The AI image that proved what you already believed

The Pattern

There are two ways to lie with an image. The old way is to take a real photo and lie about what it shows. The new way is to make the photo itself. In 2025 and into 2026, the second method stopped being a novelty and became a workflow. Fact-checkers at Misbar rounded up a year of it: AI-generated images tied to political, humanitarian, and security events, circulated as real, often in war-torn areas where verification is hardest and the appetite for confirmation is highest.

Two examples frame the problem. An AI-generated image showed mixed martial artist Sean Strickland wearing a T-shirt linking US President █████ and Jeffrey Epstein as he was escorted from a UFC event at the White House. It was not real. In early 2026, social media was flooded with AI-generated images of Venezuelan president Nicolás Maduro’s capture that never happened, blending fabricated visuals with genuine footage until millions of viewers could not tell which was which.

Why a fake image works better than a false caption

A miscaptioned real video has a weakness: the real footage exists somewhere in an archive, and a reverse image search can find it. An AI-generated image has no archive. There is no original to match it against, because it was never photographed. The only evidence of its falseness is in the generation artifacts, the watermark, or the absence of corroborating coverage, and all of those require more effort than a scroll.

This is the structural shift. The old lie borrowed truth’s body and changed its name. The new lie manufactures the body too. And it is tuned to what the audience already suspects. The Strickland image worked because the █████-Epstein association is something a portion of the audience already believes and wants confirmed. The Maduro images worked because people who wanted to believe his regime was collapsing were given a picture of exactly that. A fake image does not have to persuade anyone. It only has to confirm.

The liar’s dividend

The deeper cost is not the fake image that goes viral. It is what the existence of fakes does to the real ones. Once it is plausible that any image could be AI-generated, every inconvenient real image can be dismissed as AI. This is what researchers call the liar’s dividend: the same technology that lets anyone fabricate evidence also lets anyone deny genuine evidence by claiming it was fabricated.

The Maduro case is instructive in the other direction. When a real video of Benjamin Netanyahu surfaced, fact-checkers at NewsGuard had to do extensive verification, including matching the background to Reuters stock footage of the café, to establish it was real, because the default assumption had shifted toward suspicion. The burden of proof has moved. Real things now have to prove themselves, and fakes only have to cast enough doubt.

How to read an image now

The habits for AI images are different from the habits for recycled footage, because the archive trick does not work. There is nothing to reverse-search back to. Instead:

• Look for corroboration, not just the image. If something consequential happened, more than one source will have it. If only social accounts have it and no news organization does, treat the absence as the evidence.
• Check for generation artifacts. Hands, teeth, text on signs, repeated patterns, and the geometry of ears and glasses are still where generators fail most visibly. These are not proof, but they are flags.
• Look for provenance, not just content. Who first posted it, and when? An image that appears simultaneously across many anonymous accounts with no original poster is a distribution pattern, not a sighting.
• Assume confirmation bias is the product. If an image shows exactly what one side of a conflict most wants to see, and it shows it perfectly, that perfection is the tell. Real footage of real events is usually partial, blurry, and badly framed. Fakes are too clean.

The honest conclusion

The threat from AI images is not that they will fool everyone about everything. It is that they will make verification feel futile. The goal of a disinformation campaign using fabricated images is not always to make you believe the fake. Often it is to make you unsure enough about everything that you stop trying to tell the difference. That is the condition the liar’s dividend produces, and it is the one worth resisting. The difference between a real image and a fake one is still recoverable. It just takes more work than the scroll, and the willingness to do that work is the thing being attacked.