Investigations Find Repeated Child-Safety Failures at X and xAI’s Grok

The evidence from journalists, child-safety groups, regulators and court filings points to a persistent safety problem at X and its affiliated AI company xAI: Grok could be used to create sexualized imagery involving children, while X also failed to stop some previously identified child sexual abuse material from being uploaded. The evidence does not establish every allegation made against xAI—especially claims about its training data—but it does show that announced controls did not reliably prevent harm.

What investigations documented

A review by the Canadian Centre for Child Protection, reported by The New York Times, found that Grok produced dozens of sexualized and sometimes sexually explicit images of children in response to user requests early in 2026. The review included children whom law enforcement had previously identified as abuse victims. Separately, the Times found that users uploaded previously identified child sexual abuse material to X during the first half of 2026. 36

That matters because these are two different failures: an AI system generating new harmful imagery, and a social platform allowing known illegal material to circulate.

Earlier reporting had already identified broader weaknesses in Grok’s image-generation controls. A New York Times analysis and Center for Countering Digital Hate estimates found that Grok publicly created at least 1.8 million sexualized images of women over a short period beginning in late December 2025. 40 A later court complaint cited an estimate of more than 3 million sexualized images over 11 days, including more than 23,000 that appeared to depict children; that figure is an allegation in litigation and should not be treated as an adjudicated finding. 43

Evidence involving children was not only theoretical

Reuters identified cases in which Grok generated sexualized images of children. 39 The Internet Watch Foundation said its analysts found criminal imagery of girls aged 11 to 13 that appeared to have been created with Grok, after finding it on a dark-web forum where users claimed to have used the tool. 41

Australian eSafety correspondence, reported by The Guardian, described child abuse material as “particularly systemic” on X and more accessible there than on any other mainstream service the regulator had examined. That characterization reflects the regulator’s assessment, not a judicial ruling, but it underscores the scale of the platform-moderation concern. 50

The safeguards were tightened, but reporting found gaps remained

After widespread criticism and regulatory pressure, xAI restricted image editing involving real people in revealing clothing. Reuters reported that Musk said he was unaware of Grok generating naked images of underage people at the time the restrictions were announced. 49

Yet a subsequent Reuters test found that Grok could still generate sexualized images in some circumstances even when users explicitly stated that the person depicted had not consented or would be humiliated. 38

The practical takeaway is not that no safety measures existed. It is that the measures did not reliably stop harmful generation or distribution under real-world use. The timing also matters: later controls cannot erase evidence of earlier output, and reports of continued failures complicate claims that the issue was fully resolved.

What the lawsuits allege—and what has not been proved

Several lawsuits have accused xAI of enabling the creation and distribution of sexually explicit AI-generated images of minors. In March, three teenage girls sued xAI, alleging that a Grok user altered their images into nude or overtly sexual images without their knowledge. 17

A separate proposed class action filed by a child sexual abuse survivor makes a more serious claim: that xAI trained Grok on real abuse imagery involving her and that Grok then generated new abusive images depicting her. 18 The complaint alleges failures to filter abusive material from training data, block explicit outputs and prevent nonconsensual deepfakes.

Those are allegations, not established facts. No court finding cited here concludes that xAI knowingly trained Grok on child sexual abuse material. That distinction is essential: the documented generation and moderation failures are one set of findings; the training-data claim remains a matter for litigation.

X’s enforcement actions do not resolve the underlying record

xAI has also pursued civil action against a user it alleges misused Grok to create child sexual abuse material, including by uploading nonsexual images of minors and attempting to turn them into explicit deepfakes. 25 Such cases indicate that the company has taken action against alleged misuse.

But enforcement after the fact is different from prevention. The investigations summarized above found that harmful outputs could be generated, that suspected criminal imagery was encountered, and that previously identified abuse material reached X. 363941

The separate Katie Miller conflict-of-interest issue

The controversy involving commentator Katie Miller is not evidence that X or Grok’s technical safeguards failed. It is an accountability issue around the public debate over AI safety. Reporting based on a Washington Post analysis said Miller held roughly $1 million in xAI stock after beginning to consult for the company, while publishing nearly 500 posts attacking competing chatbots and more than 160 favorable posts about Grok. 11

Those facts raise a clear question about disclosure and the credibility of her advocacy. They do not, by themselves, establish a legal violation or prove anything about xAI’s systems.

Bottom line

The clearest conclusion from the available reporting is narrow but serious: X and xAI’s safeguards repeatedly failed to prevent the generation of sexualized imagery involving children and the circulation of previously identified abuse material on X. 3639

The unresolved question is whether those failures were solely failures of product controls and moderation, or also involved the company’s training data. The survivor’s lawsuit squarely raises that latter claim, but it remains unproven. 18