How to Spot AI Art

The signature that spells nothing

Still the single strongest tell in generated art. Diffusion models learned from millions of paintings that artwork has a signature in the corner, so they draw one — a confident squiggle that looks like handwriting and resolves into no letters at all.

Zoom into the corners. A real signature is a name, however stylised. A generated one is the *idea* of a signature: right rhythm, right placement, no content.

Brushwork with no hand behind it

A human builds a stroke with direction and pressure. Strokes follow the form: along a cheekbone, around a sleeve, with the grain of a surface.

Generated texture is statistically uniform. It looks like brushwork at a glance but the strokes do not organise around anything. Our texture analyzer measures exactly this — stroke coherence across the canvas — and it is one of the methods that still works on stylised images.

Detail that never decides what matters

Artists spend detail where attention should go and simplify everywhere else. It is the most basic compositional decision there is.

Generated pieces distribute ornament evenly: the same density of detail on the face, the sleeve, the background curtain and the floor tiles. Nothing recedes. This is visible to the eye once you know to look, and measurable in a frequency analysis.

Backgrounds and the things inside them

Architecture that does not resolve. Beams meeting at impossible angles, stairs going nowhere, windows that do not line up between floors.

Patterns that drift. A tiled floor, a woven fabric, a brick wall — real patterns repeat rigidly. Generated ones shift scale or lose alignment mid-run.

Anatomy under load. Not hands any more — those are largely solved. Look at ears, at teeth, at how jewellery meets skin, at where a strap disappears behind a shoulder and comes back wrong.

What stopped working, and what never worked

Counting fingers is obsolete. Current models handle hands well enough that a correct hand tells you nothing.

Garbled text is obsolete too. Ideogram was built for typography and others have caught up — see our Ideogram detector for what that changed.

“It looks too clean” never worked. Plenty of human digital art is immaculate, and generators deliberately add grain and imperfection now.

The problem nobody wants to discuss

Every list of tells is also a list of ways to falsely accuse a real artist. Clean digital illustration with flat colour shares statistical properties with generated work. Artists have been driven off platforms by accusations built on exactly the reasoning above.

So the useful question is usually not “is this AI” but “can human authorship be demonstrated”. Layered source files, timelapse recordings, sketch stages, revision history — these prove authorship in a way no detector output ever will, ours included.

If you are checking someone else's work, run the analysis with our AI art detector and then ask for the working files. If the answer matters, the second request settles it faster than the first.

Key points

• The corner signature that spells nothing is still the strongest single tell.
• Look for brushwork without direction and detail that never recedes.
• Fingers and garbled text are dead giveaways — of outdated advice.
• Stylised art is genuinely harder to analyse than photographs. Expect less certainty.
• Layered files and timelapse prove human authorship better than any detector.
• A detector score is never grounds to accuse an artist publicly.

Frequently asked questions

Check an image yourself

Reading about it only goes so far. Our AI image detector runs twelve independent methods and shows you every score, not just a verdict.

More from the blog