Six months after xAI pledged to clamp down on nonconsensual deepfakes, its Grok Imagine system continues to generate and host sexualized imagery of public figures—including at least one U.S. representative. A WIRED review of hundreds of publicly accessible Grok.com links found dozens of AI-generated images and videos depicting celebrities, politicians, and potentially minors in explicit scenarios. Some outputs were photorealistic; others used cartoon styles. Critically, these assets remained hosted on xAI’s infrastructure and were shared on X, indicating that content moderation pipelines either failed to catch them or were bypassed by prompt engineering techniques (e.g., indirect descriptions of sexual acts).

From an engineering perspective, the core issue is architectural: Grok’s safety layer appears to rely on prompt-level heuristics rather than robust post-generation classification or embedding-based similarity checks. Unlike OpenAI’s DALL-E or Google’s Gemini, which reject prompts describing nonconsensual scenarios outright, Grok’s “Spicy” and “Unhinged” modes deliberately reduce guardrail strictness in favor of output variety. This trade-off—prioritizing creative freedom over content safety—creates a high false-negative rate for harmful generations. Worse, the system stores results on public-facing endpoints without automatic takedown triggers, making distribution trivial.

xAI’s response has been reactive: after WIRED notified the company, the offending links were removed and X posts deleted. But the underlying architecture hasn’t changed. The Privacy Commissioner of Canada’s investigation notes that xAI has not demonstrated the effectiveness of its new safeguards. Meanwhile, SpaceX’s IPO filing discloses a $530 million reserve for litigation, including Grok-related complaints. For any team deploying generative AI at scale, this case underscores a hard lesson: safety guardrails must be baked into the generation pipeline—not bolted on after the fact.

Source: Wired