Is this image AI generated?
Upload a picture and the detector tells you whether it carries the signals of an AI generator. It gives you a warning and the evidence behind it — never a certificate of truth.
or drag one here — JPG, PNG or WebP, up to 24 MB
Your image is sent to the detector and is not published anywhere. Nothing you upload appears on this site.
Check a whole OLX listing
Paste the link to an OLX Indonesia advert and every photo in it is checked at once — useful when a listing looks too polished for the price.
What the numbers actually are
Measured on 16 August 2026 through the deployed pipeline — the same path a visitor's upload takes, including the re-encode and the tile check. Not on a clean benchmark at a threshold picked afterwards.
Per source, because one average hides the spread
A single headline number publishes the mix of the test set as if it were a property of the model. These are the slices behind the 78.1% and the 4%.
| Source | Strong warning | Images |
|---|---|---|
| Stable Diffusion | 93.9% | 115 |
| DALL·E | 81.3% | 198 |
| Midjourney | 84.0% | 131 |
| AI logos and graphics | 81.5% | 65 |
| Mixed AI sources | 54.2% | 144 |
| Polished photography (real) | 3.0% | 197 |
| Screenshots (real) | 5.0% | 199 |
| Second-hand listings, Vinted (real) | 1.0% | 100 |
| Second-hand listings, Marktplaats (real) | 2.0% | 100 |
| Cinema trailer frames (real) | 3.5% | 200 |
Where it fails
Measured against OpenFake, a public benchmark of 20 generators we did not choose ourselves: 75.4% caught, with 1.8% false warnings on real photographs from ImageNet and DOCCI. Unfamiliar generators are harder, and some are invisible.
- Flux 2 is the blind spot. 22.5% caught on 40 images — better than the 7.5% of a day earlier, still by far the weakest. More than three quarters of that family still passes unnoticed.
- Newer versions undo earlier gains. Midjourney 6 scores 81.3% here; Midjourney 7 scores 72.5%. Every new generator release moves the ground.
- Mixed and edited material is the weakest ordinary case at 54.2% — a photograph with an AI element pasted in is far harder than a fully generated scene.
- Video is judged on 15 sampled frames. It says nothing about motion, lip-sync, or the moments between the frames it looked at.
How to read a warning
- A strong warning means the model found generator signals — not that the image is fake, and not that someone lied.
- No warning is not a clean bill of health. About one in five AI images passes unnoticed, and on unfamiliar generators it is closer to two in five.
- Absence of provenance data is not evidence. Most legitimate photographs carry no Content Credentials at all.
- Before any decision that affects a person, check the source, not just the score.
NU