What 11,000 Image Checks Look Like
Published 21 September 2026 · from our own server logs, 29 August to 21 September
Most writing about AI image detection argues about accuracy on test sets. We have something different: a record of what people actually upload when they want to know whether a picture is real. Over 24 days we served 11,353 checks on images real people brought to us. Here is what that looks like, including the parts that are inconvenient for us.
A third confidently real, a quarter confidently AI, and a real middle
We report a probability rather than a yes or no. Grouped into bands, the 11,353 checks fell out like this:
| Where the result landed | Share of checks |
|---|---|
| Strongly looks real | 32% |
| Leans real | 16% |
| Genuinely uncertain | 13% |
| Leans AI-generated | 16% |
| Strongly looks AI-generated | 23% |
The middle band is the honest part. Roughly one check in eight does not separate. A tool that never returned an uncertain answer would not be more accurate; it would just be hiding the same ambiguity behind a confident label.
Note what this is not: it is not the mix of images on the internet. People check things they already find suspicious, so a detector's inbox is skewed by definition. It is a picture of what doubt looks like in practice.
The bigger finding: provenance is already gone
Alongside the pixel checks we read the metadata of images uploaded to our metadata checker. That data is blunter:
- Only 29% still carried any EXIF at all. No camera, no date, no settings, nothing.
- 6% carried Content Credentials, the signed provenance record that cameras and AI tools have begun attaching.
- 68% of metadata readings were inconclusive — not "this is suspicious", simply nothing left to read.
This is what the verification problem actually looks like in 2026. The standards work — C2PA, signed capture, edit histories — but by the time an image reaches the person who needs to check it, it has usually passed through a screenshot, a messaging app, or a social platform, and every one of those strips the record. The provenance was not defeated by a forger. It was removed in transit, by the ordinary business of sharing a picture.
That has a practical consequence. If you are trying to establish whether an image is real, ask for the original file. It is the single highest-yield thing you can do, and it costs one message. Our guide on proving a photo is real walks through what to do with it once you have it.
What people upload, and how fast it comes back
The median upload is a 0.35 MB image; one in twenty is over 4.9 MB, which is usually someone sending an original straight off a camera. Most uploads are JPEG, then PNG.
The large files matter more than they look. They are the people doing the right thing — sending the original rather than a screenshot — and they are the uploads most likely to be slow or to hit a size limit. We changed our own metadata tool this week because of exactly that: it had been refusing 50-megapixel originals on a page that asks for the original.
What we take from this
Three things, and they are not flattering to the idea of a detector as a verdict machine.
An uncertain answer is a real answer. One in eight checks deserves one, and reporting it honestly is worth more than a confident guess.
A detector is one signal. It reads pixels. It cannot tell you who published an image first, when it was taken, or whether the caption is true. The three routes to an image's origin are complementary, not competing.
The provenance ecosystem has a distribution problem, not a technology problem. Six per cent is a very small number, and it will not grow by adding more standards. It grows when platforms stop stripping what is already there.
These figures come from our own server logs between 29 August and 21 September 2026. They record the result band, file size and format. They hold no image data, no filenames and no personal information; uploaded images are deleted after analysis. Our accuracy figures, measured separately on a held-out set, are published on the numbers page.
Check an image yourself
Free, no sign-up, and the image is deleted after the check.
Check an image →Frequently asked questions
How many of the images people check are AI-generated?
In 11,353 checks over 24 days, 23% came back in the highest confidence band for AI-generated and 32% in the highest band for real. The rest sat between, including 13% the model was genuinely unsure about. That is the mix of what people bring to a detector, not the mix of images on the internet — people mostly check things they already suspect.
Do photos still carry their metadata?
Mostly not. Of images uploaded to our metadata checker, only 29% carried any EXIF at all and 6% carried Content Credentials. Screenshots, re-saves, messaging apps and social platforms strip provenance before the image reaches whoever needs to verify it.
What does an uncertain result mean?
That the pixel evidence genuinely does not separate. About one check in eight lands there. It is not a failure of the tool so much as an honest reading. The cases where our detector is weakest are listed on our numbers page. Treat it as a prompt to look at the file's metadata and its source rather than as a verdict.
Is a detector enough on its own?
No, and this data is the reason. A detector reads pixels; it cannot tell you who published an image first or when it was taken. The images that are hardest to judge are exactly the ones that have been through the most hands, and those are the ones where provenance has been stripped.
Where do these numbers come from?
Our own server logs, covering every check between 29 August and 21 September 2026. They record the probability band, file size and format. They contain no image data, no filenames and no personal information, and uploaded images are deleted after analysis.