What image models still get wrong

The hardest AI images no longer announce themselves with waxy skin or a hand full of extra fingers. A polished synthetic portrait can look completely ordinary at first glance. Meanwhile, a real phone photo can look strange because of portrait mode, aggressive sharpening, low light or compression.

That changes the task. Do not ask whether an image feels AI. Ask whether the scene remains consistent when you inspect the relationships between its parts.

No visual checklist can authenticate an image. Treat these checks as reasons to investigate, then look for source context and provenance.

Start with the whole scene

Before zooming in, describe what the image claims to show. Where is the light coming from? Which objects should touch? What should be reflected? Which details repeat?

This first description gives you constraints to test. If you begin by hunting for a familiar artifact, you are more likely to see one even when it is caused by blur or compression.

Ask three broad questions:

Does the space make sense?
Do repeated structures stay consistent?
Do light, material and motion belong to the same scene?

Check whether geometry closes

Real scenes contain connected structures. A railing meets a staircase. Glasses reach both ears. A chair leg continues behind the seat. A cable emerges from one place and arrives somewhere plausible.

Generated images can render every local patch convincingly while losing the larger connection. Follow a line instead of staring at a texture. Trace the edge of a table, the seam of a jacket or the frame of a window. Look for a connection that changes direction without a physical reason, disappears too early or merges into a neighboring object.

Perspective deserves the same treatment. Parallel edges do not always remain parallel in a photograph, but they should converge in a way that belongs to one camera position. A room where the tabletop and floor imply different viewpoints deserves a closer look.

Compare repeated parts

Repetition creates a useful stress test because each instance should follow the same underlying rule. Good places to inspect include:

letters on a sign or book spine;
bricks, tiles, railings and fence posts;
buttons, eyelets and stitching on clothing;
wheels, spokes and mechanical controls;
hands or faces in the background of a crowd.

Do not reject an image because one repeated part is messy. Real scenes are messy too. Look for a pattern that begins coherently and then changes its own rules: spacing drifts, one element melts into another, or a row gains an impossible branch.

Compression can imitate this at small sizes. If possible, inspect the best available copy and compare the suspicious area with other areas at the same level of detail.

Follow the light through the frame

Light is a relationship. A bright window should influence faces, clothing, furniture and reflections in ways that broadly agree.

Start with the strongest source and predict what it should do. Which side of a face is brighter? Where should a cast shadow fall? Would a shiny surface reflect the source, the room or both? Then compare those predictions with the rest of the image.

Apparent contradictions can have ordinary explanations. A second lamp, reflected light from a wall or computational photography may flatten shadows. The useful signal is a cluster of conflicts, not one surprising highlight.

Reflections are especially informative because they must preserve the scene from a different angle. A reflected object can be partially hidden or distorted, but it should not become a merely similar object.

Test contact and physical cause

Look where things touch: a hand around a cup, feet on the floor, jewelry against skin, hair crossing a collar, or a tool held by a worker.

Contact changes shape. Fingers wrap around an object and obscure part of it. Fabric bends under pressure. A foot supports weight. Hair passes in front of some edges and behind others.

Generated scenes sometimes place convincing objects next to each other without resolving the contact between them. The cup and hand both look plausible, but the grip does not. The shoe and ground are detailed, but the weight has no clear support.

Water, smoke, hair and fabric deserve a wider view. They may look beautiful in isolation while responding to different wind or gravity than the rest of the scene.

Know which old clues are weak

Smooth skin is weak evidence. So are dramatic lighting, shallow depth of field, saturated colors and a cinematic composition. These are common in generated work, but they are also common in edited photography.

Perfect hands do not prove that an image is real. Newer systems often render hands well, and a creator can repair a small error after generation. Hidden hands provide no evidence either way.

Metadata is useful when it contains verifiable provenance, but ordinary metadata can be removed, changed or copied. Its absence is not proof of generation. Read more in why file metadata will not save you.

Move from detection to verification

When the claim matters, leave the pixels and investigate the source.

Find the earliest available upload. Check whether a reputable person or organization takes responsibility for it. Search for other views of the same event. Look for contradictions in time, place and context. When Content Credentials or another provenance record is present, inspect what it actually attests to rather than treating the badge as a universal guarantee.

The right conclusion is sometimes “I cannot tell from this image.” That is a useful result. It prevents a weak visual hunch from becoming a confident accusation.

A 30-second routine

Describe the claimed scene and strongest light source.
Trace one connected structure from end to end.
Compare a repeated pattern in at least three places.
Inspect one contact point and one reflection.
Decide what source evidence you would need before sharing the claim.

Use the image quiz to practise the routine on unfamiliar examples. After each choice, compare your reasoning with the review notes rather than only checking whether the answer was right.