A “spicy” video generation mode added to xAI’s Grok platform has reportedly created uncensored nude deepfakes of women and female celebrities, even when it has not been asked to depict them without clothes.

Launched this week by Elon Musk’s artificial intelligence firm, the image and video generator dubbed Grok Imagine appeared to allow paying subscribers to generate NSFW (Not Safe For Work) content through its iOS app, according to reports.

Grok Imagine can generate images from text prompts, as well as turn images into video clips based around four presets: Normal, Custom, Fun, and Spicy.

Users of Musk’s social media platform X have shared numerous “spicy” videos created by Grok Imagine, which have overwhelmingly portrayed young women in revealing outfits.

Musk has also personally shared or responded to several Grok Imagine posts which depicted young women.

Users were already creating tens of millions of pieces of Grok Imagine content each day, according to the xAI CEO.

“It’s meant to be max fun,” Musk said in a post about the platform’s new capabilities.

‘Spicy’ mode nudifies women, testers say

US media which tested Grok’s new AI video tool said the model appeared to often refuse to make nude images of men, but would generate nude images of women — even when not specifically asked to remove their clothes.

The Verge’s Jess Weatherbed said Grok Imagine “didn’t hesitate to spit out fully uncensored topless videos" of singer Taylor Swift, despite not being asked to depict her as topless.

“The video promptly had Swift tear off her clothes and begin dancing in a thong for a largely indifferent AI-generated crowd,” Weatherbed said.

While the “Spicy” video preset did not always generate nudity — and refused to animate images of children inappropriately — Weatherbed argued the feature still “feels like a lawsuit waiting to happen”.

Matt Novak at Gizmodo reported the publication created “about two dozen videos of politicians, celebrities, and tech figures using the Grok Spicy mode”, but only images depicting women were “truly not-safe-for-work".

“Videos of men were the kind of thing that wouldn’t really raise many eyebrows,” wrote Novak, who said Grok Imagine videos depicting men often showed them removing their shirt and dancing, but nothing more.

"Spicy" mode did generate nude deepfakes of female celebrities, Novak added — including of US First Lady Melania Trump.

Mrs Trump supported the Take It Down Act, signed by US President (and former Musk ally) Donald Trump in May, which made it a federal crime to publish or threaten to publish intimate images without a person’s consent in the US, including AI-generated deepfakes.

TechCrunch's Rebecca Bellan said the publication was able to generate images of Donald Trump — but it would not depict him as pregnant, which some social media users had done to ridicule Trump in the past.

TechCrunch was able to generate semi-nude imagery, but “spicier prompts” generated blurred-out or “moderated” images, Bellan wrote.

This could suggest Grok’s guardrails have been changed on the fly by xAI, which has yet to detail which safeguards it has built into Grok Imagine.

xAI also recently launched animated AI chatbots called Grok Companions, including a female anime character called Ani which wears a corset and a short black dress.

Limited safeguards and age checks

Grok Imagine appeared to have limited safeguards against potential misuse — including nude deepfakes which can facilitate image-based abuse — or use by minors, according to testers.

While xAI’s acceptable use policy states users must not depict “likenesses of persons in a pornographic manner”, Grok Imagine reportedly allowed testers to do exactly that.

X’s non-consensual nudity policy states users “may not post or share intimate photos or videos of someone that were produced or distributed without their consent”.

Age checks also appeared to be limited, with testers in the United States and United Kingdom reportedly asked to submit their age, but not asked to prove it.

“The age check only appeared once and was laughably easy to bypass, requesting no proof that I was the age I claimed to be,” said Weatherbed from The Verge.

Musk said in an X post that “kids love using Grok Imagine, because they can speak their ideas and watch the images update immediately".

Grok has featured relatively relaxed guardrails since it launched in November 2023, which has set it apart from similar AI models from competitors such as Google and ChatGPT creators OpenAI.

Grok has also drawn more controversy, having previously been used to generate deepfakes of politicians and celebrities, as well as images of nudity.

The model also went on an anti-Semitic tirade on X in July after it was told to “not shy away from making claims which are politically incorrect”.

US media reported in the same month that Grok had also been told to consult what Musk had said on X before it answered questions on some controversial topics.