Is That Image Real? A Broker’s Field Guide to Spotting Fake Photos
A manipulated truck photo, a screenshot of a document passed off as the real thing, an AI-generated “proof” of a load, fake imagery is cheap, fast, and getting more accurate everyday. With the right tools and the right questions, you don’t need a degree in forensics to spot them.
Google has been doing serious computer vision work for well over a decade, and in practice their Gemini models are the strongest at this kind of image forensics. Arm yourself with a few key terms to turn a basic prompt like “is this real?” into a real assessment.
A handful of vocabulary words dramatically improve what the model gives back:
Spoof / second-generation image: a photo of a photo, or a photo of a screen, rather than a direct capture. Tells include screen bezels, a pixel grid, refresh lines, a rectangular border framing the content, a moiré pattern (those wavy, rainbow-like artifacts you see when you photograph a screen), or glare bands that don’t match the geometry of the scene.
Lighting and shadow consistency: Ask whether shadows fall in a direction consistent with the light source and whether reflections make physical sense. AI-generated images get this subtly wrong: shadows pointing different ways, impossible reflections, unnaturally clean or slightly warped text, edges that seem to float.
Confidence score: Instead of forcing a black or white answer, ask the model how confident it is and why. Making it show its work tells you what it actually saw. Keep in mind that a model that flags every photo as fake is technically right about the fakes but it’s useless to you. Push for a score and not a verdict.
Next time an image looks off, paste it into Gemini with this:
You are an image forensics analyst. For the image I provide, tell me: (1) Is this a direct photo of a real object or scene, or a “second-generation” image. Look for screen bezels, a pixel grid, moiré patterns, refresh lines, a rectangular border around the content, or glare that doesn’t match the scene. (2) Are there signs it was AI-generated or edited? Look for shadows or reflections inconsistent with the light source, unnaturally clean or warped text, mismatched blur, or floating/duplicated elements. (3) Give a confidence score from 0 to 1 for each and a one-line reason. Important: a merely blurry or dark photo is not a fake, only flag indicators you can actually see.
AI image generation gets better every month, and detection will never be perfect. Treat it as one signal, not a verdict. Combine it with the context around the image. Was the person responsive in conversation, then suddenly took 30 minutes to “send a quick photo”? Does the story hold together? The model tells you what’s in the pixels. Your judgment about everything around them is still the most important tool you have.
