Illustrations

AI Art Detector

Check whether an illustration, concept piece or digital artwork was made with generative AI. Works on output from Midjourney, Stable Diffusion, DALL·E, Flux and similar models.

Check artwork

Upload the image or paste it straight from your clipboard.

Ready
IDLE

Drop an image to analyse

or click to browse, or paste with CtrlV

JPEG · PNG · WebP · AVIF · GIF · TIFF · BMP — max 12 MB

Read this before trusting an art result

Artwork is the honest weak spot of every AI detector, including this one. That is not a limitation we can engineer away, and it is worth understanding why.

Style is not evidence. A human artist working in a flat, cel-shaded or airbrushed style produces smooth gradients, clean linework and minimal sensor noise — precisely the properties the detector associates with generation. The reverse holds too: a human can paint over AI output, and a generative model can imitate a specific artist convincingly. On stylised work, treat a mid-range score as genuinely uninformative.

Where detection does hold up on art is the technical layer rather than the visual one. Unmodified exports straight from a generator carry the same structural signature regardless of how the image looks: particular encoder settings, dimensions that are round multiples of the model's latent grid, characteristic upsampling ripple, and often a metadata trail. That is why the tool returns the evidence alongside the score — on artwork you should read the evidence, not the number.

If you are using this to settle a dispute

Accusations of AI use in art communities have real consequences, and a single score is not a sufficient basis for one. What a result can support is a decision to ask a question. The stronger evidence, in rough order of reliability, usually sits outside the image entirely:

  1. 01

    Process evidence

    Layered source files, timelapse, sketches and work-in-progress stages are strong proof of authorship. Their total absence across a long history is weaker evidence than people assume, but it is not nothing.

  2. 02

    File provenance

    Original exports with intact metadata, versus a stripped file of unknown origin. A C2PA manifest naming a generative tool is close to conclusive.

  3. 03

    Consistent structural tells

    Errors that repeat the same way across several pieces, especially where someone would have to make the identical mistake repeatedly.

Full method and failure cases are documented on accuracy. For photographs rather than illustrations, use the AI photo detector.