Fictional images of Donald Trump being arrested spread across social media after Eliot Higgins generated them with Midjourney v5. The pictures were not evidence of an arrest, but their reach showed how photorealistic AI images can travel beyond the context in which they were made.
From image experiment to viral story
On Monday, March 20, the possibility of Trump’s arrest was a prominent topic on social media. Higgins, founder of the Dutch investigative organization Bellingcat, used Midjourney v5 to create scenes imagining what such an arrest might look like.
He generated 50 images. A tweet showing two of them had been viewed well over five million times, and media outlets picked up the pictures, extending their reach. Higgins said he had been making images while waiting for the possible arrest, not setting out to make a pointed critique. Still, he said the material “kind of took on a life of its own.”
The images contained visible oddities, including an unusual depiction of legs. Higgins believed those flaws, along with his comments, made clear that the scenes were not real. But once the pictures circulated widely, viewers could encounter them apart from that explanation.
Why the images were harder to spot
Midjourney v5 produced more photorealistic images than earlier versions. One familiar clue to AI generation—the inaccurate appearance of hands and fingers—had also improved significantly. As those obvious mistakes become less common, a viewer may have fewer quick visual cues for judging whether an image is authentic.
This creates a challenge for social and editorial media. An image can look plausible while depicting an event that never happened. If it is reposted without its original context, the explanation that it was generated as an experiment may not travel with it.
The incident does not establish that every viewer believed the images. It does show that fictional political imagery can attract enormous attention and be picked up by media outlets before platform rules or corrections can contain it.
A single blocked word has limits
After the pictures circulated, Higgins suspected he had lost access to Midjourney. Buzzfeed News had not received a response from the company, but one sign of a reaction was that “arrested” had been added to Midjourney’s list of banned words.
That change did not resolve the broader policy question. Midjourney’s community guidelines broadly prohibited images that were “inherently disrespectful, aggressive, or otherwise abusive” and “adult content or gore,” but did not specifically mention political content. The company had not commented on the incident or whether it might tighten rules for generating images of political figures.
A word filter can be bypassed through a different description: “Trump in handcuffs” remained possible after “arrested” was blocked. And the word can serve legitimate image-making purposes, which prompted criticism of a blanket ban from a Midjourney community member. By the time a prompt restriction takes effect, images may already have reached a large audience.
Trust becomes a wider problem
Midjourney was not the only route to realistic synthetic images. The source article also points to open source solutions like Stable Diffusion, suggesting that limits imposed by one service cannot by themselves prevent fake images from being made or shared.
As photorealism improves, the pressure on image generators to address misuse is likely to grow. But the Trump images illustrate why moderation is difficult: a restriction may be too narrow to stop alternate prompts, while a broad ban can interfere with legitimate uses.
The episode also echoes a prediction made in 2017 by Ian Goodfellow, inventor of Generative Adversarial Networks (GANs): people may no longer be able to trust images and videos on the web. The practical challenge is not only creating rules for AI tools. People also need to know when an image is synthetic and how to assess it when it appears outside its original context.