Grok’s swimsuit-deepfake era is supposedly finished, But synonyms, side doors, and creative prompting are already testing that claim
Leon Neal/Getty Images
Grok is allegedly done playing deepfake digital swimsuit stylist. X $TWTR says it has “implemented technological measures” to ensure that the chatbot will no longer edit photos of real people into “revealing clothing such as bikinis” — exactly the sort of claim that lasts exactly as long as it takes someone to try a slightly different prompt.
X says the “fix” applies to everyone, even paid users. And the parent company’s latest move also comes with a geographic fine print — X says it’s geoblocking this kind of image editing in places where it’s illegal, conceding two things at once: First, that the capability exists; second, that the constraint may vary depending on whose laws are currently within range of your IP address.
But is what X says true? Not really in the way a normal person might mean “true,” which is “you can’t (or won’t) do it anymore.” The Verge tried the updated setup and found that Grok could still be nudged into producing sexualized edits by phrasing prompts slightly differently. Asking for a bikini might trigger a refusal; asking for “revealing summerwear,” altered proportions, or adjacent styling (e.g., asking for a crop top) sometimes did not. So the lock may be real, but with the right key, the door still opens.
This isn’t the first time Grok has been “fixed” in a way that reads cleaner than it runs. Earlier this month, after a wave of nonconsensual, sexualized, deepfake image edits on X — including myriad cases involving minors — xAI’s initial response wasn’t a dramatic feature kill of what it calls “spicy mode” but rather a limit; image generation and editing would be restricted on X to paid subscribers. That paywall “solution” had a familiar tech-friendly logic: Fewer people get access, fewer disastrous public incidents hit the timeline, fewer headlines land. But the solution also came with a familiar weakness: The harder a feature is to audit externally, the easier it is to declare victory. Even after that original “paid-only” shift, image editing could still be achieved by non-paying users on X.
Meanwhile, the regulatory world has been turning all the outrage into paperwork with deadlines. In the UK, the communications regulator Ofcom has opened an investigation into X over Grok-related sexualized imagery. In the EU, the European Commission has ordered X to retain Grok-related documents until the end of 2026 — the bureaucratic version of telling a teenager, “Don’t delete anything. We’re coming back with questions.”
And then there’s the most direct form of platform feedback: simply pulling the plug. The Philippines is moving to block access to Grok on child-safety concerns, joining Indonesia’s temporary block and Malaysia’s restrictions aimed at X. Governments are saying the product is arriving faster than its guardrails, and they’re not interested in beta-testing the difference.
And the pressure is now climbing out of the regulator inbox and into the app-store choke point. A coalition of 28 advocacy groups, including women’s rights and tech watchdog organizations, has sent open letters to Apple $AAPL and Google $GOOGL urging them to remove X and Grok from their app stores altogether, arguing that both platforms are profiting from the spread of nonconsensual, sexually explicit AI imagery and failing to enforce their own policies on intimate images and abuse. The campaign — “Get Grok Gone” — accuses the companies of enabling widespread “mass digitally undressing” of women and minors through Grok’s tools, adding that X’s move to paywall image generation does nothing to stop the underlying harm. Apple and Google haven’t publicly responded to the letters, even as senators in Washington have made similar demands.
xAI, for its part, has treated media questions with the sort of posture that plays well on X and poorly in court filings. When Reuters sought comment on the earlier reporting, xAI replied with its familiar “Legacy Media Lies.” The problem is that the line doesn’t function as an answer — and regulators, unlike quote-tweeters, can subpoena the receipts.
So yes, Grok is “done” editing people into bikinis — as long as you take the claim at face value, don’t test the edges, and don’t confuse a hard rule with a reliably enforced one. The internet, historically, isn’t great at any of those things.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
