The AI check in VeriSight returns a single number between 0 and 100%. It is tempting to read a result of 92% as “there is a 92% chance this receipt is fake” and move on. That reading is wrong in a way that matters, both when the number is high and when it is low. This guide explains what the score measures, where it goes wrong, and how to use it in a finding that will survive scrutiny.
What the score measures
The score is a model's estimate of how strongly an image shows the patterns it learned from AI-generated and AI-edited images. It looks at the pixels, not at what the receipt says. Three things follow from that:
- It is not a fraud probability. An image can be AI-edited for an innocent reason, and a fraudulent receipt can be a real photo of a real piece of paper that was later altered by hand.
- It is not a share of the image. 60% does not mean 60% of the pixels were generated. A small AI edit to one field can score lower than a fully generated image, because most of the pixels around it are genuine.
- It covers AI generation and AI editing only. Edits made with conventional image software, such as a total typed over with a text tool, are what the tamper check and the visual review are for. Face swaps are not covered by any of the checks.
The four bands
To make the number easier to act on, VeriSight sorts it into four bands. The band edges are reading conventions, not legal thresholds. A score of 49% and a score of 51% are not meaningfully different, even though they fall in different bands.
| Score | Band and what to do next |
|---|---|
| 0–14% | No AI signal detected This check found nothing. It does not clear the image: the tamper check, the visual review and the content checks still apply. |
| 15–49% | Weak or inconclusive AI signal Do not conclude either way. Get the original file if you analyzed a copy, and lean on the other checks and on independent records. |
| 50–84% | Likely AI-generated or AI-edited Treat it as a lead. Look for corroboration before you raise it with anyone, starting with the card transaction and the original file. |
| 85–100% | Very likely AI-generated or AI-edited A strong lead. Corroborate and document before acting, and keep the analyzed file exactly as you received it. |
The generator list
When there is something to attribute, the report also lists the generators the image most resembles, each with its own score. Read this as a family resemblance, not as proof of which tool was used. It is most useful as a consistency check: a confident attribution backs up a high overall score, while a list in which no single generator stands out is a reason to hold a moderate score more loosely.
Why a score can be too low
A low score means the model did not find the patterns it looks for. It does not mean they were never there. The most common reasons a manipulated image scores low:
- The file is a copy of a copy. Forwarding through messaging apps or email recompresses images and strips fine detail, and that detail is where much of the signal lives. See why the original file matters.
- It is a screenshot or a downscaled image. Both resample every pixel and leave fewer traces of how the image was made.
- It was re-photographed. A generated receipt printed out or shown on a screen and then photographed with a phone is, as far as the pixels go, a real photo. Moiré patterns, screen glare or a visible screen edge are clues worth noting.
- Only a small part was edited. As above, most of the image is genuine and the score reflects that.
- The generator is new. Detection keeps pace with new generators through continuous model updates, so output from a brand-new generator can score low for a while.
Why a score can be too high
High scores on innocent images are less common, but they happen, and they are the ones that can hurt an employee who did nothing wrong.
- Built-in AI photo features. Many phones and gallery apps now offer AI enhancement, upscaling, object removal and background effects, and some apply processing automatically in low light. An image touched by these features is AI-edited in a literal sense, and the model is right to say so. It is not evidence that the content was falsified.
- Beautification filters and heavy editing apps. Common on meeting photos, and they can push scores up across the whole image.
- Images that were never photographs. Rendered e-receipts, exported PDFs and designed graphics are not what the model mostly sees. Treat a high score on one with extra caution and check the content against the issuer instead.
How much weight the score should carry
A score is strongest when it agrees with evidence it did not produce. Some combinations that come up often:
| What you have | How to read it |
|---|---|
| High score, no matching card transaction | A strong combination. The image and the payment record point the same way. |
| High score, the tamper check outlines the total | Two independent models flag the same image. Check the outlined field in the views. |
| High score, everything else matches | Ask how the photo was taken and request the original before going further. |
| Low score on a forwarded screenshot | Says little. Recompression erodes evidence, so it weakens a negative result more than a positive one. |
| Low score, but the sums do not add up | The content check wins. A receipt can be fabricated without AI. |
Writing it up
A finding should let someone who was not there understand what you analyzed, what came back and what you did with it. Name the file and where it came from, quote the score and the band rather than paraphrasing them, and say what you corroborated it with.
The AI check returned 92% (very likely AI-generated or AI-edited) on receipt_0412.jpg, received from the employee as an email attachment on October 3, 2026. The file is a JPEG of 3024 × 4032 pixels and appears to be an original phone photo. We treated the result as a lead. The corporate card statement shows no transaction at this merchant between September 28 and October 3, and the restaurant confirmed it has no record of the receipt number.
When you raise it with the employee, do not present the score as proof. Ask open questions: where the original is, which card paid, who attended. The score earned you the question; the answers and the records make the finding.