Signs of AI-generated images in 2026
Two years ago you looked at the hands. That no longer works. Here are the signs that still give AI images away, and an honest explanation of why the eye is losing the race.
Updated 2026-09-11 · Veriflied
What no longer works
Six fingers, melted faces, impossible teeth and gibberish text in the foreground were 2022-2023 mistakes. Modern models produce correct hands, legible text and photorealistic faces in most attempts. Checklists from that era give false confidence: an image without the old mistakes is not therefore genuine.
What still gives it away
Small background text
Foreground text has become good; signs, labels and newspapers in the background still misspell or repeat characters.
The details of physics
Reflections showing the wrong thing, shadows with separate light sources, liquid and glass behaving wrongly.
Patterns and crowds
Crowds with repeated faces, brickwork losing its rhythm, fabric where the weave changes direction.
Uniform micro-texture
Real photos carry sensor noise that varies with light; AI images often show a strikingly even smoothness across the whole frame.
The margins
Models spend their effort in the centre. Corners and edges more often hold smeared detail and half-finished objects.
Jewellery, logos, text on clothing
Small structured details with rules (chain links, letters, symmetric buckles) still frequently break.
Why the eye loses
Every model generation removes the mistakes people have learned to look for, because those are exactly what the models are trained away from. Studies consistently show humans performing close to chance on modern photorealistic AI images, even when warned in advance. That does not make the list above worthless — it still catches sloppy attempts — but it cannot carry a decision alone.
What the eye cannot see, statistics can: generator pixel patterns are measurable even when the image looks perfect. That is what a detector does, and why human + detector + metadata is the standard in professional verification.
Frequently asked questions
Can you still spot AI images by eye?
Sloppy attempts, yes. Careful attempts with modern models, rarely. Expect the eye to catch the bad ones and miss the good ones.
Which subjects are hardest to judge?
Portraits and product shots — constrained subjects the models have seen millions of. Complex scenes with many people, text and physics remain harder for the models.
Will these signs keep working?
No, the list has a shelf life of months to a few years. That is why detectors are retrained continuously while checklists quietly expire.
Try it on your own document
Create an account and analyse your first 10 images or documents for free. You get a report with the concrete findings, not just a score. All processing in the EU.