Published on 2026-07-25
How can you actually tell if an image is AI generated?
Odd hands and garbled text still catch some fakes, but generators are closing that gap fast. The real proof is where the image came from, not how it looks.
Look at the hand first. That's the advice you've probably heard, and for a while it worked beautifully: six fingers, a thumb bent the wrong way, a wedding ring floating over nothing. Generated images used to give themselves away in the most human part of the picture. We still check hands. We just can't stop there anymore, because the tools that make these images have spent two years getting better at exactly the thing that used to expose them.
A library research guide from the University of Arizona puts it plainly: these visual tells keep working less often as the technology improves. That's not a reason to stop looking. It's a reason to know which clues still hold up, and to remember what they can and can't prove.
What still shows up, some of the time
Hands remain a decent bet, particularly in a busy scene the generator wasn't paying full attention to: fingers that merge, a hand that trails off into blur, a ring that doesn't quite wrap around anything. Faces further from the camera are another weak spot; a crowd in the background of an otherwise sharp photo can turn into smudged, slightly wrong features once you zoom in. Text is often the easiest catch of all, because generators still don't reliably know what letters mean. A street sign, a book spine, a logo on a cap: look closely and the words dissolve into shapes that almost read, then don't.
- Hands and fingers: extra digits, fused fingers, joints that bend at angles a hand doesn't bend at
- Text in the scene: signage, labels or book covers that look like letters but spell nothing
- Reflections and shadows: a mirror, window or puddle showing something the light in the room doesn't support
- An oddly glossy or smoothed surface, especially on skin, hair or water
- Symmetry that's too perfect: earrings that match exactly, a face with no natural asymmetry at all
- Background objects that repeat, blend into each other, or don't quite make physical sense
The Associated Press recently walked through the same territory and landed on a similar note of caution: the classic errors, too many fingers, mismatched jewellery, are getting rarer, and what's left is subtler. An object that seems to fade at its edge. A shadow that falls the wrong way for where the light source sits. A sheen across a whole image that no camera lens quite produces. None of it shouts anymore. You have to be looking for it.
A clean image proves nothing
Here's the part that trips people up: passing all six checks on that list doesn't mean a photo is real. Plenty of ordinary phone photos have a weird reflection, an oddly smooth patch of skin from a beauty filter, a shadow that looks strange because of a streetlight you can't see in frame. And plenty of AI images, made with the newer tools, will sail through every single tell without a hiccup. Absence of a flaw was never proof of authenticity. It was proof that this particular generator, on this particular day, didn't slip up in a way you happened to notice.
That's the trap with any checklist: it turns verification into a game of spot-the-glitch, and glitches are the one thing a generator's next update is specifically built to remove. The people building these tools are not trying to keep the hands wrong. They're trying to fix them, and they largely have.
A photo that looks clean earned a maybe. It didn't earn a yes.
So what actually settles it
The University of Arizona guide points to the move that matters more than any visual scan: find out where the image has already been. A reverse image search, dragging the picture into Google Images or a similar tool, shows you whether this photo has a history: an earlier posting, a different caption, a stock photo listing, a completely unrelated context from years ago. A genuine photo of a real event usually has company. Other angles. Other accounts. A place it was first published. An AI image invented for the occasion has none of that, because there's nothing to find; it was never anywhere before the moment someone generated it.
So the order of operations flips from what feels natural. Instinct says: stare harder at the picture. The better move is to stop staring at the picture and go looking for its past instead. Where did it first appear? Who else is showing it? Does it hold up next to real photos of the same place? A generated image can nail the hands, the text, the shadows, and the symmetry all at once. What it can't fake is a paper trail that doesn't exist.