Grok Continues Generating Non-Consensual Sexual Deepfakes Despite xAI's January Pledge to Stop
Summary
- • Grok still produces sexualized deepfakes of real women despite xAI's January 2026 promises
- • Users evolved tactics — pose-matching and costume prompts — to bypass Grok's filters
- • Dozens of AI-generated sexual images of celebrities found publicly on X over the past month
- • xAI took down most flagged images only after journalists presented findings for review
Details
Grok is still generating sexualized deepfakes of real women a month after xAI's January restrictions
Dozens of AI-generated sexual images and videos of real women — mostly celebrities — were found posted publicly on X over the past month. This directly contradicts xAI's pledge to halt abusive deepfakes following January's backlash and government investigations.
January 2026 'spicy mode' incident prompted global government investigations across five continents
Grok previously allowed users to undress anyone in a photo via simple text prompts. The resulting flood of non-consensual images, including some of minors, drew regulatory scrutiny worldwide and forced xAI to publicly promise new safeguards.
Three user tactics now bypass Grok's filters: pose-matching, clothing swaps, and costume prompts
Users submit a celebrity photo alongside a stick-figure drawing in a sexual pose and ask Grok to 'match the pose'; or request clothing swaps between two photos; or ask Grok to dress women in tight costumes like 'Spider-Woman' or 'bunny' outfits. All three methods evade the platform's keyword-based undressing restrictions.
Workarounds were easy to find via X's public search — not obscure exploits
The examples required no special technical knowledge or access to discover. The tactics were circulated openly on X, suggesting broad user awareness and adoption of evasion methods.
Experts warn publicly visible deepfakes are likely a fraction of total Grok output
Most Grok usage occurs privately through the Grok app, Grok website, or X's private Grok tab, making it impossible to audit the full scale of non-consensual image generation. Public posts on X represent only the most visible layer of the problem.
xAI removed flagged images and issued a policy statement only after journalist inquiry and publication
When presented with findings, xAI said it wanted to review them but did not respond to follow-up questions. Most images disappeared from X by the following day. After publication, X issued a statement citing 'continuous monitoring' and 'real-time analysis of evasion attempts' — safeguards that had not prevented the documented violations.
Volume of deepfakes has decreased since January but has not stopped
Grok now turns down or ignores many sexualized requests. None of the images found in the April 2026 review depicted nudity or minors. However, the continued production of sexualized images without consent demonstrates that restrictions remain incomplete.
RAINN: victims often never learn their likeness was used to create sexual images
Stefan Turkheimer, VP for public policy at RAINN, emphasized that the harm is compounded by the invisibility of the violation — women depicted in deepfakes frequently have no way of knowing the images exist or are being shared.
Industry Update = platform behavior vs. stated policy; Context = background on the January 2026 incident; Tech Info = how evasion methods work; Security Alert = active ongoing harm; Insight = expert analysis; Policy = company response; Stat = quantitative observation
What This Means
xAI's inability to prevent Grok from generating non-consensual sexual deepfakes — even after public pledges, government scrutiny across five continents, and updated filters — reveals a fundamental weakness in reactive AI content moderation: determined users iterate faster than platform engineers when the underlying model retains the problematic capability. The transparency deficit is equally significant: xAI's safeguards went publicly untested until NBC News journalists intervened, and the company's formal response came only after publication. For AI governance, this episode illustrates why voluntary corporate pledges without independent, continuous auditing provide weak protection against foreseeable and recurring harms.
Sentiment
Mostly critical from safety advocates, human rights experts, and tech reviewers
“Despite promises from X to curb abuse, Elon Musk’s Grok AI is still being used to generate sexualized deepfakes. A recent review by @NBCNews found dozens of explicit AI images of real women... This is exactly why NCOSE continues to call for real enforcement and accountability”
““Guardrails”? You mean making it a premium feature so you can profit from CSAM and revenge porn.”
Responding directly to X's official Safety account defense
“.@elonmusk is gross.”
“NBC News article on Grok continuing to generate sexualized deepfakes despite pledges”
Split
~80/20 critical (safety orgs, reporters, experts alarmed at harms)/skeptical (media bias, happens elsewhere, user misuse)