Strong warning on AI images
78.1%95% interval 74.8–81.1%, n=653.Current evidence
Bounded claims.
Visible limits.
Current lane: aidetect-v16-image-384-v1+owenslab+siglip-probe-routed-v11+dualview-orig-v1, measured 16 August 2026. Every number below comes from the routed lane the product actually runs, not from a model in isolation.
Still images.
653 AI images and 795 real ones, scored through the deployed lane.
False warning on real images
4.0%95% interval 2.9–5.6%, n=795.The spread matters more than the average.
A single figure hides which generators the lane knows. These are the same measurement, split by source:
| slice | n | strong warning |
|---|---|---|
| DALL·E | 198 | 81.3% |
| Stable Diffusion | 115 | 93.9% |
| AI logos and graphics | 65 | 81.5% |
| Midjourney | 131 | 84.0% |
| mixed AI sources | 144 | 54.2% |
| polished quality photography | 199 | 1.0% |
| polished featured photography | 197 | 3.0% |
| screenshots | 199 | 5.0% |
| human-made graphics | 200 | 7.0% |
| second-hand listings, Vinted | 100 | 1.0% |
| second-hand listings, Marktplaats | 100 | 2.0% |
The last four rows are real photographs and graphics, where any warning is a false one.
Video, 15 frames.
A video draws a warning when at least 3 of 15 sampled frames score strongly. On 23 videos that rule caught 82.6% (95% interval 62.9–93.0%) at 1.4% false warnings.
That interval is wide because 23 videos is a small set, and it is published wide rather than rounded away. Sampled video says nothing about motion, lip-sync, or the moments between sampled frames.
On 200 frames from 70 official trailers across 6 distributors unseen in training, 3.5% drew a false warning.
The weakest spot, stated plainly.
On 150 August 2026 Midjourney images the lane caught 81.3% (95% interval 74.4–86.7%) — against 67.1% on older Midjourney in the table above.
Same generator, newer output. The newest generation is where this lane is weakest, and that number is kept out of the headline instead of averaged into it.
Historical and superseded evidence
Earlier versions of this page published 84.0% on a Midjourney slice and 19.3% false warnings on trailer frames. Both described a model measured on its own, not the routed lane that serves requests; through the lane the same images give the figures above. The pre-registration documents that produced them remain in the repository as model-selection records, not as product claims.
What we do not claim.
- No single overall accuracy percentage; the slices differ too much to average honestly.
- No proof of origin, authenticity, or authorship.
- No claim that a quiet result means human-made.
- No temporal deepfake analysis from sampled video.
- Text and audio are not supported public product lanes.