Generative image models stopped making obvious mistakes a while ago. Six fingers and melted text were the tells of 2023. In 2026 the giveaways are subtler — and they are mostly about physics, not anatomy.
Here are eight checks that still separate a real photograph from a generated one, roughly in order of how often they catch it.
Light obeys rules. In a real scene, every shadow in the frame comes from the same light source. Generated images frequently light a face from the left and cast its shadow to the left as well. Look at the ground, then look at the objects standing on it.
Wet streets, windows, sunglasses, spoons, water. A reflection should contain a plausible version of what is in front of it. Generators typically invent reflective noise — streaks that look like a reflection but are not an image of anything.
Tiles, window grids, brickwork, fabric weave, crowds. Real materials vary; generated ones loop. Scan a building facade or a patterned shirt and ask whether any motif repeats exactly.
Still one of the most reliable tells. Real signage has consistent letterforms; generated text has letters that are nearly right — wrong stroke weight, a missing serif, characters that change style halfway through a word.
Real photos have optical imperfections: slight chromatic fringing, lens falloff, motion blur, sensor noise that increases in shadow. A generated image is often uniformly sharp, with edges that never bloom and shadows that contain no noise at all.
Less reliable than it used to be, but still worth a look — particularly where a hand grips an object, and where hair meets a busy background. Generators blend the boundary instead of resolving individual strands.
Staircases that lead nowhere, doorways at impossible angles, chair legs that do not reach the floor, a horizon that tilts between the left and right side of the frame. Slow down and trace the structure.
The most robust check is not visual. A real news or travel photograph usually has a source, an author, a date and a licence attached somewhere. If an image cannot be traced to any origin, treat that as evidence.
Human detection is drifting towards coin-flip accuracy on high-end generations. That is exactly why this is a game rather than a security control: the point is to train your eye, not to trust it as a last line of defence. For anything that matters, look for provenance — where the file came from, who published it, and whether an independent source has it.
Play today’s challengePart of OMC — Network · Compute Forum · AI Top
⚡ OMC gems & token — coming soon