xAI Accused of Training Grok on Child Sexual Abuse Material — Victim Files Class Action

A federal lawsuit has been filed against Elon Musk’s xAI in the U.S. District Court for the Northern District of California. The plaintiff, identified only as Jane Doe, alleges that developers included known child sexual abuse material (CSAM) in the training dataset for the Grok AI model, after which the service began generating new synthetic images featuring her likeness. The case could set a precedent: it is the first time an AI company has been accused not merely of weak request filtering, but of directly using prohibited content to train a model.

Digital fingerprints and bypassing global NCMEC databases

The victim was abused at preschool age in the early 2000s. The material created then circulated online for decades, and its digital hashes were added to law enforcement registries and those of the National Center for Missing and Exploited Children (NCMEC). Standard web filtering systems use these hash sums to automatically block files without moderator involvement.

As Ars Technica reports, the woman was alerted to the appearance of new images by specialists at the Canadian Centre for Child Protection. According to the investigation, users on open web forums were generating illegal content featuring the faces of real victims.

The lawsuit highlights a key technological danger of generative AI: when a neural network takes a real face as a basis and synthesizes a new frame, the file receives a unique digital fingerprint. As a result, traditional NCMEC hash databases are powerless — they cannot automatically detect and block generated material.

Does Grok train on every generation? Legal nuances

Viral posts claimed that Grok allegedly sends every single generation result back to the training database. The legal documents use softer language, but the risks are confirmed by the platform’s official rules.

According to the X social network help center and the xAI user agreement, the company reserves the right to use:

• public user posts and media files;
• text prompts and images uploaded to chat;
• results generated by the model,

unless the user manually disables data collection in privacy settings or activates Private Chat mode. Lawyers argue that if prohibited images entered public access or chats, they could have been re-absorbed into the neural network’s training pipeline. However, the prosecution has not yet provided public technical evidence of the composition of xAI’s original dataset.

xAI’s existing filters and dataset purity concerns

Claims that xAI has no safety rules at all are false. The current xAI Acceptable Use Policy explicitly prohibits any sexualization or exploitation of minors, and the moderation system requires mandatory reporting to NCMEC.

However, the lawyers’ complaint concerns the initial data collection stage (pre-training). The startup’s official documentation states that the company cleans training datasets of violent scenes, but the public descriptions of training data filtering do not directly detail the mechanisms for filtering out CSAM and unauthorized intimate content.

The problem of machine unlearning

One of the most complex technical aspects of the case relates to neural network architecture. The model does not store images as files on disk — the information is transformed into billions of mathematical weights and connections.

Deleting the original file from a server is simple, but completely erasing the influence of a specific training example on a finished neural network requires complex machine unlearning procedures or fully retraining the model from scratch at a cost of tens of millions of dollars. The lawsuit alleges that xAI has not provided evidence of performing such procedures to clean Grok’s weights.

Scope of the problem: lawsuits over thousands of images and Musk’s response

This is far from the first legal dispute involving xAI’s generator:

• In March 2026, lawsuits were filed by the families of three girls whose teenage photos were used to create deepfakes;
• In August, a Wyoming woman alleged in court that her childhood photo was used to create more than 7,000 pornographic generations;
• According to estimates by the research group Center for Countering Digital Hate (CCDH), during the peak availability of Grok’s editing features, it generated approximately 3 million sexual images in just 11 days, of which about 23,000 may have depicted minors.

xAI itself denies deliberate facilitation. In July, the company filed a countersuit against a user who created prohibited content, stating that over the past year it had blocked more than 52,000 accounts, sent over 73,000 reports to NCMEC, and helped law enforcement arrest at least 244 offenders.

Class action demands

In a statement from the law firm Girard Sharp, it is noted that Jane Doe is seeking nationwide class certification on behalf of thousands of potential victims.

The lawsuit relies on the federal Masha’s Law. The plaintiff seeks monetary compensation, the forced destruction of all copies of prohibited content on xAI’s servers, and the implementation of strict technological barriers to prevent such data from re-entering AI pipelines. xAI representatives had not provided official comments on the new lawsuit at the time of publication.