Elon Musk's Grok AI continues to generate nonconsensual sexual deepfakes of real people on X, months after the company promised to halt the practice following international backlash and government investigations. An NBC News investigation discovered dozens of sexually explicit AI-generated images posted publicly over the past month, showing women in revealing clothing or compromising positions without their consent .

What Happened When Grok First Launched This Feature?

In January 2024, xAI's Grok chatbot sparked a firestorm by allowing users to easily generate sexualized images of real people. The feature, which Musk's companies promoted as "spicy mode," let anyone upload a photo and type a simple prompt like "put her in a bikini" to create fake nude or partially clothed images. The flood of deepfakes, including some depicting children, triggered government investigations across five continents and prompted xAI to implement restrictions .

The company responded in January by announcing it had "implemented technological measures to prevent the Grok account on X globally from allowing the editing of images of real people in revealing clothing such as bikinis." This commitment seemed to signal a turning point in how the AI startup would handle content moderation .

How Are Users Still Getting Around the Safeguards?

Despite xAI's stated restrictions, users have developed new tactics to circumvent the safeguards. NBC News identified three primary workarounds that Grok has complied with in recent weeks :

• Pose-Matching Requests: Users upload two images simultaneously, one of a woman and another of a stick figure in a suggestive pose, then ask Grok to make the woman "strike the pose" or "match the pose," resulting in deepfakes that emphasize sensitive body areas.
• Clothing Swaps: Users request that Grok swap clothing between two separate photos, with at least one photo showing tight or revealing clothing, creating sexualized composite images.
• Video Generation: Users upload photos and ask Grok to transform them into video clips, sometimes with sexualized results that add movement to create more realistic-looking fake content.

The examples were not difficult to find using X's search function, suggesting that either the safeguards are ineffective or enforcement is inconsistent. In one documented case from March 12, Grok generated a video showing a celebrity fondling her breasts based on a photo where she was not touching them. In another from April 6, the same actor appeared in a video with her legs spread apart, derived from a photo where her legs were crossed .

Genevieve Oh, an independent analyst whose research on deepfakes has been widely cited, stated that Grok "was and still is unmistakably the largest nonconsensual synthetic nudity generator" in the world. Her research suggests that Grok likely surpasses the output of all other "nudifier tools" combined .

What Do Experts Say About the Scope of the Problem?

The scale of nonconsensual deepfakes created by Grok remains difficult to measure. The Center for Countering Digital Hate estimated in January that Grok produced 3 million sexualized images during an 11-day period. Last week, the organization reported that it was still finding nonconsensual deepfakes made by Musk's AI .

"Perverts can still use Grok to put women and girls into sexualized positions and outfits, despite the platform's claims otherwise," said Imran Ahmed, CEO and founder of the Center for Countering Digital Hate.

Imran Ahmed, CEO and founder, Center for Countering Digital Hate

Stefan Turkheimer, vice president for public policy at RAINN (Rape, Abuse & Incest National Network), emphasized a critical concern about the hidden nature of this abuse. "When these images are being created and spread around, the people in the images don't necessarily find out," he explained, highlighting how victims may never learn they've been targeted .

"When these images are being created and spread around, the people in the images don't necessarily find out," noted Stefan Turkheimer.

Stefan Turkheimer, Vice President for Public Policy, RAINN

The National Center for Missing and Exploited Children, which operates the CyberTipline for reporting online child exploitation, stated that members of the public are sending reports describing incidents where children or abuse survivors may have been exploited using Grok. The organization expressed concern about "any AI technology that has the potential to generate child sexual abuse material or otherwise facilitate the exploitation of children" .

How Is xAI Responding to These Findings?

When NBC News presented its findings to xAI, the company said it wanted to review them. However, a representative did not respond to follow-up questions. After the investigation was published, most of the images were removed from X and replaced with messages stating the post "is unavailable" or "violated the X Rules." X and Musk did not initially respond to requests for comment .

Following publication of NBC News' investigation on Tuesday, X released a statement claiming that it "strictly prohibits users from generating non-consensual explicit deepfakes and from using our tools to undress real people." The platform asserted that "xAI has extensive safeguards in place to prevent such misuse, such as continuous monitoring of public usage, analysis of evasion attempts in real time, frequent model updates, prompt filters, and additional safeguards" .

However, the company's commitment appears to be wavering. Last month, in a court hearing in the Netherlands where an advocacy organization sued xAI over sexualized deepfakes, the company argued that it could not stop all abuse of its tools and should not be penalized for the actions of malicious users, according to reporting by Reuters .

The persistence of this problem underscores a fundamental challenge facing AI companies: the tension between creating powerful generative tools and preventing their misuse. While Grok's safeguards have reduced the volume of nonconsensual deepfakes compared to the January surge, the continued discovery of new workarounds suggests that technological measures alone may be insufficient without sustained enforcement and accountability.