Whoever puts their name on a story or a photograph doesn't need a percentage: they need to be able to explain why.
An image sent in by a source, a portrait in a press release, a text that reads a little too well: verification always happens under time pressure and heavy responsibility. This page walks through the workflow we use ourselves — the eye first, then the tools — and what an AI detector like Uthentic adds beyond a bare score.
Always ask for the original file, not a screenshot or a forwarded copy from a chat app: every step compresses the image and strips out the data you need to verify it. Note who sent it to you and when.
An image saved from social media has already lost its metadata: missing EXIF is not proof of a fake.
Light and shadows consistent with a single source, reflections that genuinely contain the surrounding scene, depth of field that matches the stated lens, and skin or fabric micro-texture that stays irregular at 100% zoom.
Beauty retouching, noise reduction and small files destroy texture even in authentic photos.
Uthentic never hands the verdict to a single model: open-source detectors, a vision model that reasons about the scene, file data and any C2PA provenance traces, a local reading of skin, and a frame integrity map, all fused together after calibration.
When the signals genuinely contradict each other, an arbitration step kicks in: no number is forced.
The report highlights the suspicious areas on the image, states the signal that raised the flag, and explains in which lighting or context conditions deception is most likely.
A flagged area is a point to check with the source, not an automatic conviction.
The report exports as a PDF with a forensic checklist and a verification code: attach it to the story, pass it to an editor or a lawyer, keep it alongside the source's answers.
No tool is 100% reliable: the report exists to justify an editorial decision, not to replace it.
This is the distinction that saves you from a correction later. A generated image has no camera behind it: no sensor signature, no lens traces, a texture that's too uniform across the whole frame. A real photograph worked on in post-production, on the other hand, keeps its camera fingerprint everywhere except in the area that was changed: a replaced sky, a removed bystander, generative fill, smoothed skin all leave a local discontinuity.
Uthentic reports this case as an authentic photograph with an edited area, instead of calling the whole image AI — the mistake that produces most of the false accusations against photographers who were simply doing their job in the darkroom.
For a press release, a statement sent by email or an outside contribution, the text analysis is separate from the photo analysis and has its own report: how predictable the writing is, how uniform the rhythm is between sentences, and which recurring phrasing is typical of language models — each one explained on its own.
Caution matters even more with text than with images: very formal, textbook-style writing can resemble a model's output. That's why the report declares insufficient evidence when the signals don't add up, instead of producing a percentage that won't hold up in a newsroom.
Images are only used while the analysis runs and are deleted right after: they never feed the training of any model. Material under embargo or protected to shield a source can therefore be checked without leaving an archive behind.