How to tell if an image is AI-generated

For a couple of years the advice was easy: count the fingers. That era is over — current image models draw hands, teeth, and pupils just fine, and anyone still relying on anatomy is exactly who the next fake will fool. What follows is the checklist we actually use when building the daily visual round, ordered from most to least reliable. The theme throughout: models have mastered how things look, and still stumble on how things work.

1. Read every piece of text, especially the small text

Big, common words usually survive — a diner sign that says LUNCH COUNTER will say LUNCH COUNTER. But move to the second tier: jukebox placards, newspaper mastheads, shop signs in the background, a poster at the frame's edge. Generated images fill these with 'alphabet soup' — letterforms arranged into almost-words, or in other scripts, characters that nearly exist. One of our AI rounds has a neon sign reading “RAインン,” with a doubled character no human sign-painter would ever produce. If any text in an image dissolves under attention, you're done: that's a generation, not a photograph.

2. Ask what the camera was doing

A photograph is a record of physical exposure, and physics leaves paperwork. Stars twinkling above a daylit Earth: impossible — no sensor can expose for both. A 'phone snapshot' with creamy telephoto background blur: those optics don't live in that lens. Fire that glows without blowing out, neon that reflects in dry pavement, three light sources casting one shadow — each is the model compositing what looks right rather than simulating light. You don't need to be a photographer; you just need to ask 'could one camera, in one moment, have captured all of this?'

3. Follow the ropes

Functional objects have to work. Rope has to be tied to something; a strap has to bear weight from somewhere; a solar panel has to attach to a structure; a bicycle chain has to reach both gears. Models render the texture of these things beautifully and their mechanics approximately. In our beached-rowboat round, the rope wraps the hull in a confident spiral and terminates in — nothing. Pick any load-bearing detail in a suspect image and trace it end to end. Machines paint the idea of function; cameras record the fact of it.

4. Suspect the over-coherent

Real scenes contain junk that serves nothing: cable runs, tape scraps, a dumpster in the gorgeous alley, a stranger looking the wrong way. NASA's real cupola photos are cluttered with handrails and kapton tape; generated space scenes are clean. When every pedestrian conveniently faces away, when the mist and the pebbles and the sunrise all cooperate, when nothing in the frame is boring — that's not luck, that's optimization. This is the visual version of the tell we describe for prose in our guide to AI writing: nothing is allowed to be boring, and that's the problem.

5. Beware the déjà vu aesthetic

Some images feel generated because, statistically, they are the average of their genre: the rainy neon alley with a clear umbrella, the golden-hour portrait with bokeh, the cottage with impossible flowers. Models gravitate toward the most-rewarded image of each category, so heavy familiarity — 'I've seen this exact photo a thousand times' — is itself evidence. Real photographs are specific: a particular Tuesday, a particular dented car, a particular unglamorous sky.

6. Provenance beats pixels

Every tell above can, in principle, be fixed by a careful operator — regenerate the sign text, crop the broken rope. What can't be conjured is history. Real images have provenance: a reverse-image-search trail, a photographer, a negative, EXIF data, a wire-service record, other photos of the same moment from other angles. Dorothea Lange's 'Migrant Mother' is documented down to the pea field and the subject's name; a generated image appeared from nowhere, fully formed, today. Content-credential standards (C2PA) are slowly making this checkable in one click. Until then: before you trust a striking image, search for where it came from. 'Nowhere' is an answer.

And the honest caveat that runs through everything we publish: none of this is proof. Detection tools for images are as unreliable as their text cousins, skilled fakers can defeat any single tell, and real photos sometimes look wrong (that's why camera-authenticity disputes end up in court). What practice builds — and what the daily visual round is for — is calibrated suspicion: knowing which details to interrogate and how hard, before you share, cite, or believe.