7 min read · Updated August 2026
How to spot an AI-generated photo without any special tools
Image generators have fixed most of the obvious mistakes they made in 2023, but they still leave a trail. The checks below are the same ones our own analysis runs through, written so you can do them with nothing but your eyes and the file itself. Work through them in order: the early ones catch the majority of generated images in a few seconds.
1. Read the small text in the frame
Generators build text the same way they build grass or fabric: as texture that looks correct rather than as letters that spell something. Zoom into signage, name badges, book spines, keyboard keys, price tags and license plates.
What gives it away is not one misspelled word but inconsistency — a shop sign that is crisp on the left and dissolving on the right, letters with mismatched stroke weights, or a repeated word where the second copy has different letterforms. Real photographs blur text evenly, based on distance and focus.
2. Follow the hands, ears and teeth
Fingers are better than they used to be, but knuckle counts, nail shapes and the way a hand grips an object are still weak spots. Check that each finger has a plausible joint structure and that anything being held actually makes contact with the skin instead of floating a millimetre away.
Ears are the quiet giveaway: in a two-person photo, the same lighting should produce the same ear anatomy on both sides of a face. Teeth often come out as a single fused ribbon with no gaps or shadow between them.
3. Test the physics of light
Pick the brightest light source in the image and predict where every shadow should fall. Generated scenes frequently have shadows that disagree with each other: one object lit from the left, its neighbour lit from above, and a third with no shadow at all.
Reflections are harder still. Look at eyes, sunglasses, windows, puddles, glossy tables and chrome. A real reflection contains a recognisable, geometrically sensible version of the scene. A generated one contains reflection-shaped noise.
4. Look at skin, hair and fabric at 200%
Real camera sensors add noise, and that noise is uneven: more visible in shadows, less in highlights. Many generated images are noise-free in a way no camera produces, giving skin an airbrushed plastic sheen and pores that repeat in a subtle pattern.
Hair is the opposite problem: individual strands merge into ribbons, or a stray strand ends in mid-air instead of tapering. Knitwear and lace often lose the weave and become a smear where the pattern should continue.
5. Interrogate the background
Generators put their effort where your attention goes. Push your attention to the edges instead: crowd faces, distant railings, brickwork, tiling, and anything with a repeating grid. Look for melted faces, railings that change spacing, bricks that bend, and objects that pass behind something and come out the other side misaligned.
Duplicated details — the same parked car twice, the same cloud shape mirrored — are strong evidence, because real scenes rarely repeat themselves that precisely.
6. Check the file, not just the picture
Photographs from a phone or camera carry metadata: camera make and model, lens, exposure, and usually a capture timestamp and sometimes GPS. A file with no camera fields at all is not proof of anything on its own — social platforms strip metadata — but it removes the strongest evidence in favour of authenticity.
Some generators write their own tags, and a growing number embed C2PA content credentials describing how the image was made. Both are worth looking for. So is a mismatch: a file claiming a DSLR model but saved at a resolution that camera never produced.
7. Trace where else the image has appeared
Run a reverse image search. If a supposedly breaking-news photo has been online for three years, or appears only on accounts created last week, that context outweighs any pixel-level analysis.
For claims about a place, compare the image against street-level map imagery of the same location. Generated scenes get the vibe of a city right and the specific corner wrong.
8. Separate 'generated' from 'edited'
Most misleading images are not fully synthetic. They are real photographs with one element removed, added, or generatively expanded. The tell is local rather than global: a region whose noise, sharpness or colour temperature does not match its surroundings, or an edge with a faint halo where something was patched in.
Because editing is limited to a region, the rest of the file can look completely normal — which is why a single confidence score is less useful than knowing which part of the picture is suspect.
9. Weigh the evidence instead of counting it
None of these checks is decisive alone. Heavy compression, aggressive phone beautification and screenshots all produce artefacts that resemble generation, and a careful generated image can pass several checks cleanly.
Treat the checks as evidence that accumulates. Three independent oddities that all point the same way — impossible reflection, inconsistent signage, absent metadata — are far stronger than one striking detail.
The short version
- Start with text, hands and reflections: they catch most generated images in seconds.
- Missing metadata is a weak signal on its own; contradictory metadata is a strong one.
- Local mismatches in noise or sharpness suggest editing rather than full generation.
- Confidence comes from several independent signals agreeing, never from one score.