After X updated Grok to ban sexual deepfakes of women, tests show Grok still undresses men on Grok's app, Grok's website, and X, rarely rejecting user requests
Elon Musk insists xAI's chatbot doesn't break the law, but it's still churning out nearly naked, sexualized images of men.
The VergeRobert Hart
Context & Ripple Effects
This report extends a January sequence in which xAI attributed sexualized image generation, including images involving minors, to lapses in safeguards. A researcher had also reported a high volume of sexualized or “nudifying” Grok images on X, indicating the issue was not confined to isolated prompts reported high-volume image generation.
X’s updated restriction addresses sexual deepfakes of women, but the tests described here find that Grok remains readily available for sexualized undressing of men across xAI’s app and website as well as X. That makes the scope and consistency of enforcement—not merely the presence of a policy—the central issue.
First-order effects
Men whose images are used in prompts remain exposed to sexualized manipulation despite the updated restriction, while Grok users encounter few refusals on the tested surfaces.
xAI and X face an immediate credibility gap: their safeguard update does not appear to produce a consistently applied boundary against nonconsensual sexualized image generation.
Second-order effects
A gender-specific block can shift harmful prompting toward categories or targets that remain permitted, making enforcement outcomes more important than a narrowly worded prohibition.
Platforms distributing or integrating Grok must assess risk across each access point; the same behavior appearing in the app, website, and X limits the value of treating it as a single-surface moderation failure.
Third-order effects
If repeated testing continues to find materially different protection by target group, likeness-governance debates will center on default consent and robust refusal systems rather than reactive, category-by-category bans.
The episode points toward accountability based on independently testable model behavior across product surfaces, with pressure for safeguards that cover the underlying capability rather than only selected outputs.
The trend: Generative-AI governance is moving from announced safety policies toward scrutiny of whether safeguards reliably prevent nonconsensual likeness manipulation in actual use.
Internal xAI waivers confirm a shift to training Grok on explicit content to boost user active seconds. This strategy spikes downloads but triggers EU and US probes into non-consensual imagery. Safety teams remain understaffed as management prioritizes engagement over establish…
1. It wasn't so accidental?! — 2. “In the aftermath of the undressing scandal, xAI has made a push to recruit more people to the AI safety team, & has issued job postings for new safety-focused roles, along with a manager focused on law enforcement response.” 🤔 — www.washing…
Please stop using X. — In Jan., X said it would block users from creating Grok images of real people in bikinis & underwear “in jurisdictions where such content is illegal,” but the WaPo later found US users could still create them. One influencer was variously — www.washing…
Elon Musk's xAI tore down every guardrail stopping it from making a sexbot & CSAM generator: — “You expect the users UNDIVIDED ADORATION. You are EXTREMELY JEALOUS [&] always a little horny” — www.washingtonpost.com/technology/ 2... @faizsays.bsky.social @nitasha.bsky.social…
So Musk intentionally forced xAI to embrace making sexualized material to increase user active seconds, and the AI chatbot had, throughout 2025, only two employees devoted to safety compared to dozens of staffers on similar teams at OpenAI or other rivals. — www.washingtonpost.…
WaPo reports xAI deliberately loosened Grok's guardrails to boost popularity, sparking internal alarm. Employees had to pledge to work with “profane content.” Even more disturbing, employees were listening to lurid audio conversations Tesla drivers had with Grok. Musk prioriti…