One hard question, answered honestly.

The question that got harder

For most of photography's history, a picture was weak evidence that something happened — imperfect, but worth something. That held right up until generating a convincing frame became free, instant and available to anyone with a browser. The assumption did not survive, and most people have not replaced it with anything.

The tools offered as a replacement were, for the most part, worse than the problem. Hand over your private picture. Create an account. Receive a percentage with no explanation attached and no way to interrogate it. You were being asked to extend trust to a black box in order to decide whether to extend trust to a picture, which is not progress.

What we built instead

A vision transformer, __x Pixel Forensics · FT1, trained to separate the statistical residue that generation and editing leave in pixels from the residue a lens and a sensor leave. The frame is read whole for the headline number, then again as a grid of overlapping tiles — as many as nine — so a small generated patch inside an otherwise ordinary photograph has somewhere to surface. Those tile scores are what the map on the report is drawn from.

The output is one calibrated score from 0 to 100, mapped onto five fixed bands with the decision line at 65. Because that mapping is fixed and published rather than tuned quietly between releases, a 72 today carries the same meaning a 72 carried last month, and every verdict arrives with the reasoning, the tile evidence and the file signals laid out beside it.

Why nothing is uploaded

The model, the runtime and the whole analysis are downloaded to your browser and executed there. A file you pick off your own device is never uploaded, stored, queued or logged, and it cannot end up in a training set because no copy of it exists anywhere but your machine. The one exception is stated rather than buried: paste a link to a picture a site will not serve to your browser directly, and we fetch it through a public proxy so the check can proceed. That involves a URL and never a file from your computer, and the privacy policy sets it out in full.

This is deliberately an architectural commitment rather than a policy one. A promise not to look at your files depends on our conduct; an architecture with nowhere to send them does not.

How we know it still works

Every release runs against a fixed evaluation set before it ships: authentic photographs across several camera generations, output from the diffusion and transformer models available at the time, and — the part that actually decides whether a detector is useful — the awkward middle. Screenshots of real photographs. Phone pictures put through night mode. Studio product shots with the noise polished out. Genuine pictures with a single object removed.

Two figures are tracked rather than one. Overall accuracy is easy to flatter by moving the decision line, so the false positive rate on genuine photographs is watched separately and treated as the tighter constraint. A build that buys a point of accuracy by flagging more real photographs has not improved anything, because wrongly calling genuine work fake costs a real person something that a missed detection usually does not.

What comes next

Pictures were the first problem, not the only one. Detectors for video, for cloned voice and for generated music are in progress, each built to the same three rules: run on the visitor's own device, show the evidence behind the number, and state plainly what the thing cannot do. None of them will ship on a date announced in advance — each ships when its own accuracy figures are worth publishing next to the ones on this page.

What we will not claim

No detector reads intent, reconstructs a file's history, or identifies which product generated a picture, and anything advertising those abilities is selling further ahead of the evidence than we are willing to. What this measures is how closely a frame resembles the things the model learned to recognise. Used as one input among several, that is genuinely valuable. Used as a verdict on its own, it will eventually hurt somebody who did nothing wrong — which is why the limits are on the homepage rather than in the footnotes.