AI-edited images: when only one region was changed
Most image fraud today is not fully generated pictures. It is real photos where one region was altered with AI: damage enlarged, an object removed, a detail added. It is called inpainting, it is harder to detect — and it leaves traces.
Updated 2026-09-11 · Veriflied
What inpainting is
You mark a region in a real photo and ask an AI model to fill it with something else: 'remove the scratch', 'make the crack bigger', 'add water damage on the floor'. The rest of the image stays genuine, with genuine sensor noise and genuine metadata. The tools are built into Photoshop (Generative Fill), phone apps and free web services — no skill required.
That is why it is the preferred method in fraud: a fully generated image must invent an entire credible scene, while inpainting only has to lie about one region in a scene that is demonstrably real.
Why it is harder to detect
A detector judging the whole image at once mostly sees genuine photo and often lands on doubt or 'real'. Metadata does not help on its own either: the file IS a real photo, just no longer a truthful one. Detection therefore has to work at region level: analyse the image in patches and find the patch whose statistics do not match the rest.
The traces an inserted region leaves
A statistical island
The generated region carries the generator's micro-structure; the rest carries the sensor's. The boundary between them is measurable even when invisible.
Noise that does not fit
Camera noise varies systematically with light and ISO. An inserted region breaks that pattern.
The logic of lighting
Shadow and light direction in the inserted region rarely match the scene perfectly.
The thumbnail inside the file
If the image is edited without updating the embedded thumbnail, the thumbnail shows the original — before the damage got 'bigger'.
The editor's fingerprint
Photoshop and apps often write themselves into metadata, and the export changes the compression profile.
What it looks like in practice
A forensic analysis shows a heatmap: cold over the genuine part, warm over the inserted region. In a claim that means concretely: 'the photo is real, but the damage area on the bumper is generated'. That is a different and far more actionable conclusion than an overall score — the investigator knows exactly what to ask about.
Frequently asked questions
Is inpainting harder to catch than fully generated images?
Yes. The signal is confined to a small region in an otherwise genuine image. That is why it requires region-level detection, not a single overall judgement.
Does detection survive sharing via messaging apps?
Compression dampens the signal but rarely removes it entirely. The combination of region analysis, noise patterns and metadata still carries an assessment.
Can you see what the region contained before?
Not from the pixels alone. But the embedded thumbnail, earlier versions of the file or the original image from another source can show the original — that is part of the forensic examination.
Try it on your own document
Create an account and analyse your first 10 images or documents for free. You get a report with the concrete findings, not just a score. All processing in the EU.