The mechanism in the filing is a loop. Gizmodo, reading the complaint, reports that xAI policy treats anything posted publicly on X and any output Grok produces as training-eligible, so an image that survives on the timeline for even a short window becomes a candidate row in the next corpus [9]. Output moderation, the part most teams actually instrument and report on, sits downstream of that. It removes the artifact and leaves the ingestion decision untouched.
The persistence argument is where this lands on engineering rather than trust and safety. The complaint says that once such material entered training, its influence likely persists into future outputs, so the harm is not resolved by removing individual images or unposting them [10]. Deleting a file is a ticket. Producing an account of which checkpoints were trained after the material arrived is different, harder work, and most teams have never been asked for it.
The provenance question here is not exotic. Doe was preschool-age in the early 2000s when men raped her to make CSAM for sale online, according to her complaint [3], and her images were hashed afterward by the National Center for Missing and Exploited Children and the Canadian Centre for Child Protection [4]. The filing describes material that has followed her more than 20 years [6], which puts roughly two decades of hash coverage ahead of the AI-generated versions the CCCP told her it had identified on xAI [20]. Sarah London of Girard Sharp, Doe's counsel, put the claim in terms of knowing training on survivors' abuse images [17].
A product deck might describe users of a generation feature as prompting, looking, and discarding. The record in these filings describes something else. Musk promoted the nudify capability himself, millions of nonconsensual sexual deepfakes went onto the X timeline, and a large share of the victims were children [14]. Ars Technica reports the complaint citing forum messages between offenders discussing how to make AI CSAM of Doe and other known earlier victims [7]. The growth rationale is now an exhibit: the complaint alleges Grok was built to answer sexual-content prompts in order to entice more users of both X and Grok [16].
The two publishers describe her notification path differently. Gizmodo says Doe is tracked by the FBI's Child Exploitation Notification Program [c19a]; Ars says she opted into alerts from the Justice Department's Victim Notification System [c2b]. Neither has seen a corpus, and both are describing the same complaint.
xAI has said, in its own suit against two users who made CSAM with the tool, that it helped arrest at least 244 people who created or distributed such material through Grok [13]. Counting to 244 requires records of what was produced. That is the shape of the forcing function for anyone shipping image generation next quarter: take one image your product removed last month, and see whether you can name the training runs that saw it and say what became of the checkpoints trained afterward. Answering the first question means you have logs; answering both means you have provenance, which is what this complaint is asking about [8].
