AI-Powered Image Authenticity Analysis
See Beyond the Image.
VeriSight uses AI to analyze images, identify potential signs of manipulation or AI generation, and support evidence-based reviews.
Sample result
AI likelihood 92%
How it works
Three steps, a few seconds
- Step 1
Upload a photo
JPEG, PNG or WebP, up to 4 MB. The image stays in memory for the length of the analysis and is never stored.
- Step 2
Run two checks
A detection model scores AI generation or AI editing. VeriSight then computes four views of the image, the cyan and magenta ink channels, an ink colour map and an error level analysis, which make typed-over text and pasted regions easier to spot.
- Step 3
Review the evidence
You get an AI likelihood score with the most probable generators, the four views to inspect side by side with notes on how to read them, and a report you can print.
Reading a result
What the two checks mean
The AI check returns a likelihood between 0 and 100% that falls into one of four bands. The visual review gives you four views computed from the image, each with a note on how to read it. Both are starting points for a reviewer, not a decision.
AI likelihood
- 0–14%No AI signal detected
The detector found little or no evidence of AI generation or AI editing in this image.
- 15–49%Weak or inconclusive AI signal
The detector found a weak signal. It is not strong enough to conclude that the image is AI-generated or AI-edited.
- 50–84%Likely AI-generated or AI-edited
The detector found clear signs of AI generation or AI editing. Corroborate with other evidence before drawing conclusions.
- 85–100%Very likely AI-generated or AI-edited
The detector found strong signs of AI generation or AI editing in this image.
Visual review
Cyan ink channel
CMYK cyan channel, brightness = ink amount (black = none, white = full). Fields printed in one pass share one tone; a field that is markedly brighter or darker than comparable fields was produced differently.
Magenta ink channel
CMYK magenta channel, brightness = ink amount. Read it like the cyan channel and compare text fields of the same kind.
Ink colour map
Dark strokes with their colour cast amplified five times; everything else faded. Text written in one pass shares one cast.
Error level analysis
The image re-saved as JPEG at quality 95 and the change in each pixel amplified. Regions saved the same number of times at the same quality change by similar amounts, so a text field or patch that is markedly brighter or darker than comparable content, or that shows a different block pattern, carries a different saving history. Bold and high-contrast strokes are always brighter than light ones, so compare like with like.
Limits you should know
Where detection falls short
- A score is a probability estimate, not proof. Treat it as one piece of evidence and corroborate before acting on it.
- The views show where to look, not whether the image was edited. Bold or high-contrast text is always brighter in the error level view, and personalised fields printed on a different printer can differ in the channel views without any edit. Compare like with like.
- Images sent through messaging apps or email are recompressed and lose detail. When you can, analyze the original file.
- Screenshots and heavily downscaled images carry fewer traces of generation or editing, so results are less reliable.
- Face swaps and other manipulated faces are not covered by this analysis.
- Detection keeps pace with new generators only through continuous model updates. A low score on output from a brand-new generator is possible.
Privacy
Nothing is kept
Each image is held in memory only for the duration of the analysis and then discarded. VeriSight has no database and no file storage. A report exists only in your browser unless you print it.
Analyze an image