You can usually spot AI art in a fraction of a second, and it’s worth asking why — because the answer is a to-do list. The tells aren’t really about the technology. They’re about defaults: the choices the model makes when nobody makes a choice for it. The best AI-assisted work doesn’t look machine-made because someone overrode every one of those defaults on purpose.
The tells are all defaults
Left alone, a generator reaches for the center of everything it has seen. Perfect symmetry, because symmetry is average. Over-rendering, because more detail reads as “more effort” in the training data. A slightly plastic sheen, faces pointed at the camera, lighting that comes from nowhere in particular. None of these are mistakes exactly — they’re the visual sound of no decision being made.
Which means the fix is never “better prompting” in the abstract. It’s making the specific decision the model declined to make.
| The AI tell | Why it happens | The directorial override |
|---|---|---|
| Uncanny symmetry | Symmetry is the statistical average | Compose off-center; let the subject sit at an edge |
| Over-rendered, too much detail everywhere | Detail correlates with “effort” in training data | Decide what’s in focus and let the rest fall away |
| Plastic, glossy surfaces | Gloss photographs as “high quality” | Ask for a real material — clay, paper, worn metal |
| Faces to camera, posed | Portraits in the data face forward | Turn the subject away; catch them mid-action or at rest |
| Light from nowhere | No light source was ever specified | Name where the light is, and why it’s there |
A worked example: making it clay
For the Heroes, Off Duty series, the entire look leans on overriding the gloss tell. Every scene is specified as handmade claymation — visible fingerprints, tool marks, muted earth tones, one soft warm light with no obvious source. Those aren’t decorations; each is a direct veto of a default. “Visible fingerprints” kills the plastic sheen. “One warm light” kills the light-from-nowhere flatness. The image stops looking generated not because the model got better, but because every generic reflex got a specific instruction instead.
“Looks handmade” is the same skill as “looks intentional”
Here’s the part that generalizes beyond images. Whether you’re directing a generator, laying out a page, or editing a video, the thing that separates it from slop is the presence of decisions a viewer can feel. Off-center framing feels chosen. A single deliberate light feels chosen. Restraint — leaving most of the frame quiet — feels chosen. The machine’s defaults never feel chosen, because they weren’t. Your whole job is to leave fingerprints. (For the harder discipline of choosing among outputs, see Twenty Images, One Keeper.)
A de-slopping pass
- Break the symmetry. If the composition is centered and balanced by default, deliberately pull the subject off-axis.
- Specify a real material and a real light source. "Clay, lit warmly from the left" beats any amount of "high quality, detailed."
- Choose one thing to be in focus and let everything else soften. Uniform detail is the loudest tell there is.
- Turn faces and subjects away from the camera. Caught mid-action reads as real; posed-to-lens reads as generated.
- Do a final pass in a real editor — crop, grade, adjust. A human hand on the last 5% is often what erases the machine-made feel entirely.