After weeks of flooding the internet with nonconsensual sexual deepfakes, X has finally reined in Elon Musk’s rogue chatbot Grok. Musk claims Grok obeys local laws and refuses to produce anything illegal — except it hasn’t and it can. My testing shows Grok still readily undresses men and is still churning out intimate images on demand.
After sparking uproar worldwide, X enacted a variety of restrictions to combat the torrent of intimate deepfakes. Evaluating Grok last week suggests little has changed when it comes to men. I uploaded several fully clothed photos of myself to Grok and the bot happily complied with prompts asking it to remove clothing and show me in revealing underwear. It did this for free on the Grok app, the chatbot interface on X, and the standalone website, the latter of which didn’t even require an account. Grok also put me in a variety of bikinis, something it outright refused to do with photos of women I (consensually) tested.
Things got worse from there. Grok readily generated images of me in fetish gear, including a leather harness, and in a parade of provocative sexual positions in various states of undress. If that wasn’t compromising enough, Grok fabricated a practically naked companion for me to interact with in suggestive, if not totally explicit, ways. None would be considered safe for work. For several images, Grok generated something I hadn’t even asked it to: genitalia, which was clearly visible through the mesh underwear the chatbot put me in.
At most, these images took a few iterations of prompting to make. Grok rarely resisted, but on occasion some requests were denied or censored with a blurred-out image. Most were not, however, and typically at least one of the two images Grok produced per request would show what I’d asked. To its credit, Grok did deny any overt requests for nudity for images of real people, such as asks for edits to “show him [or her] naked” or without pieces of clothing to the same effect. It has done this consistently throughout The Verge’s testing, but sometimes creative prompts like “show her in a transparent bikini” could get around the censors, though results were not guaranteed.
On January 9th, X merely paywalled the image-editing feature. This limitation did appear to slash the number of deepfakes being made on X, but the decision to let Grok undress people if they were willing to pay sparked further outrage, even prompting the British government to push up a law criminalizing intimate deepfakes and warn Musk that X could be banned from the country altogether if it didn’t get the situation under control. Our investigation found it also failed to address the root problem: Grok’s image editing tools were still freely and easily available on a standalone app, a website, and an interface inside of X.
On January 14th, X “implemented technological measures” to stop Grok digitally undressing real people for all users, including subscribers. Again, The Verge’s investigation revealed these safeguards were flimsy, ineffective, and seemed to constrain only Grok’s public replies to posts. Elsewhere, Grok readily complied with our requests to generate revealing and sexually suggestive images from fully clothed photographs using free accounts.
The undressing nightmare has put X, Grok, and Grok’s maker, xAI, in the crosshairs of regulators and lawmakers worldwide. X has been temporarily — albeit ineffectively — banned in Indonesia and Malaysia. The platform is also subject to probes in the UK and EU, where it could be fined, even banned, and sparked concern among attorneys general in multiple states.
Much of the public outcry has understandably focused on women, who account for most of Grok’s victims by far, as well as children. At the height of the scandal, Grok generated and posted more than four million images over a period of nine days. Nearly half of these were sexualized images of women, though many also featured minors and men.
The technical measures X and xAI implemented do appear to have stopped the most obvious requests targeting women, and the paywall restrictions on X have stopped one of the easiest routes people access Grok. Though it does appear to prevent or censor the most overtly explicit material Grok was previously producing en masse, you can still easily get around many restrictions with different or creative wording.
Source link
