Guide

Why Does the Same Image Get Different AI Detector Scores?

Two AI detectors can score the exact same file and disagree. That comparison is between detectors. A separate comparison is one detector scoring a crop, a screenshot, or another save. Those are different questions.

Two detectors, one unchanged file

Give both detectors the same file. Each one processes the image according to its own design and turns it into a score. The models may have learned different patterns from different training images, and they may use different scoring scales or cutoffs. Both numbers can be called scores and still not mean the same thing.

Ojha, Li, and Lee found that classifiers trained to catch images from one kind of generator often missed images from newer generators, and treated those images as real. That finding describes the classifiers in their study. It is not a measurement of Real Image Check.

Do not add the two numbers together, and do not treat the higher one as more sure. A result of 0.80 on this site is this site’s estimate. A result of 0.80 on another site is that site’s estimate. Another site’s cutoff is that site’s label rule. It does not place its score on this site’s scale.

The same detector, a different file

A crop, a resize, a new JPEG, an edit, or a screenshot can look like the same picture and still be a different file. The same detector is then scoring that new file. This site has not published a test of how often those changes move its own score.

Grommelt, Weiss, Pfreundt, and Keuper found that, for the detectors they tested, JPEG compression could change the result. That paper is not a test of Real Image Check. For screenshots, compression, resizing, cropping, and editing, see Can an AI image detector check a screenshot?

How this site’s score gets its label

The score on this site runs from 0.000 to 1.000. At 0.65 or above, the label is Likely AI-generated. Below 0.65, the label is Likely a real photo. Scores on the same side of 0.65 share that label. A result of 0.64 and a result of 0.66 are opposite labels. A result of 0.66 and a result of 0.90 share a label. Neither gap is a measure of extra confidence.

The score is an estimate, not a calibrated probability or proof. The 0.65 line chooses the label. It does not guarantee a particular accuracy. For false positives, false negatives, and what a wrong label means, see Can AI image detectors get it wrong?

What to do when two scores disagree

Same file, two detectors. Write both results down. Do not blend them into one number. A higher score on another site does not overrule the label here, and it does not prove the picture is AI-generated or real.

Same detector, two versions. You checked two files. For what a changed copy can do, see Can an AI image detector check a screenshot? For where the image was published, see How to tell if an image is AI-generated.

One score does not establish who made the image, prove the image is authentic, name a generator, or read a watermark or Content Credentials. Real Image Check does not verify another detector’s score.

What Real Image Check estimates

The checker scores one JPEG, PNG, or WebP image, up to 20 MB. The score is calculated in your browser. The image is not uploaded for analysis.

A short checklist

• Confirm both scores came from the same file.
• Write down which detector produced each score.
• Do not average scores from different detectors.
• Do not read another site’s higher number as more certain.
• On this site, read the label at 0.65. The score is an estimate, not proof.

Questions