Test results
How often it's right, and where it isn't.
We test Find the AI on photos and videos it never trained on: real ones from phones and public datasets, and AI ones from today's generators, plus copies squeezed the way social apps do it. These are the model's own numbers, without the proof we sometimes find in a file (Content Credentials catch every original Sora file we tested, for example).
We set the cut-offs so real things rarely read as AI. The real-file numbers are measured as if we had never seen that kind of file: each source's rates come from cut-offs set without it. The price of caution is that some AI slips through as Inconclusive or No AI detected, most of all from the newest video models. Updated October 25, 2026.
Photos
Real photos marked AI
Lower is better.
| Source | Items | Very likely AI | Likely AI or more |
|---|---|---|---|
| iPhone photos | 80 | 0% | 0% |
| Our own phone photos | 42 | 0% | 0% |
| Everyday photos (COCO) | 400 | 0% | 0.5% |
| Android phone photos | 280 | 0.7% | 1.4% |
| All | 802 | 0.3% | 0.7% |
AI photos caught
Higher is better.
| Made with | Items | Very likely AI | Likely AI or more |
|---|---|---|---|
| DALL-E 3 | 80 | 90% | 100% |
| Hunyuan Image 2.1 | 32 | 84% | 100% |
| Janus 7B | 80 | 99% | 100% |
| Lumina | 46 | 78% | 100% |
| Flux 2 Dev and Flex | 160 | 82% | 99% |
| ChatGPT (GPT-4o) | 80 | 50% | 96% |
| HiDream I1 | 78 | 69% | 96% |
| Flux 1 | 160 | 84% | 95% |
| Grok (Aurora) | 80 | 79% | 95% |
| Recraft V2 | 66 | 74% | 94% |
| Google Imagen 4 Ultra | 72 | 51% | 93% |
| Google Imagen 3 | 80 | 75% | 93% |
| Stable Diffusion 3 | 80 | 61% | 91% |
| Wan 2.2 image | 80 | 61% | 90% |
| Midjourney 5.2 | 78 | 69% | 90% |
| Nano Banana (Gemini) | 80 | 65% | 89% |
| Flux 2 Pro | 80 | 65% | 88% |
| Runway Frames | 80 | 44% | 86% |
| Ideogram | 64 | 50% | 86% |
| Nano Banana Pro (Gemini) | 58 | 40% | 84% |
| Halfmoon | 58 | 29% | 78% |
| Flux 1.1 Pro | 80 | 49% | 69% |
| Recraft V3 | 32 | 41% | 69% |
| Seedream 3 | 32 | 13% | 56% |
| All | 1816 | 66% | 91% |
Videos
Real videos marked AI
Lower is better.
| Source | Items | Very likely AI | Likely AI or more |
|---|---|---|---|
| Activity videos (HAR) | 40 | 0% | 0% |
| Our own phone videos | 198 | 0% | 0% |
| Web videos (AIGVDBench real set) | 150 | 0% | 0.7% |
| Everyday videos (VidOR) | 100 | 7.0% | 11% |
| Workout videos | 40 | 0% | 15% |
| All | 528 | 1.3% | 3.4% |
AI videos caught
Higher is better.
| Made with | Items | Very likely AI | Likely AI or more |
|---|---|---|---|
| Jimeng (Dreamina) | 19 | 89% | 100% |
| Runway Gen-3 | 19 | 58% | 100% |
| Pika | 67 | 73% | 91% |
| OpenAI Sora (first version) | 75 | 75% | 91% |
| Runway | 32 | 84% | 91% |
| Wan | 105 | 74% | 87% |
| Luma Ray2 | 44 | 59% | 86% |
| Google Veo 3.1 | 53 | 68% | 81% |
| Luma | 31 | 71% | 81% |
| Moonvalley Marey | 36 | 44% | 78% |
| Seedance | 38 | 61% | 76% |
| Vidu | 23 | 35% | 65% |
| Our own renders (Wan, LTX and others) | 77 | 49% | 62% |
| Kling | 99 | 44% | 58% |
| Google Veo 3 | 142 | 34% | 56% |
| HunyuanVideo | 54 | 24% | 44% |
| OpenAI Sora 2 | 92 | 6.5% | 25% |
| All | 1006 | 51% | 69% |
How we test
AI photos come from public research sets of each generator's output; real photos from phones (iPhone, Pixel, Galaxy, Xiaomi), the COCO dataset and our own cameras. Videos come from AIGVDBench, Rapidata's public sets and other public collections, everyday-video datasets and our own phone footage. The phone models and files in the test set are never used for training. Video scores are cross-validated by leaving whole sources out.
Numbers on a test set are not a promise about every file. A result is evidence to weigh with where the file came from. How it works