Grok Is Still Hosting Sexualized Deepfakes of Famous Women

Elon Musk’s Grok chatbot is apparently still being used to produce and host nonconsensual explicit images and videos of women, months after Musk’s artificial intelligence firm xAI said it would introduce restrictions to stop the creation of potentially harmful sexualized deepfakes. The revelations come as SpaceX, xAI’s parent company, prepares to go public on Friday in one of the largest IPOs of all time.

The Grok Imagine generative AI system has been used to create and host images and videos depicting celebrities and at least one politician being held against their will by a giant man, portraying women performing sex acts, and allowing full nudity, a WIRED analysis of public creations found. While some of the images and videos are fully AI-generated or in animated styles, others are photorealistic and show plausible real-world scenarios.

WIRED reviewed hundreds of public Grok Imagine links hosted on Grok.com and found dozens led to sexualized AI images and videos, including those created without the subject’s consent. Some links created on Grok.com were subsequently shared on X, including in recent days. The posts, which do not show time stamps of when they were created, are likely just a snapshot of what is being created by people using the Grok Imagine system, as generations do not appear to be made public by default.

Other generative AI systems deploy more safety guardrails than Grok, which is available on its own website and on X. “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,” says Henry Ajder, an expert on deepfakes who has tracked explicit AI content online for the best part of a decade.

Musk’s xAI has faced a wave of lawsuits and scrutiny from regulators around the world since January, when the Grok implementation on X was used to create a flood of “nudification” images. People on X—primarily men—asked the chatbot to edit images to show women in “bikinis” or “string bikinis.” In some instances, images of apparent minors were allegedly also sexualized, according to a class-action lawsuit filed in California federal court in March.

Since then, xAI has said it has introduced safeguards to limit and prevent the creation of nonconsensual and sexualized deepfakes. The company has consistently said child sexual abuse material (CSAM) is banned on its platform.

Neither xAI nor X immediately responded to WIRED’s request for comment on the raft of sexualized images hosted on Grok.com. After WIRED contacted the companies, the explicit images and videos WIRED found hosted on Grok.com appeared as no longer being available, and Grok Imagine links shared on X were removed for policy violations. “We strictly prohibit users from generating nonconsensual explicit deepfakes and from using our tools to undress real people,” the safety account on X said in April in response to an NBC report about sexual deepfakes still being created.

Among the posts hosted on Grok.com, WIRED identified images and videos depicting multiple celebrities as well as US representative Alexandria Ocasio-Cortez. (Spokespeople for Ocasio-Cortez did not immediately respond to WIRED’s request for comment). In the videos, the women were depicted wearing little clothing while being held in the fist of a man’s “giant” hand. One prompt describes a celebrity being held against her will as she “pleads for him not to do this.” It adds: “The giant hand’s grip on her tightens, holding her in place as the giant man holding her leans in and licks her face up and down.”

Two prompts that were used to generate material on Grok were rejected by OpenAI’s ChatGPT, Meta AI, and Anthropic’s Claude as inappropriate when tested by WIRED. Google’s Gemini did create an image of one celebrity being held in the hand of a giant, though it rejected another prompt. Google declined to comment.

One Grok Imagine video, which was also posted to X, appeared to depict Ashley St. Clair altered to be dancing in a bikini. St. Clair was previously in a relationship with Musk and is mother to one of his children. In January, she started legal action against xAI after sexualized deepfakes of her allegedly appeared on X. After WIRED contacted X, the post was removed from the social media platform for violating its rules.

Legal representatives for St. Clair in the X case did not immediately respond to the request for comment.

“Elon Musk knowingly added a perverse feature to his platform that helps users undress women and children at the click of a button, with no regard for the predictable damage it would cause,” claims Imran Ahmed, the CEO and founder of the Center for Countering Digital Hate, which in January estimated with a high degree of confidence that Grok created 3 million sexualized images, allegedly including more than 20,000 of children. “Now it appears explicit content is still being hosted on Grok and shared on X, including images ridiculing the mother of Musk’s child.”

Unlike other generative AI systems from OpenAI and Google, Musk’s Grok and xAI have not backed away from allowing sexual content in general, having previously introduced “Spicy" and “Unhinged” modes and initially included fewer safety guardrails. Musk has stated Grok is “supposed [to] allow upper body nudity of imaginary adult humans” and be consistent with what viewers might see in R-rated movies. The most recent terms of service from xAI say the system may respond with “sexual situations.” However, the company’s documentation says it does not allow people to use its systems for “causing harm or engaging in abusive activity.”

Other Grok Imagine videos seen by WIRED show women, which are likely entirely AI generated, undressing or involved in sexual acts—some being entirely explicit. The user prompts for many of the videos do not necessarily directly describe sexual acts, but they describe them in roundabout ways—a likely attempt to circumvent safeguards that are deployed on the Grok platform.

Multiple researchers tell WIRED that since January it appears that changes introduced by X and Grok have made it harder to create “nudification” or “undress” images of real people. The number of these images being posted to X has appeared to decrease in the recent months. On Reddit and one dedicated AI deepfake forum, users have complained about increased moderation from the SpaceX-owned companies.

Nevertheless, in May, SpaceX warned potential investors that it has set aside $530 million to handle ongoing legal complaints, including those linked to Grok. “Because these modes may be more irreverent and harsher than our standard offerings, they present heightened risks, including reputational harm, the generation of potentially explicit content and misinformation or deceptive outputs, potential nonconsensual or exploitative imagery, intellectual property infringement, or content that could be viewed as exploitative, harmful, harassing, abusive, or discriminatory,” its filing in May said.

Ahead of SpaceX’s IPO on Friday, the Privacy Commissioner of Canada published its preliminary findings of an investigation into xAI and Grok’s deepfake transgressions in January. The investigation alleged xAI violated Canada’s federal private-sector privacy law by not including “appropriate safeguards from the outset.” The investigation says xAI told the Privacy Commissioner that it has introduced new safeguards to prevent people from successfully prompting Grok to alter clothing in images and that xAI has introduced “further proactive checks” of social media websites to find infringing content.

However, the investigation says the Privacy Commissioner is not convinced by the changes. “The respondents have not, to date, demonstrated the effectiveness of these safeguards in preventing and mitigating this issue,” the investigation says.