A trustworthy answer to one hard question.

Why we built it

AI image generators went from novelty to everywhere in the space of a couple of years. That is genuinely useful, and it quietly broke an assumption most of us still carry: that a photograph is evidence of something that happened. A convincing fake now takes seconds to make and costs nothing.

The tools meant to help were mostly the opposite of reassuring. Many demand an account, upload your private images to a server you know nothing about, then hand back a single percentage with no explanation. You are asked to trust a black box about whether you can trust an image. We wanted the opposite: a result you can actually reason about, produced without ever taking custody of your files.

How the detector works

At the core is a vision transformer, __x Pixel Forensics · FT1, trained to recognise the statistical fingerprints that generation and editing leave in pixels. The image is analysed as a whole, then again as a grid of overlapping regions, so a small AI edit hidden inside an otherwise real photo still shows up. Those region scores drive the heatmap you see in the report.

The output is a single calibrated score from 0 to 100, mapped to five fixed bands with a decision line at 65. Because the mapping is fixed and published, a 72 today means exactly what a 72 meant last week, and you can see the reasoning, the region evidence and the file signals behind every verdict.

Private by design

The model, the runtime and all of the analysis are downloaded to your browser and run there. Images you choose from your device are never uploaded, stored, queued or logged. The one exception is transparent: if you paste a link to an image that a site refuses to share directly, we fetch it through a public proxy so the check can run. That involves the URL, never a file from your computer, and it is spelled out in our privacy policy.

How we check that it still works

Every release is run against a fixed evaluation set before it ships: authentic photographs from several camera generations, output from the diffusion and transformer models available at the time, and, the part that actually matters, the awkward middle. Screenshots of real photos. Phone shots processed by night mode. Studio product images with the noise smoothed out. Real pictures with one object removed.

Two numbers are watched rather than one. Overall accuracy is easy to move by shifting the decision line, so the false positive rate on genuine photographs is tracked separately and treated as the harder constraint. A build that gains a point of accuracy by flagging more real photographs is not an improvement, because the cost of wrongly calling a genuine picture fake lands on a person.

Honest about the limits

No detector is right on every image, and anyone claiming otherwise is selling something. Heavy compression, screenshots, upscaling and aggressive filters can all push a real photo toward the middle of the scale, and a determined forger can work to defeat any single model. That is why the score is presented as a likelihood alongside its evidence: a strong input to a human decision, never a verdict to be quoted on its own.

The specific failure modes are written down rather than glossed over, and a result should never be used on its own in a legal, employment or disciplinary decision. Read accuracy and limitations before you act on a score that affects somebody.

Three things we will not do

We will not start uploading your images. Local analysis is the architecture, not a feature flag. If a future capability could only work server-side, it does not ship.

We will not put detection behind a paywall. The score, the region map and the exports stay free. Everything the tool does is on the pricing page, and it costs nothing.

We will not quote an accuracy figure we cannot support. The benchmark number on this site is from a held-out research set under controlled conditions, and it is labelled that way everywhere it appears, because real-world performance is lower.