How to tell if art is AI generated

Read the file first, briefly

The checker above reads the same fields it reads for a photograph, and two of them come up often on artwork. Adobe attaches Content Credentials to anything made with Firefly and to images touched by Generative Fill, Generative Expand or the neural filters in Photoshop, so an image that passed through those tools can be carrying a signed record of it. Stable Diffusion front ends write the prompt, seed, sampler and model name straight into PNG text chunks, and people export and post those files without ever opening them.

The IPTC digital source type field is the third one worth knowing. A value of trainedAlgorithmicMedia is the software declaring that a generative model made the image, and it survives in files that platforms have not stripped.

Process evidence, which is where drawn work differs

This is the part that has no equivalent in photography or in prose, and it is the reason an accused illustrator is usually in a better position than an accused essay writer.

• The layered document. A working PSD, PROCREATE, KRA or CSP file with named layers, masks, a sketch layer underneath and a colour flats layer is the record of a decision sequence. A flattened export is not.
• Time-lapse replay. Procreate records every stroke of every painting automatically and exports it as video from Actions, then Video, then Export time-lapse video. Most artists have never touched the setting, which means the recording exists whether or not they expected to need it.
• The discards. Thumbnails, a composition that was abandoned, a version with the arm in the wrong place. Generated work has no failed attempts because there was nothing to fail at.
• Physical originals. Pencil under paint, an indented line where a stylus pressed through paper, a scan at a higher resolution than anything posted. These are difficult to produce after an accusation and trivial to produce before one.

Looking alone was measured, and it is not good

A Microsoft Research team published the largest experiment of its kind in 2025. Around 287,000 image judgements from more than 12,500 people, across photographs, portraits and art, produced an overall accuracy of 62%. People did best on human faces. They did worst on natural and urban landscapes, which is to say on the subjects where they had the least idea what a mistake would look like.

Sixty-two percent is better than a coin and nowhere near enough to accuse somebody. Illustration is harder still than the categories in that study, because a stylised image gives you no ground truth: a hand drawn in a personal style is supposed to be wrong about anatomy, and a flat background is a choice rather than a failure.

Signs specific to drawn and painted work

• Strokes with no tool behind them. Paint that thins in ways no brush thins, a hard edge next to a soft one with no reason for either, texture that sits on top of the form rather than following it.
• Line weight that carries no intent. Human line work thickens where the artist leaned in and thins where the hand lifted. Generated line work varies at random.
• Decoration that dissolves under magnification. Lace, chainmail, keyboard keys, book spines on a shelf: patterns that read at thumbnail size and turn to soup at full resolution.
• A smear where a signature should be. Models learned that finished artwork carries a mark in the corner and reproduce the shape of one without the letters.
• Perspective that is locally right and globally wrong. Each window is drawn correctly; the building they sit in does not close.
• Style that is consistent within one image and unrepeatable across two. Ask for another piece of the same character from a different angle.

Questions that come up in the comments

What the platforms require

Both of the large portfolio sites moved to labelling rather than banning. DeviantArt requires submitters to say whether AI tools were involved, applies the label automatically when an image's metadata says so, and runs its own detection over work that arrives undeclared. ArtStation requires a tag on generative work, and on marketplace products that tag is a condition of sale.

Labels catch the honest and the careless. Somebody who strips metadata and ticks the wrong box gets through both systems, which is the same limit every disclosure regime has. What they do give you is a place to report, and a policy to point at when you do.

If it is your work being questioned

• Stop exporting and stop flattening. Keep the working file exactly as it is.
• Export the time-lapse today if your app records one, and put it somewhere with a date on it.
• Gather the discards: thumbnails, the version you hated, the reference photographs you shot yourself.
• Offer a live demonstration rather than an argument. Draw the character again, on stream or on a call, and the question usually ends there.