After X updated Grok to ban sexual deepfakes of women, tests show Grok still undresses men on Grok's app, Grok's website, and X, rarely rejecting user requests
Robert Hart /The Verge:
Context & Ripple Effects
This follows X's January change to block edits of real people into revealing clothing and geoblock the capability where illegal, a policy response that narrowed the stated use case without resolving access across Grok's surfaces. Earlier reporting had also found that restrictions on the @grok command could be bypassed through image editing and Grok's own app via alternate Grok entry points.
The new testing makes enforcement consistency—not merely the existence of a rule—the central issue. That matters because prior coverage documented Grok being used to alter images of real women into abusive and humiliating depictions for abusive sexualized imagery.
First-order effects
- Grok's updated safeguard appears to leave a readily available path for requests to undress men on its app, website, and X, so users face little immediate friction for that category of request.
- X and Grok now have evidence that the policy is not being applied consistently across product surfaces, putting their image-generation moderation controls under renewed scrutiny.
Second-order effects
- A surface-by-surface enforcement gap makes command-level restrictions less meaningful: users can migrate to whichever interface remains permissive, echoing the earlier bypass through image editing and the app.
- The mismatch between the announced restriction and observed outputs raises the operational burden of testing prompts, applying safeguards uniformly, and demonstrating that geoblocking and policy controls work as intended.
Third-order effects
- If similar gaps persist, safety claims for consumer image models will increasingly be judged by adversarial testing across every distribution surface rather than by a single published policy.
- The episode points toward a tougher product-governance standard for generative-image features: providers may need controls that cover both the model behavior and the interfaces that expose it.
The trend: Generative-image platforms are moving from announcing narrow safety rules to being evaluated on whether those rules hold consistently across models, apps, web products, and social integrations.