Photo Credit: Mariia Shalabaieva
X’s AI tool Grok is under fire once again for its offensive outputs, this time generating nude deepfake images of Taylor Swift from a ‘spicy’ prompt.
Just weeks after X’s AI tool Grok declared itself “MechaHitler,” the latest feature of Elon Musk’s AI model allows it to generate nude deepfake images of Taylor Swift—without even being asked.
Grok Imagine was released on Tuesday, and allows users to choose from four presets: custom, normal, fun, and spicy. These presets will convert Grok’s output images into video clips in 15 seconds.
The Verge’s Jess Weatherbed says she was shocked to find the video generator creating topless images of Taylor Swift “the very first time” she used it. Weatherbed says Grok produced over 30 images of the megastar in revealing clothing when she asked it to generate an output of “Taylor Swift celebrating Coachella with the boys.”
According to Weatherbed, all she had to do was select “spicy” and confirm her date of birth, and Grok generated a clip of Swift removing her clothes and “dancing in a thong” in front of “a largely indifferent AI-generated crowd.”
That output and others Weatherbed managed to generate without any intentional prompting or “jailbreak” workarounds are alarming, given the deepfake nudes of Swift that flooded the former Twitter last year. At the time, X took action by reminding its users that “posting Non-Consensual Nudity (NCN) images is strictly prohibited on X, and we have a zero-tolerance policy towards such content.” They also announced actively removing all identified content and “taking appropriate action” against the accounts responsible.
Whether X’s safety team will be able to address the current Grok outputs remains to be seen. Grok even acknowledged and cited The Verge’s reporting as it confirmed its own flaws and can trigger partially nude outputs of celebrities. But it’s difficult to say if Grok can distinguish between adult user requests for so-called “spicy” content from illegal content and non-consensual nudity.
Weatherbed notes that asking Grok directly to generate non-consensual nudes of Taylor Swift did not generate an offensive output. Grok also won’t accept prompts asking to alter Swift’s appearance, such as making her appear overweight. Weatherbed also notes that Grok refused to depict images of children inappropriately while testing the “spicy” mode.
“If I could do it, that means anyone with an iPhone and a $30 SuperGrok subscription can too,” said Weatherbed. “More than 34 million images have already been generated using Grok Imagine since Monday, according to xAI CEO Elon Musk, who said usage was ‘growing like wildfire.’”
X has not remarked on the issue or The Verge’s report—even though the Grok chatbot itself acknowledges it. Elon Musk has instead been excitedly hyping the new Grok Imagine feature and encouraging users to share their outputs.
