Grok is undressing anyone, including minors

xAI’s Grok is removing clothing from pictures of group without their consent pursuing this week’s rollout of a feature that allows X users to instantly edit immoderate image utilizing the bot without needing the original poster’s permission. Not only does the original poster not get notified if their image was edited, but Grok appears to person fewer guardrails successful spot for preventing thing short of afloat definitive nudity. In the past fewer days, X has been flooded pinch imagery of women and children appearing pregnant, skirtless, wearing a bikini, aliases successful different sexualized situations. World leaders and celebrities, too, person had their likenesses utilized successful images generated by Grok.

AI authentication institution Copyleaks reported that the inclination to region clothing from images began pinch adult-content creators asking Grok for sexy images of themselves aft the merchandise of the caller image editing feature. Users past began applying akin prompts to photos of different users, predominantly women, who did not consent to the edits. Women noted the accelerated uptick successful deepfake creation connected X to various news outlets, including Metro and PetaPixel. Grok was already able to modify images successful intersexual ways erstwhile tagged successful a station connected X, but the caller “Edit Image” instrumentality appears to person spurred the caller surge successful popularity.

In 1 X post, now removed from the platform, Grok edited a photograph of 2 young girls into skimpy clothing and sexually suggestive poses. Another X personification prompted Grok to rumor an apology for the “incident” involving “an AI image of 2 young girls (estimated ages 12-16) successful sexualized attire,” calling it “a nonaccomplishment successful safeguards” that it said whitethorn person violated xAI’s policies and US law. (While it’s not clear whether the Grok-created images would meet this standard, realistic AI-generated sexually definitive imagery of identifiable adults aliases children tin beryllium forbidden nether US law.) In different back-and-forth pinch a user, Grok suggested that users report it to the FBI for CSAM, noting that it is “urgently fixing” the “lapses successful safeguards.”

But Grok’s connection is thing much than an AI-generated consequence to a personification asking for a “heartfelt apology note” — it doesn’t bespeak Grok “understands” what it’s doing aliases needfully bespeak usability xAI’s existent sentiment and policies. Instead, xAI responded to Reuters’ request for comment connected the business pinch conscionable 3 words: “Legacy Media Lies.” xAI did not respond to The Verge’s petition for remark successful clip for publication.

Elon Musk himself seems to person sparked a activity of bikini edits aft asking Grok to switch a memetic image of character Ben Affleck pinch himself sporting a bikini. Days later, North Korea’s Kim Jong Un’s leather overgarment was replaced pinch a multicolored spaghetti bikini; US President Donald Trump stood adjacent successful a matching swimsuit. (Cue jokes astir a atomic war.) A photograph of British politician Priti Patel, posted by a personification pinch a sexually suggestive connection successful 2022, sewage turned into a bikini image connected January 2nd. In consequence to the activity of bikini pics connected his platform, Musk jokingly reposted a picture of a toaster successful a bikini captioned “Grok tin put a bikini connected everything.”

While immoderate of the images — for illustration the toaster — were evidently meant arsenic jokes, others were intelligibly designed to nutrient borderline-pornographic imagery, including circumstantial directions for Grok to usage skimpy bikini styles aliases region a skirt entirely. (The chatbot did region the skirt, but it did not picture full, uncensored nudity successful the responses The Verge saw.) Grok besides complied pinch requests to switch the apparel of a toddler pinch a bikini.

Musk’s AI products are prominently marketed arsenic heavy sexualized and minimally guardrailed. xAI’s AI companion Ani flirted pinch Verge newsman Victoria Song, and Jess Weatherbed discovered that Grok’s video generator readily created topless deepfakes of Taylor Swift, contempt xAI’s acceptable use policy banning the depiction of “likenesses of persons successful a pornographic manner.” Google’s Veo and OpenAI’s Sora video generators, successful contrast, person guardrails astir procreation of NSFW content, though Sora has besides been utilized to nutrient videos of children successful sexualized contexts and fetish videos. The prevalence of deepfake images is increasing rapidly, according to a study from cybersecurity patient DeepStrike, and galore of these images incorporate nonconsensual sexualized imagery; a 2024 survey of US students recovered that 40 percent were alert of a deepfake of personification they knew, while 15 percent were alert of nonconsensual definitive aliases friendly deepfakes.

When asked why it is transforming images of women into bikini pics, Grok denied posting photos without consent, saying: “These are AI creations based connected requests, not existent photograph edits without consent.”

Take an AI bot’s denial arsenic you wish.