How to Tell AI-Generated Images From Hand-Made Work

Artists are increasingly asked to judge whether an image was generated, whether as a contest judge, a client, a moderator, or simply a curious viewer. The craft knowledge taught across this library turns out to be the best toolkit for the job, because generated images tend to fail in the places where a trained artist makes deliberate decisions. This guide covers what to look for, and just as importantly, how to avoid accusing a real artist on thin evidence.

Look at construction, not surface

A generated image is usually strongest at the surface: texture, lighting mood, polish. It is weakest at underlying structure. Check the things an artist has to construct. Do both eyes sit on the same wrapping line, as in head proportions from any angle? Do the hands have a consistent number of fingers and joints that bend the right way? Do earrings, buttons, and straps match on both sides? Structural contradictions under a polished surface are a classic generation pattern.

Read the edges and the light

A painter controls edges on purpose: hard at the focal point, soft and lost elsewhere, as explained in hard and soft edges. Generated images often have edges that are uniformly soft, or that sharpen in places with no reason to be the focus. Light is similar. Check whether cast shadows agree with the highlights, whether the single highlight in an eye is on the same side in both eyes, and whether reflected light comes from something actually in the scene.

Check the places models stop paying attention

• Backgrounds: architecture, fences, and furniture that melt, bend, or change logic halfway across the frame.
• Repeated elements: windows, teeth, fingers, chain links, or leaves that multiply or merge.
• Text and symbols: lettering that almost reads but does not, or logos that are close to real but wrong.
• Texture logic: brush texture that appears with no stroke direction, or that is identical in the sky and the skin.
• Hair and fabric: strands that pass through each other or folds that do not follow any pull or gravity, unlike hair drawn as flowing masses.

Ask for process, politely

The most reliable way to settle a doubt is not forensic squinting. It is asking for process. A working artist can usually show roughs, a layered file, or a timelapse within minutes, which is exactly the record described in how to document your art process. If the context is a contest or a commission, make process evidence part of the rules in advance so nobody is singled out.

What about detection tools?

Automated detectors exist for images, and some are useful as a first pass. But they produce false positives on real art, especially on stylized, smooth, or heavily edited pieces, and false negatives on edited generations. Treat any score as a prompt to look closer and to ask for process, never as a verdict. Provenance data, where present, is a better signal than a score; see Content Credentials explained.

A fair review checklist

1. Check construction: eyes, hands, symmetry of paired objects.
2. Check edges and light for deliberate focus and consistent sources.
3. Scan backgrounds, repeats, text, and texture direction.
4. Look for provenance data on the original file.
5. Ask for process evidence before reaching any conclusion.
6. If still unsure, say so. Uncertain is an honest answer.

Used this way, the eye you train drawing eyes, heads, hair, and edges becomes a fair and careful judge, and it protects real artists from being wrongly accused as much as it catches generated work.

Sources and further reading

Primary references for the facts and definitions in this piece, all from independent publishers:

• Ha et al., Organic or Diffused (2024): how automated detectors and artists compare at telling human art from generated images.
• NIST AI 100-4, Reducing Risks Posed by Synthetic Content (2024): detection, provenance and watermarking as complementary approaches.