The result
Measured on 2026-11-19: the detector was right on 184 of 203 pictures (91%). It caught 83 of 98 AI made pictures (85%) and cleared 101 of 105 human made pictures (96%). A picture counts as right when the detector's score is on the correct side of 50%.
| Pictures | Tested | Right | Score |
|---|---|---|---|
| AI made, photo-like (people, animals, food, places, objects) | 35 | 30 | 86% |
| AI made, art (paintings, drawings, comics, illustrations, renders) | 43 | 33 | 77% |
| AI made, anime and manga style | 20 | 20 | 100% |
| Human made photos (Commons featured pictures) | 59 | 56 | 95% |
| Human made paintings (museum photos and scans) | 23 | 22 | 96% |
| Human made illustrations and drawings | 18 | 18 | 100% |
| Human made anime style drawings | 5 | 5 | 100% |
| All test pictures | 203 | 184 | 91% |
The detector
The model is ai-image-detector-deploy by Sloan Haywood, a SwinV2 vision model trained to tell AI made pictures from human made ones, published under the Apache 2.0 licence. Its author trained it on a dataset he lists as 100,000 to 1 million pictures, tagged AI, art and pictures; the dataset is private and the author does not publish which generators it holds, so we measured the model ourselves. The site serves its own copy of the same weights, made smaller for faster loading: exported to ONNX and with the weights of every matrix stored in 4 bits instead of 32 (about 103 MB instead of 190 MB for the 8-bit copy we used at launch, and 750 MB for the original), as the Apache 2.0 licence allows, with the licence and a note of the changes beside the file (model notes). It runs with onnxruntime-web (MIT licence) in your browser. The 4-bit copy was measured on the same pictures before it went live: 91% right, against 91% for the 8-bit copy. Before choosing it we ran two other free models whose licences allow this use (Apache 2.0) on the same 203 pictures, in Python rather than the browser: AI-Anime-Image-Detector-ViT, trained only on anime pictures, was right on 69% and caught only 45% of the AI anime pictures; a SigLIP based detector (ai-vs-human-image-detector) caught more AI pictures (93%) but called 61% of the human illustrations and 48% of the human paintings AI made, which would wrongly flag artists. So the site uses one model for all pictures and states its results.
The test pictures
All test pictures come from Wikimedia Commons, where every file has a known uploader, a stated source and a licence that allows this use. They were read on 2026-11-19.
- AI made: pictures in Commons categories for named generators (Midjourney, DALL-E, ChatGPT, Gemini and Nano Banana, Imagen, Flux, Stable Diffusion, Adobe Firefly, Ideogram, Microsoft Bing Image Creator, Grok, Meta AI, Leonardo, NovelAI, Seedream, Recraft, Reve, and StyleGAN faces from thispersondoesnotexist.com) and categories for AI made photographs, food pictures, line drawings and manga. Each was looked at by hand: files that were not wholly AI made or not a single picture (charts, a screenshot of a chat, grids and collages, real photos run through AI upscalers or edited with AI, photos whose origin was unclear) were removed, and each AI picture was sorted by eye into photo-like, art or anime style.
- Human made: Commons featured pictures of people, landscapes, architecture, plants, birds, mammals and objects, with no date after 2021 in their file details; featured pictures of paintings (France, Japan, the Netherlands, the United States); featured illustrations and drawings; and anime style drawings made by people for Commons (Moe anthropomorphism and Wikipe-tan; the cosplay, train and mug photos in those categories were counted as photos). Files whose details mention AI were left out.
Pictures were checked as Commons serves them at up to 1024 pixels wide, the size most people see online. The site's three sample pictures are part of this set.
How it was run
The test runs the site itself in a headless Chrome browser, the same code and the same model file visitors get, picture by picture, and writes every score to a file kept with the site's code. This run scored 203 pictures in 901 seconds (4.4 s each) in headless Chrome on a 2 core cloud machine, one thread.
What this test does not show
- How it does on generators released after the model was trained, or on pictures made to fool detectors.
- How it does on screenshots and pictures squeezed by chat apps (weaker, in general).
- How it does on real photos edited with AI tools.
- The test is a few hundred pictures, not millions; each number could move by several points on a different set.
Pictures it got wrong
- Chinees decoratief vaasje op glas 20-12-2020. (actm.) 02.jpg (human made; the detector said 94% AI made)
- Earwax on swab.jpg (human made; the detector said 97% AI made)
- Nighthawks by Edward Hopper 1942.jpg (human made; the detector said 95% AI made)
- CWT30 DSC1036 23.jpg (human made; the detector said 98% AI made)
- Boanthropy.jpg (AI made; the detector said 6% AI made)
- Baobab Totem 2025.png (AI made; the detector said 1% AI made)
- ChaoticHarmony explosion in post-painter abstraction (Nano Banana 2).webp (AI made; the detector said 9% AI made)
- Algorithmically-generated landscape artwork of forest with Shinto shrine using negative prompt for green trees.png (AI made; the detector said 7% AI made)
- Guzpata.jpg (AI made; the detector said 9% AI made)
- Aula virtual.jpg (AI made; the detector said 1% AI made)
- Cold mountains of middle east.jpg (AI made; the detector said 2% AI made)
- Comet colonized by extraterrestrials 134 million years ago; long-term underground system still intact (AI art).jpg (AI made; the detector said 1% AI made)
- Hacking goat (DALL-E generated).jpg (AI made; the detector said 1% AI made)
- Doctor showing form.jpg (AI made; the detector said 1% AI made)
The full list of pictures with their links, licences and scores is kept with the site's code, and the test is run again when the model changes.