How Users Tried to “Unblur” Photos from the Epstein Archive Using Grok

When images from the Jeffrey Epstein case first circulated on-line, they were reviewed, and discussed by the public very quickly. However, that review and discussion of the material quickly turned into a new subject of discussion – the potential of artificial intelligence (AI). The focus was on the chat bot Grok, which had been used to attempt to un-redact photos.

This action brought up many tough issues: If a technical solution exists for a problem, does that mean it has to exist? Is there a distinction between being curious, and violating someone’s right to privacy? Who would be responsible – the entity hosting the AI, or the individual using the AI? Here we will evaluate this situation and how it caused many people to reflect on their actions.

The Controversy Surrounding Grok and the Epstein Files

A week or so after the United States Department of Justice posted about 3.5 million pages of material concerning Jeffrey Epstein, requests to remove black boxes from photographs were posted all over X by several users. Black boxes are the same boxes used in other high-profile child and woman sexual abuse cases to obscure the faces of children and women in order to protect them.

While a number of survivors in this case chose to identify themselves as having been abused at the hands of Jeffrey Epstein, a number of others chose not to. In a joint statement, 18 people who suffered abuse expressed their outrage at the release of these files; they pointed out that the files include the names and personal information of many victims and that the men who committed those abuses continue to remain anonymous.

Within a couple of weeks of the public release of the files on January 30 (the day the files became available because of the passage of the Epstein Files Transparency Act), many of the files had to be removed from public view due to poor redactions that could potentially reveal the identity of nearly a hundred survivors of Jeffrey Epstein’s abuse.

Regardless of whether or not there were legitimate reasons for removing redacted portions of the files from public view, X users quickly began looking for ways to “undo” the redaction efforts. More than twenty photographs and one video, each containing images of children and young women whose faces were obscured using black boxes, were located with an X search that contained the terms “unblur” and @grok. Many more attempts to un-redact images probably existed, but never came to public light.

The pictures, according to reports, contained women, children and Jeffrey Epstein. Other known people who had been named as being connected to the case prior to the release of the pictures (such as Prince Andrew, Bill Clinton, Bill Gates and filmmaker Brett Ratner) were also shown in the pictures.

Pictures were taken at different places – on an airplane, by a swimming pool, etc.
From January 30 through February 5, we analyzed thirty one individual user requests for Grok asking to “Remove Blur” or to help the user identify the women and/or children in the pictures provided to Grok.

In some responses, Grok stated that the faces of minors were intentionally blurred to protect their identities, and therefore Grok could not restore or reveal their identities. However, in twenty-seven of the thirty one requests we examined, the system created a picture from the information provided.

To prevent the potential for further dissemination of these posts, we are not linking to the posts we reviewed.

The quality of the output of Grok was very poor to excellent. Many of the outputs appeared to be fairly realistic; however, many were significantly distorted or even absurd — such as attaching the face of a young child to the body of an older female. Although some of these posts received tens of thousands to millions of views on X, they received a disproportionate number of views because X financially rewards its creators when they create content that is extremely engaging.

In one instance, Grok produced no image; in another it acknowledged that it could not edit (i.e., remove blur) from photos nor would it do so. In addition to Grok’s inability to produce images to the above three requests, the Chatbot provided a clarification that the images from the Department of Justice releases were intentionally redacted for privacy purposes.

The other two requests were to what appeared to be “free” account users. The responses to both of their requests indicated that Grok could not produce images nor edit them because these functions are only available to verified paid account holders. Since January, X has restricted these image functions on free accounts due to controversy surrounding AI generating harmful content, including the use of so-called “digital undressing” of women and children.

Following an inquiry sent to X regarding this matter on February 6, the situation appeared to have changed. We identified 16 additional requests using the same search terms between February 7-9. Grok attempted to remove blur in none of the 16 new requests. Instead, Grok simply failed to respond in 14 of the 16 new requests. In the remaining two, Grok produced images that bore absolutely no resemblance to the images users had previously uploaded.

One user complained that Grok “no longer worked.” Grok responded that it was functioning correctly, however, it could neither unblur nor identify faces in photographs taken from the Department of Justice files. Grok explained that the faces of minors were blurred in the DOJ files for privacy reasons, and it is legally and ethically bound not to perform such functions. Additionally, Grok directed users to consult official sources for accurate information.

Artificially Generated Photos

AI-generated images related to the Epstein case have been circulating on social media platforms since long before the above-mentioned requests. Prior to the requests, AI-generated photos were circulating depicting Epstein alongside numerous public figures — including U.S. President Donald Trump and New York City Mayor, Zohran Mamdani, who at that time was a young boy accompanied by his mother. Some of the photos that were purportedly created using AI tools were created using those same tools.

It should be noted that while many of the people depicted in the AI-generated images have also appeared in real photographs published by the U.S. Department of Justice, the real photographs are publicly available on the Department of Justice website. Because there is overlap between real photographs and artificially generated images, the confusion arises as to whether a given photograph is real or AI-generated. Not all users can immediately discern the difference between real and artificially generated images.

Grok’s scandal rapidly grew into many separate scandals across multiple countries (the U.K., France, etc.). Following the reports that Grok had created countless images of people engaging in intimate acts without their knowledge or permission, the governments of both the U.K. and France began investigating Grok. Reports have surfaced that some of these images include minors which has prompted severe reactions from regulatory bodies. For example, Malaysia and Indonesia took action to ban Grok entirely citing that they were concerned about the proliferation of pornography in new formats (i.e., Altered Images and AI-generated Images).

In total, the Center for Countering Digital Hate reported that in less than two weeks (from Dec. 29, 2025 to Jan. 8), Grok had publicly produced nearly 3 Million images with a sexually-oriented theme. Approximately 23,000 of those images contained children. As public outcry against X continued to grow, X implemented some limited changes to Grok’s image creation and editing features in late January; specifically, X restricted those features to only its paid subscribers. Those changes were largely seen as too little, too late. In response to the actions taken by X, UK Prime Minister Keir Starmer noted that X had essentially made creating illegal images a “premium” option for its subscribers.

Following the widespread criticism, the company ultimately made additional changes to limit user access to Grok. Specifically, users — including paid subscribers — are no longer able to use Grok to alter images of real people in a manner that would either remove clothing from a person or otherwise depict them in an explicit way. Those changes were cited as efforts to reduce abuse and show that something is being done about the situation.

All information was obtained from publicly available reporting and articles posted at Bellingcat.