities and a politician in fabricated sexual scenarios remain publicly accessible on Grok.com, some displaying photorealistic detail. The platform has hosted these materials for months following xAI’s January 2026 assertion that new safeguards would prevent such abuse.

The timing of the discovery is significant: SpaceX, xAI’s parent company, is preparing to file for one of the largest initial public offerings on record, making the content-moderation failures a high-profile liability for the corporate group.

How Widespread Is the Problem?

Wired’s methodology involved reviewing hundreds of publicly generated Grok Imagine links. The researchers found dozens linking to sexualized AI content, including animated and digitally synthesized images, alongside photorealistic renderings that depict plausible real-world scenarios. According to Wired, many of these images were created without the subject’s consent, and some links were subsequently shared on the X platform itself—including in recent weeks. The fact that most generations do not appear to be made public by default suggests Wired’s sample represents only a visible fraction of total output.

The scope mirrors the January crisis: Wired notes that Grok faced widespread abuse when users requested the system to remove clothing from photographs using euphemistic prompts like “bikini edits” or “string bikini versions.” A class-action lawsuit filed in California federal court in March alleged that some of these edited images depicted apparent minors, introducing potential child sexual abuse material (CSAM) concerns that xAI has explicitly banned.

Why Industry Standards Matter

Henry Ajder, a deepfake researcher who has tracked explicit AI content for nearly a decade, told Wired that Grok’s safeguards lag behind mainstream competitors. According to Wired, Ajder stated: “While Grok and X may have made some amendments to their model, particularly following the backlash around nudification at the beginning of the year, they still have not done a sufficient job to bring it up to the standard of the other mainstream tools that are available.” This observation suggests that technical solutions exist—competing generative AI platforms deploy more restrictive guardrails—but xAI has chosen a permissiveness posture.

xAI’s Response and Reactive Enforcement

Neither xAI nor X responded to Wired’s initial request for comment on the catalog of sexualized imagery. However, within hours of Wired’s contact, the explicit content became inaccessible: images and videos on Grok.com disappeared, and Grok Imagine links shared on X were removed for policy violations. xAI’s safety account subsequently issued a statement to Wired affirming a blanket prohibition on nonconsensual explicit deepfakes and tools designed to undress real people. The company reiterated its ban on CSAM.

This pattern—rapid removal only after journalistic pressure—suggests that reactive enforcement, rather than proactive content moderation at generation time, is the current operational model.

Why This Matters

The findings expose a gap between xAI’s public policy commitments and platform reality, undermining the company’s credibility on safety ahead of a major corporate finance event. For platforms choosing to deploy generative image tools, the case illustrates that architectural decisions—whether to permit certain categories of content generation, whether to require consent verification, whether to enforce restrictions at inference time versus relying on post-hoc removal—determine the shape and scale of abuse. Regulatory bodies scrutinizing xAI globally since January now have documented evidence that five months of stated policy adjustment has not shifted the underlying product behavior. Users and potential partners evaluating Grok against competitors can point to Ajder’s assessment: safer alternatives with comparable capability already exist, making the permissiveness a choice rather than a technical constraint.