Why the old checklist stopped working
Every widely shared list of tells has the same problem: it describes the weaknesses of the models that existed when it was written. Those weaknesses are public, measurable and embarrassing, which makes them precisely the things the next release is tuned to fix.
Hands are the clearest example. Counting fingers was genuinely useful for about eighteen months, then it became a way to be confidently wrong in both directions: generated hands are usually correct now, and real hands in motion blur into shapes that look wrong to somebody who has been told to count.
The same has happened to eye reflections, ear asymmetry and dental geometry, in roughly that order. Anything that can be described in a viral post gets benchmarked, and anything benchmarked gets fixed.
-
2019
Backgrounds dissolved Early face generators produced a sharp face on a smear. Fixed once models learned whole scenes.
-
2022
Ears and earrings mismatched Asymmetric jewellery was a strong signal until training data covered it.
-
2023
Hands and fingers The most quoted tell of all. Largely solved within two release cycles.
-
2024
Text in the frame Signage and labels became legible, removing an easy giveaway.
-
2026
What is left is statistical The remaining differences are in texture and noise, below what an eye resolves.
Look away from the face
Generation effort concentrates where attention concentrates. A model trained on portraits gets very good at the region people look at, and stays weaker on everything the training signal treated as background.
That is where the durable checks live. Ask what is written on the lanyard, whether the pattern on the shirt continues correctly behind the arm, whether the chain of a necklace passes behind the neck and comes out in the right place, and whether the room behind the shoulder is a place that could exist.
Repeating detail is the other reliable family. Wallpaper, brickwork, foliage and crowd faces are expensive to generate coherently, so they tend to repeat, drift or lose their vanishing point in ways an eye can catch even when the portrait in front of them is flawless.
- Text away from the centre. Badges, signage, book spines, packaging. Central text is now usually correct; peripheral text is still where it falls apart.
- Continuity through an occlusion. A strap, a chain or a stripe that goes behind something and comes back wrong.
- Crowds and repeated objects. The second and third row of a crowd, or the fourth chair in a row, get less attention than the first.
- Skin at high magnification. Real skin has pores, stray hairs and asymmetric blemishes. Generated skin is often smooth in a way no lighting explains.
- Where hair meets everything. The boundary between hair and background remains one of the hardest regions to render coherently.
What a detector sees that you cannot
The reason to run a check rather than rely on inspection is that the remaining differences are below the resolution of human vision. A camera sensor produces a particular kind of noise; a generation process produces a different one, and neither is visible at normal viewing size.
This is also why a detector and your eyes can disagree. A picture with an obvious visual error may score low because its texture is photographic, and a picture that looks perfect may score high because its texture is not. Both readings are informative, and neither overrules the other.
| Visual inspection | Pixel analysis | |
|---|---|---|
| Works on | Composition and logic errors | Texture and noise structure |
| Survives compression | Yes | Degrades |
| Survives a screenshot | Yes | Degrades badly |
| Catches a small edit | Only if you notice it | Yes, via the region map |
| Gets worse over time | Rapidly | Gradually |
The last row is the useful one. Visual tells decay with every model release; a detector trained across many generators decays more slowly, because it learned what synthetic texture has in common rather than what one product got wrong in one year.
The case that matters most: a real face, altered
Most consequential fakes are not wholly generated portraits. They are photographs of real people with one thing changed: a different face on the same body, an object added to the hand, a badge altered, a second person removed from the frame.
A whole-frame score handles this badly by design, because ninety per cent of the image really is a photograph and the average reflects that. The region map is what surfaces it, and reading the map rather than the number is the single habit that most improves accuracy on faces.
Eleven tiles read as photographic. One does not, and it is the one covering a face.
A workable habit
Give the portrait five seconds of peripheral inspection: text, continuity, repeated pattern, hair boundary. Then run a check and read the map. The two together catch far more than either alone, and the whole thing takes under a minute.
Where the stakes are real, add the step that no amount of pixel analysis replaces: find the same face somewhere with a history. A profile with three years of posts, tagged photographs from other people and a consistent employer is evidence of a kind no generator produces.
The tells that have not moved
Two things have resisted four years of improvement, and both are about coherence across distance rather than local detail. The first is lighting: whether every surface in the frame is lit by the same sources, from the same directions, with shadows of the same softness.
A generated portrait usually gets the face right and the surroundings approximately right, and the mismatch shows in the shadows. A hard shadow under the chin next to a soft shadow on the wall behind describes two different rooms.
The second is depth. Reflective surfaces, glass, water and mirrors have to agree with the geometry of everything else, and they often do not. A window behind a subject showing a scene that could not be there is the kind of error that survives model upgrades, because it needs a model of the world rather than a model of faces.
Neither is a test you can apply mechanically, and both take longer than counting fingers did. That is the trade: the checks that still work are the ones that require thinking about the picture as a place rather than scanning it for a defect.