Some social media users pointed out differences between the mugshot of Charlie Kirk’s suspected shooter and a supposed FBI “person of interest” image, raising questions about whether officials really have the right person.

The problem? The supposed FBI photo you may have seen online was altered with artificial intelligence.

On Sept. 12, authorities released the mugshot of Tyler Robinson, a 22-year-old Utah resident, who they say shot and killed Kirk Sept. 10 during an event at Utah Valley University campus. The day before the FBI released images of a “person of interest” from a stairwell security camera after the shooting.

Some were not convinced this is the same person. 

“The guy on the left with the glasses is not Tyler, okay. These are two different people here, the mouth is different, the chin is different, the face is shaped different, the nose is sharp on the left, his nose is sharp, his lips are different,” says a Sept. 12 TikTok post that shows an altered FBI image and a mugshot of Robinson. “I mean c’mon, he even has a Michael Jackson chin and this one almost has that chin, but he doesn’t, there are two different people here.”

Other TikTok and X posts also shared altered images to make a similar point.

The person in the supposed FBI image is not Robinson because it was altered by X’s AI-powered chatbot Grok.

On Sept. 11, multiple X users said they asked Grok to enhance some of the FBI images for a side-by-side comparison. But Grok did not merely enhance some of the facial features — the tool created images of a different person. 

PolitiFact also noticed that Grok distorted the design on the suspect’s shirt. Grok turned the design of a U.S. flag with what appeared to be an eagle, to a white patch with random shapes and patterns; it also distorted the shape of his black sunglasses.

The requests got even more bizarre after a user asked Grok to make one of the images better quality. Grok turned the man into a lookalike for Venezuelan President Nicolas Maduro. 

The confusion did not stay on X. The Washington County Sheriff's Office in Utah shared one of the “enhanced” images with a Grok watermark Sept. 11  with the caption, “Much clearer image of the suspect compared to others we have seen in the media." The office later edited the post saying the image appeared to be an AI enhanced photo.

PolitiFact asked an AI expert why Grok struggled with requests to enhance the FBI’s images..

“Despite advancements in the technology, AI models often struggle with enhancing images because they repeat patterns and biases from the content they are trained on,” said Krystyna Sikora, a research analyst for the Alliance for Securing Democracy at the German Marshall Fund.

If a model is trained on low-quality or distorted images, the model will produce the same unrealistic or over-processed results when asked to enhance photos. Sikora said some of these results are over-smoothing, removing fine details and textures and intensifying lighting in an unnatural way.