An image hits your group chat: a politician in a weird pose, a disaster photo that feels slightly off, a product shot that's a little too perfect. Is it real? In 2026 you don't have to guess. There's a method — five checks, about two minutes, and it beats any single "AI detector" you'll find online.
First: read everything readable
Text is still the best tell. Zoom into license plates, street signs, shop fronts, T-shirts, labels — anywhere letters live. Generators have gotten scary good at faces and lighting, but they keep mangling written language into almost-words. A sign that says "STOPP" or a menu that dissolves into glyphs settles it on the spot. Malwarebytes puts garbled text at the top of its checklist, and honestly, so should you.
Second: check the physics
Reflections are the runner-up giveaway. Mirrors, sunglasses, windows, water — they should show a consistent version of the scene, and generated images regularly blow it, reflecting things that aren't there or missing things that are. Give hands a look while you're at it; extra or fused fingers haven't gone away. Then zoom to the boundaries — hair against background, jewelry against skin. That's where diffusion models smear.
Third: reverse image search
Drop the image into Google Images or Bing before you believe anything about it. You'll find out if it existed before, when it first showed up, and where. This catches both flavors of fake at once — synthetic images with no history, and the older scam of a real photo wearing a false caption. If a "yesterday" image first appeared in 2019, you're done, no forensics required.
Fourth: check the watermark
Here's the newest layer. Google DeepMind's SynthID is an invisible watermark woven right into the pixels, and it survives cropping, filters, brightness tweaks — even screenshots and re-compression. Thanks to a partnership announced this year, it now covers images from ChatGPT and DALL·E too. Tools like the Gemini app and OpenAI Verify can read the signal. When a suspicious image came from one of the big generators, this check can end the debate outright.
Fifth: run a detector, but don't worship it
Automated detectors compare color patterns, shapes, and textures against the fingerprints of generated imagery. Fine — use one. Just know the numbers: against cutting-edge 2026 models, accuracy runs around 60–70% in testing Facia reported. Against older Stable Diffusion-era stuff, they do much better. So treat the verdict as one vote, not a ruling. These things flag real photos, too.
The part nobody likes to say
No tool can guarantee an answer. None. That's exactly why the stack works — each check fails differently, so a fake that slips past one usually trips another. Text, physics, history, watermark, detector: five different trapdoors.
And don't skip the cheapest check of all — context. Who posted this? Is anyone credible corroborating it? An extraordinary image from a single anonymous account has earned your doubt before you've zoomed into a single finger. Two minutes of this won't make you unfoolable. It'll make you a much harder target than the person who just hit share.
Image: George Milton, via Pexels
