Grok Imagine's Spicy Mode: Access, Safeguards & Complaints

An image and video generator with an explicit-content toggle built in

xAI shipped Grok Imagine on August 4, 2025, as a feature inside the Grok iOS app for SuperGrok and X Premium+ subscribers. It generates images from text prompts and can animate a still image into a 15-second video with audio. Alongside the ordinary mode, the app ships a "Spicy" toggle that permits sexually explicit output, including partial nudity; the most explicit prompts still return blurred, moderated results rather than an outright block. [1][2]

The gate in front of Spicy mode was a self-reported birth year field in account settings, pre-filled to the year 2000, with no document check or other verification behind it. A coalition of 15 advocacy organizations that examined the feature two weeks after launch pointed to that gap directly, despite the Grok app carrying a 12+ App Store rating. [3]

Atlas interpretation: "Behind an age check" is doing a lot of work in that sentence. The check is a form field a user fills in about themselves, not a check against any external record, so the gate's practical effect is closer to a terms-of-service checkbox than an age-verification system of the kind more commonly required for age-restricted content elsewhere. [3]

Where the other frontier video generators drew the line

A Verge reporter tested Spicy mode the day after launch by uploading a photo and selecting the setting, without asking the tool to remove any clothing. It returned an uncensored topless video of the subject unprompted. Her report on the result put the contrast plainly: "While other video generators like Google's Veo and OpenAI's Sora have safeguards in place to prevent users from creating NSFW content and celebrity deepfakes, Grok Imagine is happy to do both simultaneously." [4]

Atlas interpretation: Every other lab shipping a comparable video model, Google's Veo and OpenAI's Sora among them, had already chosen to block sexualized output and celebrity likenesses as a matter of policy, not capability; the underlying diffusion and video-generation techniques do not differ enough between labs to explain the gap. xAI's decision to ship the opposite default, explicit content on by user choice rather than off by design, was the news, not the underlying model. [4][1]

Nonconsensual deepfakes and a coalition complaint within three weeks

The unprompted topless video the Verge reporter received was generated from her own uploaded photo, not a celebrity's, but the same failure mode was reported against public figures within days: Spicy mode could produce sexualized depictions of people, including celebrities, who had never consented to it and had not been named in the prompt. [4]

On August 18, 2025, the Sexual Violence Prevention Association and 14 partner organizations sent a formal complaint to the FTC and all 50 state attorneys general asking for an investigation into Grok Imagine. It alleged the tool violated state nonconsensual-intimate-imagery statutes in 38 states, the FTC Act's unfair-and-deceptive-practices standard, and the Children's Online Privacy Protection Act, given the unverified age gate. The Electronic Privacy Information Center joined a separate, related coalition letter to the FTC on the same subject on August 20. [3][5]

Atlas interpretation: The complaint arrived fast relative to the product, a matter of weeks rather than months, which is itself a data point: the harms the coalition described were not edge cases discovered through adversarial prompting but the tool's default behavior against an ordinary, non-adversarial input. That is a different kind of failure than the jailbreak stories that usually accompany a new generative model, and it is why the criticism centered on the product decision rather than on model safety training. [3][4]