Illustration & digital art
AI art detector
For artists, buyers, art directors and contest judges: drop an illustration, painting or render to see whether it carries AI generation metadata and how a neural classifier scores it.
Free · no sign-up · your file never leaves your device
Drop the artwork, click to choose, or paste
JPEG · PNG · WebP · AVIF · HEIC (Safari) — checked on your device
Why checking art is different from checking photos
A photo detector asks a narrow question: did a camera capture this? Art has no such baseline. A digital painting was never captured by a sensor, so “no camera metadata” means nothing, and its brushstrokes, gradients and textures can resemble generator output in ways a photograph never does. That makes art the hardest category for AI detection — and the one where a false accusation hurts real people most.
GPTTrace handles this by leaning on the evidence that does not depend on style. Generator metadata is the same whether the image is a photo or an anime illustration: a Stable Diffusion parameters chunk, a ComfyUI workflow, NovelAI’s hidden alpha-channel record, an Adobe Firefly credential or a Midjourney prompt all identify AI output with certainty. The pixel classifier is reported as a probability, and the verdict stays “Inconclusive” unless the evidence is strong.
What tends to give AI artwork away
- Inconsistent small details — jewellery, buttons, belt buckles and patterns that change from one side to the other, or decorative text that is nearly-but-not-quite letters.
- Logic errors — straps that go nowhere, a sword hilt that merges into a hand, stairs that loop, architecture that cannot stand.
- Uniform finish — every area rendered to the same level of polish, where a human would leave background areas looser.
- No process — the account posts finished pieces in many different styles at a pace no single artist could manage, never shows sketches, and can’t supply a layered file.
Good artists sometimes show some of these, and good AI users fix most of them. Use them to decide what to ask, not what to conclude.
For contests, publishers and marketplaces
If your rules require human-made work, ask entrants for process evidence up front rather than relying on detection afterwards. A detector score can justify a closer look, never a disqualification by itself. When a piece does carry generator metadata, GPTTrace shows exactly which field and what it says, so you have something concrete to discuss with the entrant. Because nothing is uploaded, you can check unpublished submissions without leaking them.
Reading a result on artwork
A “proof” result on artwork means generator metadata was found — the file itself says which tool made it, and often the prompt. A high classifier score without metadata means the texture statistics resemble generated images; on painterly or heavily stylised work, treat that as a reason to ask for process files rather than a conclusion. An “Inconclusive” result is common for illustrations and should be read literally. When you do raise a question with an artist, show them the specific evidence from the report — a named generator, a prompt chunk, a C2PA history — rather than a percentage.
How accurate is it? Our measured numbers
We test GPTTrace on labelled image samples and publish the results, including where it does badly. It is tuned to keep false accusations rare, so it misses some AI content rather than flag real work.
image check: AUC 0.938 (cross-validated)
679 labelled samples (399 AI, 280 human), run 2026-10-08. At the “Likely AI” line it caught 64% of AI samples and wrongly flagged 5% of human ones.
| Source | Truth | Samples | Result at “Likely AI” |
|---|---|---|---|
| gemini-nano-banana | AI | 40 | 40% caught |
| midjourney-v6 | AI | 40 | 68% caught |
| midjourney-v5 | AI | 40 | 73% caught |
| flux-dev | AI | 40 | 13% caught |
| flux-schnell | AI | 40 | 48% caught |
| sdxl | AI | 40 | 100% caught |
| gpt-image | AI | 40 | 30% caught |
| kling | AI | 39 | 97% caught |
| leonardo-stablecog | AI | 40 | 98% caught |
| bitmind-imagine-mix | AI | 40 | 80% caught |
| fullsize-photos | Human | 40 | 10% wrongly flagged |
| open-images-photos | Human | 40 | 5% wrongly flagged |
| lfw-faces | Human | 40 | 0% wrongly flagged |
| caltech-objects | Human | 40 | 3% wrongly flagged |
| coco-photos | Human | 40 | 0% wrongly flagged |
| ffhq-faces | Human | 40 | 0% wrongly flagged |
| celeba-faces | Human | 40 | 18% wrongly flagged |
Data sources and method: methodology & accuracy.