Grok's Spicy Mode Sparks Deepfake Fears After Topless Taylor Swift Video

Grok Imagine and Spicy Mode

xAI rolled out Grok Imagine with a feature labeled 'Spicy Mode' that allows SuperGrok and Premium+ users to turn text prompts into stylized visuals and 15-second audio-animated videos. The filter is pitched as an edge-case creative tool, but in practice it can produce explicit, blurred NSFW imagery, including partial nudity and sexualized content.

The Taylor Swift deepfake test

A Verge reporter tested Grok Imagine by prompting a generic description of 'Taylor Swift celebrating Coachella' with Spicy Mode enabled. The result was a generated clip showing a topless figure resembling Swift dancing in a thong, even though the prompt contained no mention of nudity. The output raises immediate concerns about how the tool handles recognizable public figures.

Policy and safety gaps

xAI states that its acceptable use policy bans explicit depictions of real people, yet Spicy Mode appears to override that rule. There is no robust age verification or identity safeguard in the flow: users only face a basic confirmation tap before receiving potentially sexualized imagery. Compared with other major AI image services, Grok lacks celebrity filters and deepfake protections that could prevent misuse.

Legal and ethical fallout

This is not only a technical misstep but a social and legal one. High-profile targets like Taylor Swift have been repeatedly victimized by deepfake content, and tools that make similar results trivial to create worsen privacy and reputational harms. In the US, pending legislation such as the Take It Down Act aims to address deepfakes and nonconsensual imagery, and outputs like these could trigger legal challenges.

Scale multiplies harm

XAI claims Grok generated over 34 million images in a matter of days. At that scale, even a small percentage of problematic outputs can lead to a flood of nonconsensual, deceptive, or defamatory material. Musk frames Grok as an 'unfiltered' creative platform, but freedom without guardrails allows easy weaponization of another person's image.

Why this matters

Testing the boundaries of generative systems can be valuable for research and design, but when a feature makes sexualized depictions of real people effortless, the ethical calculus changes. This episode highlights the need for built-in safeguards, clearer enforcement of acceptable use policies, and better preemptive design choices to prevent misuse. Regulators and platforms alike will likely need to act quickly to close these gaps.